Thread: Full page writes improvement, code update
Hi, Here's an update of a code to improve full page writes as proposed in http://archives.postgresql.org/pgsql-hackers/2007-01/msg01491.php and http://archives.postgresql.org/pgsql-patches/2007-01/msg00607.php Update includes some modification for error handling in archiver and restoration command. In the previous threads, I posted several evaluation and shown that we can keep all the full page writes needed for full XLOG crash recovery, while compressing archive log size considerably better than gzip, with less CPU consumption. I've found no further objection for this proposal but still would like to hear comments/opinions/advices. Regards; -- Koichi Suzuki
Attachment
On Tue, 2007-03-27 at 11:52 +0900, Koichi Suzuki wrote: > Here's an update of a code to improve full page writes as proposed in > > http://archives.postgresql.org/pgsql-hackers/2007-01/msg01491.php > and > http://archives.postgresql.org/pgsql-patches/2007-01/msg00607.php > > Update includes some modification for error handling in archiver and > restoration command. > > In the previous threads, I posted several evaluation and shown that we > can keep all the full page writes needed for full XLOG crash recovery, > while compressing archive log size considerably better than gzip, with > less CPU consumption. I've found no further objection for this proposal > but still would like to hear comments/opinions/advices. Koichi-san, Looks interesting. I like the small amount of code to do this. A few thoughts: - Not sure why we need "full_page_compress", why not just mark them always? That harms noone. (Did someone else ask for that? If so, keep it) - OTOH I'd like to see an explicit parameter set during recovery since you're asking the main recovery path to act differently in case a single bit is set/unset. If you are using that form of recovery, we should say so explicitly, to keep everybody else safe. - I'd rather mark just the nonremovable blocks. But no real difference - We definitely don't want an normal elog in a XLogInsert critical section, especially at DEBUG1 level - diff -c format is easier and the standard - pg_logarchive and pg_logrestore should be called by a name that reflects what they actually do. Possibly pg_compresslog and pg_decompresslog etc.. I've not reviewed those programs, but: - Not sure why we have to compress away page headers. Touch as little as you can has always been my thinking with recovery code. - I'm very uncomfortable with touching the LSN. Maybe I misunderstand? - Have you thought about how pg_standby would integrate with this option? Can you please? - I'll do some docs for this after Freeze, if you'd like. I have some other changes to make there, so I can do this at the same time. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Simon; Thanks a lot for your comments/advices. I'd like to write some feedback. Simon Riggs wrote: > On Tue, 2007-03-27 at 11:52 +0900, Koichi Suzuki wrote: > >> Here's an update of a code to improve full page writes as proposed in >> >> http://archives.postgresql.org/pgsql-hackers/2007-01/msg01491.php >> and >> http://archives.postgresql.org/pgsql-patches/2007-01/msg00607.php >> >> Update includes some modification for error handling in archiver and >> restoration command. >> >> In the previous threads, I posted several evaluation and shown that we >> can keep all the full page writes needed for full XLOG crash recovery, >> while compressing archive log size considerably better than gzip, with >> less CPU consumption. I've found no further objection for this proposal >> but still would like to hear comments/opinions/advices. > > Koichi-san, > > Looks interesting. I like the small amount of code to do this. > > A few thoughts: > > - Not sure why we need "full_page_compress", why not just mark them > always? That harms noone. (Did someone else ask for that? If so, keep > it) No, no one asked to have a separate option. There'll be no bad influence to do so. As written below, full page write can be categolized as follows: 1) Needed for crash recovery: first page update after each checkpoint. This has to be kept in WAL. 2) Needed for archive recovery: page update between pg_start_backup and pg_stop_backup. This has to be kept in archive log. 3) For log-shipping slave such as pg_standby: no full page writes will be needed for this purpose. My proposal deals with 2). So, if we mark each "full_page_write", I'd rather mark when this is needed. Still need only one bit because the case 3) does not need any mark. > - OTOH I'd like to see an explicit parameter set during recovery since > you're asking the main recovery path to act differently in case a single > bit is set/unset. If you are using that form of recovery, we should say > so explicitly, to keep everybody else safe. Only one thing I had to do is to create "dummy" full page write to maintain LSNs. Full page writes are omitted in archive log. We have to LSNs same as those in the original WAL. In this case, recovery has to read logical log, not "dummy" full page writes. On the other hand, if both logical log and "real" full page writes are found in a log record, the recovery has to use "real" full page writes. > - I'd rather mark just the nonremovable blocks. But no real difference It sound nicer. According to the full page write categories above, we can mark full page writes as needed in "crash recovery" or "archive recovery". Please give some feedback to the above full page write categories. > > - We definitely don't want an normal elog in a XLogInsert critical > section, especially at DEBUG1 level I agree. I'll remove elog calls from critical sections. > - diff -c format is easier and the standard I'll change the diff option. > - pg_logarchive and pg_logrestore should be called by a name that > reflects what they actually do. Possibly pg_compresslog and > pg_decompresslog etc.. I've not reviewed those programs, but: I wasn't careful to have command names more based on what to be done. I'll change the command name. > > - Not sure why we have to compress away page headers. Touch as little as > you can has always been my thinking with recovery code. Eliminating page headers gives compression rate slightly better, a couple of percents. To make code simpler, I'll drop this compression. > > - I'm very uncomfortable with touching the LSN. Maybe I misunderstand? The reason why we need pg_logarchive (or pg_decompresslog) is to maintain LSN the same as those in the original WAL. For this purpose, we need "dummy" full page write. I don't like to touch LSN either and this is the reason why pg_decompresslog need some extra work, eliminating many other changes in the recovery. > > - Have you thought about how pg_standby would integrate with this > option? Can you please? Yes I believe so. As pg_standby does not include any chance to meet partial writes of pages, I believe you can omit all the full page writes. Of course, as Tom Lange suggested in http://archives.postgresql.org/pgsql-hackers/2007-02/msg00034.php removing full page writes can lose a chance to recover from partial/inconsisitent writes in the crash of pg_standby. In this case, we have to import a backup and archive logs (with full page writes during the backup) to recover. (We have to import them when the file system crashes anyway). If it's okay, I believe pg_compresslog/pg_decompresslog can be integrated with pg_standby. Maybe we can work together to include pg_compresslog/pg_decompresslog in pg_standby. > > - I'll do some docs for this after Freeze, if you'd like. I have some > other changes to make there, so I can do this at the same time. I'll be looking forward to your writings. Best regards; -- Koichi Suzuki
On Wed, 2007-03-28 at 10:54 +0900, Koichi Suzuki wrote: > As written below, full page write can be > categolized as follows: > > 1) Needed for crash recovery: first page update after each checkpoint. > This has to be kept in WAL. > > 2) Needed for archive recovery: page update between pg_start_backup and > pg_stop_backup. This has to be kept in archive log. > > 3) For log-shipping slave such as pg_standby: no full page writes will > be needed for this purpose. > > My proposal deals with 2). So, if we mark each "full_page_write", I'd > rather mark when this is needed. Still need only one bit because the > case 3) does not need any mark. I'm very happy with this proposal, though I do still have some points in detailed areas. If you accept that 1 & 2 are valid goals, then 1 & 3 or 1, 2 & 3 are also valid goals, ISTM. i.e. you might choose to use full_page_writes on the primary and yet would like to see optimised data transfer to the standby server. In that case, you would need the mark. > > - Not sure why we need "full_page_compress", why not just mark them > > always? That harms noone. (Did someone else ask for that? If so, keep > > it) > > No, no one asked to have a separate option. There'll be no bad > influence to do so. So, if we mark each "full_page_write", I'd > rather mark when this is needed. Still need only one bit because the > case 3) does not need any mark. OK, different question: Why would anyone ever set full_page_compress = off? Why have a parameter that does so little? ISTM this is: i) one more thing to get wrong ii) cheaper to mark the block when appropriate than to perform the if() test each time. That can be done only in the path where backup blocks are present. iii) If we mark the blocks every time, it allows us to do an offline WAL compression. If the blocks aren't marked that option is lost. The bit is useful information, so we should have it in all cases. > > - OTOH I'd like to see an explicit parameter set during recovery since > > you're asking the main recovery path to act differently in case a single > > bit is set/unset. If you are using that form of recovery, we should say > > so explicitly, to keep everybody else safe. > > Only one thing I had to do is to create "dummy" full page write to > maintain LSNs. Full page writes are omitted in archive log. We have to > LSNs same as those in the original WAL. In this case, recovery has to > read logical log, not "dummy" full page writes. On the other hand, if > both logical log and "real" full page writes are found in a log record, > the recovery has to use "real" full page writes. I apologise for not understanding your reply, perhaps my original request was unclear. In recovery.conf, I'd like to see a parameter such as dummy_backup_blocks = off (default) | on to explicitly indicate to the recovery process that backup blocks are present, yet they are garbage and should be ignored. Having garbage data within the system is potentially dangerous and I want to be told by the user that they were expecting that and its OK to ignore that data. Otherwise I want to throw informative errors. Maybe it seems OK now, but the next change to the system may have unintended consequences and it may not be us making the change. "It's OK the Alien will never escape from the lab" is the starting premise for many good sci-fi horrors and I want to watch them, not be in one myself. :-) We can call it other things, of course. e.g. ignore_dummy_blocks decompressed_blocks apply_backup_blocks > Yes I believe so. As pg_standby does not include any chance to meet > partial writes of pages, I believe you can omit all the full page > writes. Of course, as Tom Lange suggested in > http://archives.postgresql.org/pgsql-hackers/2007-02/msg00034.php > removing full page writes can lose a chance to recover from > partial/inconsisitent writes in the crash of pg_standby. In this case, > we have to import a backup and archive logs (with full page writes > during the backup) to recover. (We have to import them when the file > system crashes anyway). If it's okay, I believe > pg_compresslog/pg_decompresslog can be integrated with pg_standby. > > Maybe we can work together to include pg_compresslog/pg_decompresslog in > pg_standby. ISTM there are two options. I think this option is already possible: 1. Allow pg_decompresslog to operate on a file, replacing it with the expanded form, like gunzip, so we would do this: restore_command = 'pg_standby %f decomp.tmp && pg_decompresslog decomp.tmp %p' though the decomp.tmp file would not get properly initialised or cleaned up when we finish. whereas this will take additional work 2. Allow pg_standby to write to stdin, so that we can do this: restore_command = 'pg_standby %f | pg_decompresslog - %p' -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Hi, Here're some feedback to the comment: Simon Riggs wrote: > On Wed, 2007-03-28 at 10:54 +0900, Koichi Suzuki wrote: > >> As written below, full page write can be >> categolized as follows: >> >> 1) Needed for crash recovery: first page update after each checkpoint. >> This has to be kept in WAL. >> >> 2) Needed for archive recovery: page update between pg_start_backup and >> pg_stop_backup. This has to be kept in archive log. >> >> 3) For log-shipping slave such as pg_standby: no full page writes will >> be needed for this purpose. >> >> My proposal deals with 2). So, if we mark each "full_page_write", I'd >> rather mark when this is needed. Still need only one bit because the >> case 3) does not need any mark. > > I'm very happy with this proposal, though I do still have some points in > detailed areas. > > If you accept that 1 & 2 are valid goals, then 1 & 3 or 1, 2 & 3 are > also valid goals, ISTM. i.e. you might choose to use full_page_writes on > the primary and yet would like to see optimised data transfer to the > standby server. In that case, you would need the mark. Yes, I need the mark. In my proposal, only unmarked full-page-writes, which were written as the first update after a checkpoint, are to be removed offline (pg_compresslog). > >>> - Not sure why we need "full_page_compress", why not just mark them >>> always? That harms noone. (Did someone else ask for that? If so, keep >>> it) >> No, no one asked to have a separate option. There'll be no bad >> influence to do so. So, if we mark each "full_page_write", I'd >> rather mark when this is needed. Still need only one bit because the >> case 3) does not need any mark. > > OK, different question: > Why would anyone ever set full_page_compress = off? > > Why have a parameter that does so little? ISTM this is: > > i) one more thing to get wrong > > ii) cheaper to mark the block when appropriate than to perform the if() > test each time. That can be done only in the path where backup blocks > are present. > > iii) If we mark the blocks every time, it allows us to do an offline WAL > compression. If the blocks aren't marked that option is lost. The bit is > useful information, so we should have it in all cases. Not only full-page-writes are written as WAL record. In my proposal, both full-page-writes and logical log are written in a WAL record, which will make WAL size slightly bigger (five percent or so). If full_page_compress = off, only a full-page-write will be written in a WAL record. I thought someone will not be happy with this size growth. I agree to make this mandatory if every body is happy with extra logical log in WAL records with full page writes. I'd like to have your opinion. > >>> - OTOH I'd like to see an explicit parameter set during recovery since >>> you're asking the main recovery path to act differently in case a single >>> bit is set/unset. If you are using that form of recovery, we should say >>> so explicitly, to keep everybody else safe. >> Only one thing I had to do is to create "dummy" full page write to >> maintain LSNs. Full page writes are omitted in archive log. We have to >> LSNs same as those in the original WAL. In this case, recovery has to >> read logical log, not "dummy" full page writes. On the other hand, if >> both logical log and "real" full page writes are found in a log record, >> the recovery has to use "real" full page writes. > > I apologise for not understanding your reply, perhaps my original > request was unclear. > > In recovery.conf, I'd like to see a parameter such as > > dummy_backup_blocks = off (default) | on > > to explicitly indicate to the recovery process that backup blocks are > present, yet they are garbage and should be ignored. Having garbage data > within the system is potentially dangerous and I want to be told by the > user that they were expecting that and its OK to ignore that data. > Otherwise I want to throw informative errors. Maybe it seems OK now, but > the next change to the system may have unintended consequences and it > may not be us making the change. "It's OK the Alien will never escape > from the lab" is the starting premise for many good sci-fi horrors and I > want to watch them, not be in one myself. :-) > > We can call it other things, of course. e.g. > ignore_dummy_blocks > decompressed_blocks > apply_backup_blocks So far, we don't need any modification to the recovery and redo functions. They ignore the dummy and apply logical logs. Also, if there are both full page writes and logical log, current recovery selects full page writes to apply. I agree to introduce this option if 8.3 code introduces any conflict to the current. Or, we could introduce this option for future safety. Do you think we should introduce this option? If this should be introduced now, what we should do is to check this option when dummy full-page-write appears. > >> Yes I believe so. As pg_standby does not include any chance to meet >> partial writes of pages, I believe you can omit all the full page >> writes. Of course, as Tom Lange suggested in >> http://archives.postgresql.org/pgsql-hackers/2007-02/msg00034.php >> removing full page writes can lose a chance to recover from >> partial/inconsisitent writes in the crash of pg_standby. In this case, >> we have to import a backup and archive logs (with full page writes >> during the backup) to recover. (We have to import them when the file >> system crashes anyway). If it's okay, I believe >> pg_compresslog/pg_decompresslog can be integrated with pg_standby. >> >> Maybe we can work together to include pg_compresslog/pg_decompresslog in >> pg_standby. > > ISTM there are two options. > > I think this option is already possible: > > 1. Allow pg_decompresslog to operate on a file, replacing it with the > expanded form, like gunzip, so we would do this: > restore_command = 'pg_standby %f decomp.tmp && pg_decompresslog > decomp.tmp %p' > > though the decomp.tmp file would not get properly initialised or cleaned > up when we finish. > > whereas this will take additional work > > 2. Allow pg_standby to write to stdin, so that we can do this: > restore_command = 'pg_standby %f | pg_decompresslog - %p' > Both seem to work fine. pg_decompresslog will read entire file at the beginning and so both two will be equivallent. To make the second option run quicker, pg_decompresslog needs some modification to read WAL record one after another. Anyway, could you try to run pg_standby with pg_compresslog and pg_decompresslog? ---- Additional recomment on page header removal: I found that it is not simple to keep page header in the compressed archive log. Because we eliminate unmarked full page writes and shift the rest of the WAL file data, it is not simple to keep page header as the page header in the compressed archive log. It is much simpler to remove page header as well and rebuild them. I'd like to keep current implementation in this point. Any suggestions are welcome. ------------------- I'll modify the name of the commands and post it hopefully within 20hours. With Best Regards; -- Koichi Suzuki
On Thu, 2007-03-29 at 17:50 +0900, Koichi Suzuki wrote: > Not only full-page-writes are written as WAL record. In my proposal, > both full-page-writes and logical log are written in a WAL record, which > will make WAL size slightly bigger (five percent or so). If > full_page_compress = off, only a full-page-write will be written in a > WAL record. I thought someone will not be happy with this size growth. OK, I see what you're doing now and agree with you that we do need a parameter. Not sure about the name you've chosen though - it certainly confused me until you explained. A parameter called ..._compress indicates to me that it would reduce something in size whereas what it actually does is increase the size of WAL slightly. We should have a parameter name that indicates what it actually does, otherwise some people will choose to use this parameter even when they are not using archive_command with pg_compresslog. Some possible names... additional_wal_info = 'COMPRESS' add_wal_info wal_additional_info wal_auxiliary_info wal_extra_data attach_wal_info ... others? I've got some ideas for the future for adding additional WAL info for various purposes, so it might be useful to have a parameter that can cater for multiple types of additional WAL data. Or maybe we go for something more specific like wal_add_compress_info = on wal_add_XXXX_info ... > > In recovery.conf, I'd like to see a parameter such as > > > > dummy_backup_blocks = off (default) | on > > > > to explicitly indicate to the recovery process that backup blocks are > > present, yet they are garbage and should be ignored. Having garbage data > > within the system is potentially dangerous and I want to be told by the > > user that they were expecting that and its OK to ignore that data. > > Otherwise I want to throw informative errors. Maybe it seems OK now, but > > the next change to the system may have unintended consequences and it > > may not be us making the change. "It's OK the Alien will never escape > > from the lab" is the starting premise for many good sci-fi horrors and I > > want to watch them, not be in one myself. :-) > > > > We can call it other things, of course. e.g. > > ignore_dummy_blocks > > decompressed_blocks > > apply_backup_blocks > > So far, we don't need any modification to the recovery and redo > functions. They ignore the dummy and apply logical logs. Also, if > there are both full page writes and logical log, current recovery > selects full page writes to apply. > > I agree to introduce this option if 8.3 code introduces any conflict to > the current. Or, we could introduce this option for future safety. Do > you think we should introduce this option? Yes. You are skipping a correctness test and that should be by explicit command only. It's no problem to include that as well, since you are already having to specify pg_... decompress... but the recovery process doesn't know whether or not you've done that. > Anyway, could you try to run pg_standby with pg_compresslog and > pg_decompresslog? After freeze, yes. > ---- > Additional recomment on page header removal: > > I found that it is not simple to keep page header in the compressed > archive log. Because we eliminate unmarked full page writes and shift > the rest of the WAL file data, it is not simple to keep page header as > the page header in the compressed archive log. It is much simpler to > remove page header as well and rebuild them. I'd like to keep current > implementation in this point. OK. This is a good feature. Thanks for your patience with my comments. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Simon, > OK, different question: > Why would anyone ever set full_page_compress = off? The only reason I can see is if compression costs us CPU but gains RAM & I/O. I can think of a lot of applications ... benchmarks included ... which are CPU-bound but not RAM or I/O bound. For those applications, compression is a bad tradeoff. If, however, CPU used for compression is made up elsewhere through smaller file processing, then I'd agree that we don't need a switch. -- --Josh Josh Berkus PostgreSQL @ Sun San Francisco
On Thu, 2007-03-29 at 11:45 -0700, Josh Berkus wrote: > > OK, different question: > > Why would anyone ever set full_page_compress = off? > > The only reason I can see is if compression costs us CPU but gains RAM & > I/O. I can think of a lot of applications ... benchmarks included ... > which are CPU-bound but not RAM or I/O bound. For those applications, > compression is a bad tradeoff. > > If, however, CPU used for compression is made up elsewhere through smaller > file processing, then I'd agree that we don't need a switch. Koichi-san has explained things for me now. I misunderstood what the parameter did and reading your post, ISTM you have as well. I do hope Koichi-san will alter the name to allow everybody to understand what it does. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Josh;
I'd like to explain what the term "compression" in my proposal means
again and would like to show the resource consumption comparision with
cp and gzip.
My proposal is to remove unnecessary full page writes (they are needed
in crash recovery from inconsistent or partial writes) when we copy WAL
to archive log and rebuilt them as a dummy when we restore from archive
log. Dummy is needed to maintain LSN. So it is very very different
from general purpose compression such as gzip, although pg_compresslog
compresses archive log as a result.
As to CPU and I/O consumption, I've already evaluated as follows:
1) Collect all the WAL segment.
2) Copy them by different means, cp, pg_compresslog and gzip.
and compared the ellapsed time as well as other resource consumption.
Benchmark: DBT-2
Database size: 120WH (12.3GB)
Total WAL size: 4.2GB (after 60min. run)
Elapsed time:
cp: 120.6sec
gzip: 590.0sec
pg_compresslog: 79.4sec
Resultant archive log size:
cp: 4.2GB
gzip: 2.2GB
pg_compresslog: 0.3GB
Resource consumption:
cp: user: 0.5sec system: 15.8sec idle: 16.9sec I/O wait: 87.7sec
gzip: user: 286.2sec system: 8.6sec idle: 260.5sec I/O wait: 36.0sec
pg_compresslog:
user: 7.9sec system: 5.5sec idle: 37.8sec I/O wait: 28.4sec
Because the resultant log size is considerably smaller than cp or gzip,
pg_compresslog need much less I/O and because the logic is much simpler
than gzip, it does not consume CPU.
The term "compress" may not be appropriate. We may call this "log
optimization" instead.
So I don't see any reason why this (at least optimization "mark" in each
log record) can't be integrated.
Simon Riggs wrote:
> On Thu, 2007-03-29 at 11:45 -0700, Josh Berkus wrote:
>
>>> OK, different question:
>>> Why would anyone ever set full_page_compress = off?
>> The only reason I can see is if compression costs us CPU but gains RAM &
>> I/O. I can think of a lot of applications ... benchmarks included ...
>> which are CPU-bound but not RAM or I/O bound. For those applications,
>> compression is a bad tradeoff.
>>
>> If, however, CPU used for compression is made up elsewhere through smaller
>> file processing, then I'd agree that we don't need a switch.
As I wrote to Simon's comment, I concern only one thing.
Without a switch, because both full page writes and corresponding
logical log is included in WAL, this will increase WAL size slightly
(maybe about five percent or so). If everybody is happy with this, we
don't need a switch.
>
> Koichi-san has explained things for me now.
>
> I misunderstood what the parameter did and reading your post, ISTM you
> have as well. I do hope Koichi-san will alter the name to allow
> everybody to understand what it does.
>
Here're some candidates:
full_page_writes_optimize
full_page_writes_mark: means it marks full_page_write as "needed in
crash recovery", "needed in archive recovery" and so on.
I don't insist these names. It's very helpful if you have any
suggestion to reflect what it really means.
Regards;
--
Koichi Suzuki
Hi, Here's a patch reflected some of Simon's comments. 1) Removed an elog call in a critical section. 2) Changed the name of the commands, pg_complesslog and pg_decompresslog. 3) Changed diff option to make a patch. -- Koichi Suzuki
Attachment
> Without a switch, because both full page writes and > corresponding logical log is included in WAL, this will > increase WAL size slightly > (maybe about five percent or so). If everybody is happy > with this, we > don't need a switch. Sorry, I still don't understand that. What is the "corresponding logical log" ? It seems to me, that a full page WAL record has enough info to produce a dummy LSN WAL entry. So insead of just cutting the full page wal record you could replace it with a LSN WAL entry when archiving the log. Then all that is needed is the one flag, no extra space ? Andreas
On Fri, 2007-03-30 at 10:22 +0200, Zeugswetter Andreas ADI SD wrote: > > Without a switch, because both full page writes and > > corresponding logical log is included in WAL, this will > > increase WAL size slightly > > (maybe about five percent or so). If everybody is happy > > with this, we > > don't need a switch. > > Sorry, I still don't understand that. What is the "corresponding logical > log" ? > It seems to me, that a full page WAL record has enough info to produce a > > dummy LSN WAL entry. So insead of just cutting the full page wal record > you > could replace it with a LSN WAL entry when archiving the log. > > Then all that is needed is the one flag, no extra space ? The full page write is required for crash recovery, but that isn't required during archive recovery because the base backup provides the safe base. Archive recovery needs the normal xlog record, which in some cases has been optimised away because the backup block is present, since the full block already contains the changes. If you want to remove the backup blocks, you need to put back the information that was optimised away, otherwise you won't be able to do the archive recovery correctly. Hence a slight increase in WAL volume to allow it to be compressed does make sense. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
> Archive recovery needs the > normal xlog record, which in some cases has been optimised > away because the backup block is present, since the full > block already contains the changes. Aah, I didn't know that optimization exists. I agree that removing that optimization is good/ok. Andreas
Simon Riggs wrote: > On Fri, 2007-03-30 at 10:22 +0200, Zeugswetter Andreas ADI SD wrote: >>> Without a switch, because both full page writes and >>> corresponding logical log is included in WAL, this will >>> increase WAL size slightly >>> (maybe about five percent or so). If everybody is happy >>> with this, we >>> don't need a switch. >> Sorry, I still don't understand that. What is the "corresponding logical >> log" ? >> It seems to me, that a full page WAL record has enough info to produce a >> >> dummy LSN WAL entry. So insead of just cutting the full page wal record >> you >> could replace it with a LSN WAL entry when archiving the log. >> >> Then all that is needed is the one flag, no extra space ? > > The full page write is required for crash recovery, but that isn't > required during archive recovery because the base backup provides the > safe base. Is that always true? Could the backup not pick up a partially-written page? Assuming it's being written to as the backup is in progress. (We are talking about when disk blocks are smaller than PG blocks here, so can't guarantee an atomic write for a PG block?) -- Richard Huxton Archonet Ltd
On Fri, 2007-03-30 at 11:27 +0100, Richard Huxton wrote: > Is that always true? Could the backup not pick up a partially-written > page? Assuming it's being written to as the backup is in progress. (We > are talking about when disk blocks are smaller than PG blocks here, so > can't guarantee an atomic write for a PG block?) Any page written during a backup has a backup block that would not be removable by Koichi's tool, so yes, you'd still be safe. i.e. between pg_start_backup() and pg_stop_backup() we always use full page writes, even if you are running in full_page_writes=off mode. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Simon Riggs wrote: > On Fri, 2007-03-30 at 11:27 +0100, Richard Huxton wrote: > >> Is that always true? Could the backup not pick up a partially-written >> page? Assuming it's being written to as the backup is in progress. (We >> are talking about when disk blocks are smaller than PG blocks here, so >> can't guarantee an atomic write for a PG block?) > > Any page written during a backup has a backup block that would not be > removable by Koichi's tool, so yes, you'd still be safe. > > i.e. between pg_start_backup() and pg_stop_backup() we always use full > page writes, even if you are running in full_page_writes=off mode. Ah, that's OK then. -- Richard Huxton Archonet Ltd
"Simon Riggs" <simon@2ndquadrant.com> writes:
> Any page written during a backup has a backup block that would not be
> removable by Koichi's tool, so yes, you'd still be safe.
How does it know not to do that?
regards, tom lane
On Fri, 2007-03-30 at 16:35 -0400, Tom Lane wrote: > "Simon Riggs" <simon@2ndquadrant.com> writes: > > Any page written during a backup has a backup block that would not be > > removable by Koichi's tool, so yes, you'd still be safe. > > How does it know not to do that? Not sure what you mean, but I'll take a stab... I originally questioned Koichi-san's request for a full_page_compress parameter, which is how it would tell whether/not. After explanation, I accepted the need for a parameter, but I think we're looking for a new name for it. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Tom Lane wrote: > "Simon Riggs" <simon@2ndquadrant.com> writes: >> Any page written during a backup has a backup block that would not be >> removable by Koichi's tool, so yes, you'd still be safe. > > How does it know not to do that? > > regards, tom lane > > ---------------------------(end of broadcast)--------------------------- > TIP 5: don't forget to increase your free space map settings > XLogInsert( ) already has a logic to determine if inserting WAL record is between pg_start_backup and pg_stop_backup. Currently it is used to remove full_page_writes when full_page_writes=off. We can use this to mark WAL records. We have one bit not used in WAL record header, the last bit of xl_info, where upper four bits are used to indicate the resource manager and three of the rest are used to indicate number of full page writes included in the record. In my proposal, this unused bit is used to mark that full page writes must not be removed at offline optimization by pg_compresslog. Regards; -- ------ Koichi Suzuki -- Koichi Suzuki
Your patch has been added to the PostgreSQL unapplied patches list at:
http://momjian.postgresql.org/cgi-bin/pgpatches
It will be applied as soon as one of the PostgreSQL committers reviews
and approves it.
---------------------------------------------------------------------------
Koichi Suzuki wrote:
> Hi,
>
> Here's a patch reflected some of Simon's comments.
>
> 1) Removed an elog call in a critical section.
>
> 2) Changed the name of the commands, pg_complesslog and pg_decompresslog.
>
> 3) Changed diff option to make a patch.
>
> --
> Koichi Suzuki
[ Attachment, skipping... ]
>
> ---------------------------(end of broadcast)---------------------------
> TIP 5: don't forget to increase your free space map settings
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com
+ If your life is a hard drive, Christ can be your backup. +
Here's third revision of WAL archival optimization patch. GUC parameter name was changed to wal_add_optimization_info. Regards; -- Koichi Suzuki
Bruce Momjian wrote: > Your patch has been added to the PostgreSQL unapplied patches list at: > > http://momjian.postgresql.org/cgi-bin/pgpatches Thank you very much for including. Attached is an update of the patch according to Simon Riggs's comment about GUC name. Regards; -- Koichi Suzuki
Your patch has been added to the PostgreSQL unapplied patches list at:
http://momjian.postgresql.org/cgi-bin/pgpatches
It will be applied as soon as one of the PostgreSQL committers reviews
and approves it.
---------------------------------------------------------------------------
Koichi Suzuki wrote:
> Bruce Momjian wrote:
> > Your patch has been added to the PostgreSQL unapplied patches list at:
> >
> > http://momjian.postgresql.org/cgi-bin/pgpatches
>
> Thank you very much for including. Attached is an update of the patch
> according to Simon Riggs's comment about GUC name.
>
> Regards;
>
> --
> Koichi Suzuki
[ Attachment, skipping... ]
>
> ---------------------------(end of broadcast)---------------------------
> TIP 6: explain analyze is your friend
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com
+ If your life is a hard drive, Christ can be your backup. +
On Tue, 2007-04-03 at 19:45 +0900, Koichi Suzuki wrote: > Bruce Momjian wrote: > > Your patch has been added to the PostgreSQL unapplied patches list at: > > > > http://momjian.postgresql.org/cgi-bin/pgpatches > > Thank you very much for including. Attached is an update of the patch > according to Simon Riggs's comment about GUC name. The patch comes with its own "install kit", which is great to review (many thanks), but hard to determine where you think code should go when committed. My guess based on your patch - the patch gets applied to core :-) - pg_compresslog *and* pg_decompresslog go to a contrib directory called contrib/lesslog? Can you please produce a combined patch that does all of the above, plus edits the contrib Makefile to add all of those, as well as editing the README so it doesn't mention the patch, just the contrib executables? The patch looks correct to me now. I haven't tested it yet, but will be doing so in the last week of April, which is when I'll be doing docs for this and other stuff, since time is pressing now. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Hi, I agree to put the patch to core and the others (pg_compresslog and pg_decompresslog) to contrib/lesslog. I will make separate materials to go to core and contrib. As for patches, we have tested against pgbench, DBT-2 and our propriatery benchmarks and it looked to run correctly. Regards; Simon Riggs wrote: > On Tue, 2007-04-03 at 19:45 +0900, Koichi Suzuki wrote: >> Bruce Momjian wrote: >>> Your patch has been added to the PostgreSQL unapplied patches list at: >>> >>> http://momjian.postgresql.org/cgi-bin/pgpatches >> Thank you very much for including. Attached is an update of the patch >> according to Simon Riggs's comment about GUC name. > > The patch comes with its own "install kit", which is great to review > (many thanks), but hard to determine where you think code should go when > committed. > > My guess based on your patch > - the patch gets applied to core :-) > - pg_compresslog *and* pg_decompresslog go to a contrib directory called > contrib/lesslog? > > Can you please produce a combined patch that does all of the above, plus > edits the contrib Makefile to add all of those, as well as editing the > README so it doesn't mention the patch, just the contrib executables? > > The patch looks correct to me now. I haven't tested it yet, but will be > doing so in the last week of April, which is when I'll be doing docs for > this and other stuff, since time is pressing now. > -- Koichi Suzuki
Koichi Suzuki <suzuki.koichi@oss.ntt.co.jp> writes:
> My proposal is to remove unnecessary full page writes (they are needed
> in crash recovery from inconsistent or partial writes) when we copy WAL
> to archive log and rebuilt them as a dummy when we restore from archive
> log.
> ...
> Benchmark: DBT-2
> Database size: 120WH (12.3GB)
> Total WAL size: 4.2GB (after 60min. run)
> Elapsed time:
> cp: 120.6sec
> gzip: 590.0sec
> pg_compresslog: 79.4sec
> Resultant archive log size:
> cp: 4.2GB
> gzip: 2.2GB
> pg_compresslog: 0.3GB
> Resource consumption:
> cp: user: 0.5sec system: 15.8sec idle: 16.9sec I/O wait: 87.7sec
> gzip: user: 286.2sec system: 8.6sec idle: 260.5sec I/O wait: 36.0sec
> pg_compresslog:
> user: 7.9sec system: 5.5sec idle: 37.8sec I/O wait: 28.4sec
What checkpoint settings were used to make this comparison? I'm
wondering whether much of the same benefit can't be bought at zero cost
by increasing the checkpoint interval, because that translates directly
to a reduction in the number of full-page images inserted into WAL.
Also, how much was the database run itself slowed down by the increased
volume of WAL (due to duplicated information)? It seems rather
pointless to me to measure only the archiving effort without any
consideration for the impact on the database server proper.
regards, tom lane
PS: there's something fishy about the gzip numbers ... why all the idle
time?
Hi, In the case below, we run DBT-2 benchmark for one hour to get the measure. Checkpoint occured three times (checkpoint interval was 20min). For more information, when checkpoint interval is one hour, the amount of the archived log size was as follows: cp: 3.1GB gzip: 1.5GB pg_compresslog: 0.3GB For both cases, database size was 12.7GB, relatively small. As pointed out, if we don't run the checkpoint forever, the value for cp will become close to that for pg_compresslog, but it is not practical. The point here is, if we collect archive log with cp and the average work load is a quarter of the full power, cp archiving will produce about 0.8GB archive log per hour (for DBT-2 case, of course the size depends on the nature of the transaction). If we run the database whole day, the amount of the archive log will be as large as database itself. After one week, archive log size gets seven times as large as the database itself. This is the point. In production, such large archive log will raise storage cost. The purpose of the proposal is not to improve the performance, but to decrease the size of archive log to save necessary storage, preserving the same chance of recovery at the crash recovery as full_page_writes=on. Because of DBT-2 nature, it is not meaningful to compare the throuput (databsae size determines the number of transactions to run). Instead, I compared the throuput using pgbench. These measures are: cp: 570tps, gzip:558tps, pg_compresslog: 574tps, negligible difference. In terms of idle time for gzip and other command to archive WAL offline, no difference in the environment was given other than the command to archive. My guess is because the user time is very large in gzip, it has more chance for scheduler to give resource to other processes. In the case of cp, idle time is more than 30times longer than user time. Pg_compresslog uses seven times longer idle time than user time. On the other hand, gzip uses less idle time than user time. Considering the total amount of user time, I think it's reasonable measure. Again, in my proposal, it is not the issue to increase run time performance. Issue is to decrease the size of archive log to save the storage. Regards; Tom Lane wrote: > Koichi Suzuki <suzuki.koichi@oss.ntt.co.jp> writes: >> My proposal is to remove unnecessary full page writes (they are needed >> in crash recovery from inconsistent or partial writes) when we copy WAL >> to archive log and rebuilt them as a dummy when we restore from archive >> log. >> ... >> Benchmark: DBT-2 >> Database size: 120WH (12.3GB) >> Total WAL size: 4.2GB (after 60min. run) >> Elapsed time: >> cp: 120.6sec >> gzip: 590.0sec >> pg_compresslog: 79.4sec >> Resultant archive log size: >> cp: 4.2GB >> gzip: 2.2GB >> pg_compresslog: 0.3GB >> Resource consumption: >> cp: user: 0.5sec system: 15.8sec idle: 16.9sec I/O wait: 87.7sec >> gzip: user: 286.2sec system: 8.6sec idle: 260.5sec I/O wait: 36.0sec >> pg_compresslog: >> user: 7.9sec system: 5.5sec idle: 37.8sec I/O wait: 28.4sec > > What checkpoint settings were used to make this comparison? I'm > wondering whether much of the same benefit can't be bought at zero cost > by increasing the checkpoint interval, because that translates directly > to a reduction in the number of full-page images inserted into WAL. > > Also, how much was the database run itself slowed down by the increased > volume of WAL (due to duplicated information)? It seems rather > pointless to me to measure only the archiving effort without any > consideration for the impact on the database server proper. > > regards, tom lane > > PS: there's something fishy about the gzip numbers ... why all the idle > time? > > ---------------------------(end of broadcast)--------------------------- > TIP 4: Have you searched our list archives? > > http://archives.postgresql.org > -- Koichi Suzuki
> In terms of idle time for gzip and other command to archive WAL offline,
> no difference in the environment was given other than the command to
> archive. My guess is because the user time is very large in gzip, it
> has more chance for scheduler to give resource to other processes. In
> the case of cp, idle time is more than 30times longer than user time.
> Pg_compresslog uses seven times longer idle time than user time. On the
> other hand, gzip uses less idle time than user time. Considering the
> total amount of user time, I think it's reasonable measure.
>
> Again, in my proposal, it is not the issue to increase run time
> performance. Issue is to decrease the size of archive log to save the
> storage.
Considering the relatively little amount of storage a transaction log
takes, it would seem to me that the performance angle is more appropriate.
Is it more efficient in other ways besides negligible tps? Possibly more
efficient memory usage? Better restore times for a crashed system?
Sincerely,
Joshua D. Drake
>
> Regards;
>
--
=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997
http://www.commandprompt.com/
Donate to the PostgreSQL Project: http://www.postgresql.org/about/donate
PostgreSQL Replication: http://www.commandprompt.com/products/
Ühel kenal päeval, T, 2007-04-10 kell 18:17, kirjutas Joshua D. Drake: > > In terms of idle time for gzip and other command to archive WAL offline, > > no difference in the environment was given other than the command to > > archive. My guess is because the user time is very large in gzip, it > > has more chance for scheduler to give resource to other processes. In > > the case of cp, idle time is more than 30times longer than user time. > > Pg_compresslog uses seven times longer idle time than user time. On the > > other hand, gzip uses less idle time than user time. Considering the > > total amount of user time, I think it's reasonable measure. > > > > Again, in my proposal, it is not the issue to increase run time > > performance. Issue is to decrease the size of archive log to save the > > storage. > > Considering the relatively little amount of storage a transaction log > takes, it would seem to me that the performance angle is more appropriate. As I understand it it's not about transaction log but about write-ahead log. and the amount of data in WAL can become very important once you have to keep standby servers in different physical locations (cities, countries or continents) where channel throughput and cost comes into play. With simple cp (scp/rsync) the amount of WAL data needing to be copied is about 10x more than data collected by trigger based solutions (Slony/pgQ). With pg_compresslog WAL-shipping seems to have roughly the same amount and thus becomes a viable alternative again. > Is it more efficient in other ways besides negligible tps? Possibly more > efficient memory usage? Better restore times for a crashed system? I think that TPS is more affected by number of writes than size of each block written, so there is probably not that much to gain in TPS, except perhaps from better disk cache usage. For me pg_compresslog seems to be a winner even if it just does not degrade performance. -- ---------------- Hannu Krosing Database Architect Skype Technologies OÜ Akadeemia tee 21 F, Tallinn, 12618, Estonia Skype me: callto:hkrosing Get Skype for free: http://www.skype.com
Koichi Suzuki <suzuki.koichi@oss.ntt.co.jp> writes:
> For more information, when checkpoint interval is one hour, the amount
> of the archived log size was as follows:
> cp: 3.1GB
> gzip: 1.5GB
> pg_compresslog: 0.3GB
The notion that 90% of the WAL could be backup blocks even at very long
checkpoint intervals struck me as excessive, so I went looking for a
reason, and I may have found one. There has been a bug in CVS HEAD
since Feb 8 causing every btree page split record to include a backup
block whether needed or not. If these numbers were taken with recent
8.3 code, please retest with current HEAD.
regards, tom lane
The score below was taken based on 8.2 code, not 8.3 code. So I don't think the below measure is introduced only in 8.3 code. Tom Lane wrote: > Koichi Suzuki <suzuki.koichi@oss.ntt.co.jp> writes: >> For more information, when checkpoint interval is one hour, the amount >> of the archived log size was as follows: >> cp: 3.1GB >> gzip: 1.5GB >> pg_compresslog: 0.3GB > > The notion that 90% of the WAL could be backup blocks even at very long > checkpoint intervals struck me as excessive, so I went looking for a > reason, and I may have found one. There has been a bug in CVS HEAD > since Feb 8 causing every btree page split record to include a backup > block whether needed or not. If these numbers were taken with recent > 8.3 code, please retest with current HEAD. > > regards, tom lane > -- Koichi Suzuki
I don't fully understand what "transaction log" means. If it means "archived WAL", the current (8.2) code handle WAL as follows: 1) If full_page_writes=off, then no full page writes will be written to WAL, except for those during onlie backup (between pg_start_backup and pg_stop_backup). The WAL size will be considerably small but it cannot recover from partial/inconsistent write to the database files. We have to go back to the online backup and apply all the archive log. 2) If full_page_writes=on, then full page writes will be written at the first update of a page after each checkpoint, plus full page writes at 1). Because we have no means (in 8.2) to optimize the WAL so far, what we can do is to copy WAL or gzip it at archive time. If we'd like to keep good chance of recovery after the crash, 8.2 provides only the method 2), leaving archive log size considerably large. My proposal maintains the chance of crash recovery the same as in the case of full_page_writes=on and reduces the size of archived log as in the case of full_page_writes=off. Regards; Hannu Krosing wrote: > Ühel kenal päeval, T, 2007-04-10 kell 18:17, kirjutas Joshua D. Drake: >>> In terms of idle time for gzip and other command to archive WAL offline, >>> no difference in the environment was given other than the command to >>> archive. My guess is because the user time is very large in gzip, it >>> has more chance for scheduler to give resource to other processes. In >>> the case of cp, idle time is more than 30times longer than user time. >>> Pg_compresslog uses seven times longer idle time than user time. On the >>> other hand, gzip uses less idle time than user time. Considering the >>> total amount of user time, I think it's reasonable measure. >>> >>> Again, in my proposal, it is not the issue to increase run time >>> performance. Issue is to decrease the size of archive log to save the >>> storage. >> Considering the relatively little amount of storage a transaction log >> takes, it would seem to me that the performance angle is more appropriate. > > As I understand it it's not about transaction log but about write-ahead > log. > > and the amount of data in WAL can become very important once you have to > keep standby servers in different physical locations (cities, countries > or continents) where channel throughput and cost comes into play. > > With simple cp (scp/rsync) the amount of WAL data needing to be copied > is about 10x more than data collected by trigger based solutions > (Slony/pgQ). With pg_compresslog WAL-shipping seems to have roughly the > same amount and thus becomes a viable alternative again. > >> Is it more efficient in other ways besides negligible tps? Possibly more >> efficient memory usage? Better restore times for a crashed system? > > I think that TPS is more affected by number of writes than size of each > block written, so there is probably not that much to gain in TPS, except > perhaps from better disk cache usage. > > For me pg_compresslog seems to be a winner even if it just does not > degrade performance. > -- Koichi Suzuki
> I don't fully understand what "transaction log" means. If it means > "archived WAL", the current (8.2) code handle WAL as follows: Probably we can define "transaction log" to be the part of WAL that is not full pages. > 1) If full_page_writes=off, then no full page writes will be > written to WAL, except for those during onlie backup (between > pg_start_backup and > pg_stop_backup). The WAL size will be considerably small > but it cannot > recover from partial/inconsistent write to the database > files. We have to go back to the online backup and apply all > the archive log. > > 2) If full_page_writes=on, then full page writes will be > written at the first update of a page after each checkpoint, > plus full page writes at > 1). Because we have no means (in 8.2) to optimize the WAL > so far, what > we can do is to copy WAL or gzip it at archive time. > > If we'd like to keep good chance of recovery after the crash, > 8.2 provides only the method 2), leaving archive log size > considerably large. My proposal maintains the chance of > crash recovery the same as in the case of full_page_writes=on > and reduces the size of archived log as in the case of > full_page_writes=off. Yup, this is a good summary. You say you need to remove the optimization that avoids the logging of a new tuple because the full page image exists. I think we must already have the info in WAL which tuple inside the full page image is new (the one for which we avoided the WAL entry for). How about this: Leave current WAL as it is and only add the not removeable flag to full pages. pg_compresslog then replaces the full page image with a record for the one tuple that is changed. I tend to think it is not worth the increased complexity only to save bytes in the uncompressed WAL though. Another point about pg_decompresslog: Why do you need a pg_decompresslog ? Imho pg_compresslog should already do the replacing of the full_page with the dummy entry. Then pg_decompresslog could be a simple gunzip, or whatever compression was used, but no logic. Andreas
Hi, Sorry, inline reply. Zeugswetter Andreas ADI SD wrote: > > Yup, this is a good summary. > > You say you need to remove the optimization that avoids > the logging of a new tuple because the full page image exists. > I think we must already have the info in WAL which tuple inside the full > page image > is new (the one for which we avoided the WAL entry for). > > How about this: > Leave current WAL as it is and only add the not removeable flag to full > pages. > pg_compresslog then replaces the full page image with a record for the > one tuple that is changed. > I tend to think it is not worth the increased complexity only to save > bytes in the uncompressed WAL though. It is essentially what my patch proposes. My patch includes flag to full page writes which "can be" removed. > Another point about pg_decompresslog: > > Why do you need a pg_decompresslog ? Imho pg_compresslog should already > do the replacing of the > full_page with the dummy entry. Then pg_decompresslog could be a simple > gunzip, or whatever compression was used, > but no logic. Just removing full page writes does not work. If we shift the rest of the WAL, then LSN becomes inconsistent in compressed archive logs which pg_compresslog produces. For recovery, we have to restore LSN as the original WAL. Pg_decompresslog restores removed full page writes as a dumm records so that recovery redo functions won't be confused. Regards; > > Andreas > -- ------------- Koichi Suzuki
> > Yup, this is a good summary. > > > > You say you need to remove the optimization that avoids the logging of > > a new tuple because the full page image exists. > > I think we must already have the info in WAL which tuple inside the > > full page image is new (the one for which we avoided the WAL entry > > for). > > > > How about this: > > Leave current WAL as it is and only add the not removeable flag to > > full pages. > > pg_compresslog then replaces the full page image with a record for the > > one tuple that is changed. > > I tend to think it is not worth the increased complexity only to save > > bytes in the uncompressed WAL though. > > It is essentially what my patch proposes. My patch includes > flag to full page writes which "can be" removed. Ok, a flag that marks full page images that can be removed is perfect. But you also turn off the optimization that avoids writing regular WAL records when the info is already contained in a full-page image (increasing the uncompressed size of WAL). It was that part I questioned. As already stated, maybe I should not have because it would be too complex to reconstruct a regular WAL record from the full-page image. But that code would also be needed for WAL based partial replication, so if it where too complicated we would eventually want a switch to turn off the optimization anyway (at least for heap page changes). > > Another point about pg_decompresslog: > > > > Why do you need a pg_decompresslog ? Imho pg_compresslog should > > already do the replacing of the full_page with the dummy entry. Then > > pg_decompresslog could be a simple gunzip, or whatever compression was > > used, but no logic. > > Just removing full page writes does not work. If we shift the rest of > the WAL, then LSN becomes inconsistent in compressed archive logs which > pg_compresslog produces. For recovery, we have to restore LSN as the > original WAL. Pg_decompresslog restores removed full page writes as a > dumm records so that recovery redo functions won't be confused. Ah sorry, I needed some pgsql/src/backend/access/transam/README reading. LSN is the physical position of records in WAL. Thus your dummy record size is equal to what you cut out of the original record. What about disconnecting WAL LSN from physical WAL record position during replay ? Add simple short WAL records in pg_compresslog like: advance LSN by 8192 bytes. Andreas
"Zeugswetter Andreas ADI SD" <ZeugswetterA@spardat.at> writes:
> But you also turn off the optimization that avoids writing regular
> WAL records when the info is already contained in a full-page image
> (increasing the uncompressed size of WAL).
> It was that part I questioned.
That's what bothers me about this patch, too. It will be increasing
the cost of writing WAL (more data -> more CRC computation and more
I/O, not to mention more contention for the WAL locks) which translates
directly to a server slowdown.
The main arguments that I could see against Andreas' alternative are:
1. Some WAL record types are arranged in a way that actually would not
permit the reconstruction of the short form from the long form, because
they throw away too much data when the full-page image is substituted.
An example that's fresh in my mind is that the current format of the
btree page split WAL record discards newitemoff in that case, so you
couldn't identify the inserted item in the page image. Now this is only
saving two bytes in what's usually going to be a darn large record
anyway, and it complicates the code to do it, so I wouldn't cry if we
changed btree split to include newitemoff always. But there might be
some other cases where more data is involved. In any case, someone
would have to look through every single WAL record type to determine
whether reconstruction is possible and fix it if not.
2. The compresslog utility would have to have specific knowledge about
every compressible WAL record type, to know how to convert it to the
short format. That means an ongoing maintenance commitment there.
I don't think this is unacceptable, simply because we need only teach
it about a few common record types, not everything under the sun ---
anything it doesn't know how to fix, just leave alone, and if it's an
uncommon record type it really doesn't matter. (I guess that means
that we don't really have to do #1 for every last record type, either.)
So I don't think either of these is a showstopper. Doing it this way
would certainly make the patch more acceptable, since the argument that
it might hurt rather than help performance in some cases would go away.
> What about disconnecting WAL LSN from physical WAL record position
> during replay ?
> Add simple short WAL records in pg_compresslog like: advance LSN by 8192
> bytes.
I don't care for that, as it pretty much destroys some of the more
important sanity checks that xlog replay does. The page boundaries
need to match the records contained in them. So I think we do need
to have pg_decompresslog insert dummy WAL entries to fill up the
space saved by omitting full pages.
regards, tom lane
On Fri, 2007-04-13 at 10:36 -0400, Tom Lane wrote: > "Zeugswetter Andreas ADI SD" <ZeugswetterA@spardat.at> writes: > > But you also turn off the optimization that avoids writing regular > > WAL records when the info is already contained in a full-page image > > (increasing the uncompressed size of WAL). > > It was that part I questioned. I think its right to question it, certainly. > That's what bothers me about this patch, too. It will be increasing > the cost of writing WAL (more data -> more CRC computation and more > I/O, not to mention more contention for the WAL locks) which translates > directly to a server slowdown. I don't really understand this concern. Koichi-san has included a parameter setting that would prevent any change at all in the way WAL is written. If you don't want this slight increase in WAL, don't enable it. If you do enable it, you'll also presumably be compressing the xlog too, which works much better than gzip using less CPU. So overall it saves more than it costs, ISTM, and nothing at all if you choose not to use it. > The main arguments that I could see against Andreas' alternative are: > > 1. Some WAL record types are arranged in a way that actually would not > permit the reconstruction of the short form from the long form, because > they throw away too much data when the full-page image is substituted. > An example that's fresh in my mind is that the current format of the > btree page split WAL record discards newitemoff in that case, so you > couldn't identify the inserted item in the page image. Now this is only > saving two bytes in what's usually going to be a darn large record > anyway, and it complicates the code to do it, so I wouldn't cry if we > changed btree split to include newitemoff always. But there might be > some other cases where more data is involved. In any case, someone > would have to look through every single WAL record type to determine > whether reconstruction is possible and fix it if not. > > 2. The compresslog utility would have to have specific knowledge about > every compressible WAL record type, to know how to convert it to the > short format. That means an ongoing maintenance commitment there. > I don't think this is unacceptable, simply because we need only teach > it about a few common record types, not everything under the sun --- > anything it doesn't know how to fix, just leave alone, and if it's an > uncommon record type it really doesn't matter. (I guess that means > that we don't really have to do #1 for every last record type, either.) > > So I don't think either of these is a showstopper. Doing it this way > would certainly make the patch more acceptable, since the argument that > it might hurt rather than help performance in some cases would go away. Yeh, its additional code paths, but it sounds like Koichi-san and colleagues are going to be trail blazing any bugs there and will be around to fix any more that emerge. > > What about disconnecting WAL LSN from physical WAL record position > > during replay ? > > Add simple short WAL records in pg_compresslog like: advance LSN by 8192 > > bytes. > > I don't care for that, as it pretty much destroys some of the more > important sanity checks that xlog replay does. The page boundaries > need to match the records contained in them. So I think we do need > to have pg_decompresslog insert dummy WAL entries to fill up the > space saved by omitting full pages. Agreed. I don't want to start touching something that works so well. We've been thinking about doing this for at least 3 years now, so I don't see any reason to baulk at it now. I'm happy with Koichi-san's patch as-is, assuming further extensive testing will be carried out on it during beta. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
"Simon Riggs" <simon@2ndquadrant.com> writes:
> On Fri, 2007-04-13 at 10:36 -0400, Tom Lane wrote:
>> That's what bothers me about this patch, too. It will be increasing
>> the cost of writing WAL (more data -> more CRC computation and more
>> I/O, not to mention more contention for the WAL locks) which translates
>> directly to a server slowdown.
> I don't really understand this concern.
The real objection is that a patch that's alleged to make WAL smaller
actually does the exact opposite. Now maybe you can buy that back
downstream of the archiver --- after yet more added-on processing ---
but it still seems that there's a fundamental misdesign here.
> Koichi-san has included a parameter setting that would prevent any
> change at all in the way WAL is written.
It bothers me that we'd need to have such a switch. That's just another
way to shoot yourself in the foot, either by not enabling it (in which
case applying pg_compresslog as it stands would actively break your
WAL), or by enabling it when you weren't actually going to use
pg_compresslog (because you misunderstood the documentation to imply
that it'd make your WAL smaller by itself). What I want to see is a
patch that doesn't bloat WAL at all and therefore doesn't need a switch.
I think Andreas is correct to complain that it should be done that way.
regards, tom lane
On Fri, 2007-04-13 at 11:47 -0400, Tom Lane wrote: > "Simon Riggs" <simon@2ndquadrant.com> writes: > > On Fri, 2007-04-13 at 10:36 -0400, Tom Lane wrote: > >> That's what bothers me about this patch, too. It will be increasing > >> the cost of writing WAL (more data -> more CRC computation and more > >> I/O, not to mention more contention for the WAL locks) which translates > >> directly to a server slowdown. > > > I don't really understand this concern. > > The real objection is that a patch that's alleged to make WAL smaller > actually does the exact opposite. Now maybe you can buy that back > downstream of the archiver --- after yet more added-on processing --- > but it still seems that there's a fundamental misdesign here. > > > Koichi-san has included a parameter setting that would prevent any > > change at all in the way WAL is written. > > It bothers me that we'd need to have such a switch. That's just another > way to shoot yourself in the foot, either by not enabling it (in which > case applying pg_compresslog as it stands would actively break your > WAL), or by enabling it when you weren't actually going to use > pg_compresslog (because you misunderstood the documentation to imply > that it'd make your WAL smaller by itself). What I want to see is a > patch that doesn't bloat WAL at all and therefore doesn't need a switch. > I think Andreas is correct to complain that it should be done that way. I agree with everything you say because we already had *exactly* this discussion when the patch was already submitted, with me saying everything you just said. After a few things have been renamed to show their correct function and impact, I am now comfortable with this patch. Writing lots of additional code simply to remove a parameter that *might* be mis-interpreted doesn't sound useful to me, especially when bugs may leak in that way. My take is that this is simple and useful *and* we have it now; other ways don't yet exist, nor will they in time for 8.3. -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
"Simon Riggs" <simon@2ndquadrant.com> writes:
> Writing lots of additional code simply to remove a parameter that
> *might* be mis-interpreted doesn't sound useful to me, especially when
> bugs may leak in that way. My take is that this is simple and useful
> *and* we have it now; other ways don't yet exist, nor will they in time
> for 8.3.
The potential for misusing the switch is only one small part of the
argument; the larger part is that this has been done in the wrong way
and will cost performance unnecessarily. The fact that it's ready
now is not something that I think should drive our choices.
I believe that it would be possible to make the needed core-server
changes in time for 8.3, and then to work on compress/decompress
on its own time scale and publish it on pgfoundry; with the hope
that it would be merged to contrib or core in 8.4. Frankly the
compress/decompress code needs work anyway before it could be
merged (eg, I noted a distinct lack of I/O error checking).
regards, tom lane
Sorry I was very late to find this. With DBT-2 benchmark, I've already compared the amount of WAL. The result was as follows: Amount of WAL after 60min. run of DBT-2 benchmark wal_add_optimization_info = off (default) 3.13GB wal_add_optimization_info = on (new case) 3.17GB -> can be optimized to 0.31GB by pg_compresslog. So the difference will be around a couple of percents. I think this is very good figure. For information, DB Size: 12.35GB (120WH) Checkpoint timeout: 60min. Checkpoint occured only once in the run. ---------------- I don't think replacing LSN works fine. For full recovery to the current time, we need both archive log and WAL. Replacing LSN will make archive log LSN inconsistent with WAL's LSN and the recovery will not work. Reconstruction to regular WAL is proposed as pg_decompresslog. We should be careful enough not to make redo routines confused with the dummy full page writes, as Simon suggested. So far, it works fine. Regards; Zeugswetter Andreas ADI SD wrote: >>> Yup, this is a good summary. >>> >>> You say you need to remove the optimization that avoids the logging > of >>> a new tuple because the full page image exists. >>> I think we must already have the info in WAL which tuple inside the >>> full page image is new (the one for which we avoided the WAL entry >>> for). >>> >>> How about this: >>> Leave current WAL as it is and only add the not removeable flag to >>> full pages. >>> pg_compresslog then replaces the full page image with a record for > the >>> one tuple that is changed. >>> I tend to think it is not worth the increased complexity only to > save >>> bytes in the uncompressed WAL though. >> It is essentially what my patch proposes. My patch includes >> flag to full page writes which "can be" removed. > > Ok, a flag that marks full page images that can be removed is perfect. > > But you also turn off the optimization that avoids writing regular > WAL records when the info is already contained in a full-page image > (increasing the > uncompressed size of WAL). > It was that part I questioned. As already stated, maybe I should not > have because > it would be too complex to reconstruct a regular WAL record from the > full-page image. > But that code would also be needed for WAL based partial replication, so > if it where too > complicated we would eventually want a switch to turn off the > optimization anyway > (at least for heap page changes). > >>> Another point about pg_decompresslog: >>> >>> Why do you need a pg_decompresslog ? Imho pg_compresslog should >>> already do the replacing of the full_page with the dummy entry. Then > >>> pg_decompresslog could be a simple gunzip, or whatever compression > was >>> used, but no logic. >> Just removing full page writes does not work. If we shift the rest > of >> the WAL, then LSN becomes inconsistent in compressed archive logs > which >> pg_compresslog produces. For recovery, we have to restore LSN as the > >> original WAL. Pg_decompresslog restores removed full page writes as > a >> dumm records so that recovery redo functions won't be confused. > > Ah sorry, I needed some pgsql/src/backend/access/transam/README reading. > > LSN is the physical position of records in WAL. Thus your dummy record > size is equal to what you cut out of the original record. > What about disconnecting WAL LSN from physical WAL record position > during replay ? > Add simple short WAL records in pg_compresslog like: advance LSN by 8192 > bytes. > > Andreas > -- ------------- Koichi Suzuki
> With DBT-2 benchmark, I've already compared the amount of WAL. The > result was as follows: > > Amount of WAL after 60min. run of DBT-2 benchmark > wal_add_optimization_info = off (default) 3.13GB how about wal_fullpage_optimization = on (default) > wal_add_optimization_info = on (new case) 3.17GB -> can be > optimized to 0.31GB by pg_compresslog. > > So the difference will be around a couple of percents. I think this is > very good figure. > > For information, > DB Size: 12.35GB (120WH) > Checkpoint timeout: 60min. Checkpoint occured only once in the run. Unfortunately I think DBT-2 is not a good benchmark to test the disabled wal optimization. The test should contain some larger rows (maybe some updates on large toasted values), and maybe more frequent checkpoints. Actually the poor ratio between full pages and normal WAL content in this benchmark is strange to begin with. Tom fixed a bug recently, and it would be nice to see the new ratio. Have you read Tom's comment on not really having to be able to reconstruct all record types from the full page image ? I think that sounded very promising (e.g. start out with only heap insert/update). Then: - we would not need the wal optimization switch (the full page flag would always be added depending only on backup) - pg_compresslog would only remove such "full page" images where it knows how to reconstruct a "normal" WAL record from - with time and effort pg_compresslog would be able to compress [nearly] all record types's full images (no change in backend) > I don't think replacing LSN works fine. For full recovery to > the current time, we need both archive log and WAL. > Replacing LSN will make archive log LSN inconsistent with > WAL's LSN and the recovery will not work. WAL recovery would have had to be modified (decouple LSN from WAL position during recovery). An "archive log" would have been a valid WAL (with appropriate LSN advance records). > Reconstruction to regular WAL is proposed as > pg_decompresslog. We should be careful enough not to make > redo routines confused with the dummy full page writes, as > Simon suggested. So far, it works fine. Yes, Tom didn't like "LSN replacing" eighter. I withdraw my concern regarding pg_decompresslog. Your work in this area is extremely valuable and I hope my comments are not discouraging. Thank you Andreas
Hi, I agree that pg_compresslog should be aware of all the WAL records' details so that it can optimize archive log safely. In my patch, I've examined 8.2's WAL records to make pg_compresslog/pg_decompresslog safe. Also I agree further pg_compresslog maintenance needs to examine changes in WAL record format. Because the number of such format will be limited, I think the amount of work will be reasonable enough. Regards; Simon Riggs wrote: > On Fri, 2007-04-13 at 10:36 -0400, Tom Lane wrote: >> "Zeugswetter Andreas ADI SD" <ZeugswetterA@spardat.at> writes: >>> But you also turn off the optimization that avoids writing regular >>> WAL records when the info is already contained in a full-page image >>> (increasing the uncompressed size of WAL). >>> It was that part I questioned. > > I think its right to question it, certainly. > >> That's what bothers me about this patch, too. It will be increasing >> the cost of writing WAL (more data -> more CRC computation and more >> I/O, not to mention more contention for the WAL locks) which translates >> directly to a server slowdown. > > I don't really understand this concern. Koichi-san has included a > parameter setting that would prevent any change at all in the way WAL is > written. If you don't want this slight increase in WAL, don't enable it. > If you do enable it, you'll also presumably be compressing the xlog too, > which works much better than gzip using less CPU. So overall it saves > more than it costs, ISTM, and nothing at all if you choose not to use > it. > >> The main arguments that I could see against Andreas' alternative are: >> >> 1. Some WAL record types are arranged in a way that actually would not >> permit the reconstruction of the short form from the long form, because >> they throw away too much data when the full-page image is substituted. >> An example that's fresh in my mind is that the current format of the >> btree page split WAL record discards newitemoff in that case, so you >> couldn't identify the inserted item in the page image. Now this is only >> saving two bytes in what's usually going to be a darn large record >> anyway, and it complicates the code to do it, so I wouldn't cry if we >> changed btree split to include newitemoff always. But there might be >> some other cases where more data is involved. In any case, someone >> would have to look through every single WAL record type to determine >> whether reconstruction is possible and fix it if not. >> >> 2. The compresslog utility would have to have specific knowledge about >> every compressible WAL record type, to know how to convert it to the >> short format. That means an ongoing maintenance commitment there. >> I don't think this is unacceptable, simply because we need only teach >> it about a few common record types, not everything under the sun --- >> anything it doesn't know how to fix, just leave alone, and if it's an >> uncommon record type it really doesn't matter. (I guess that means >> that we don't really have to do #1 for every last record type, either.) >> >> So I don't think either of these is a showstopper. Doing it this way >> would certainly make the patch more acceptable, since the argument that >> it might hurt rather than help performance in some cases would go away. > > Yeh, its additional code paths, but it sounds like Koichi-san and > colleagues are going to be trail blazing any bugs there and will be > around to fix any more that emerge. > >>> What about disconnecting WAL LSN from physical WAL record position >>> during replay ? >>> Add simple short WAL records in pg_compresslog like: advance LSN by 8192 >>> bytes. >> I don't care for that, as it pretty much destroys some of the more >> important sanity checks that xlog replay does. The page boundaries >> need to match the records contained in them. So I think we do need >> to have pg_decompresslog insert dummy WAL entries to fill up the >> space saved by omitting full pages. > > Agreed. I don't want to start touching something that works so well. > > > We've been thinking about doing this for at least 3 years now, so I > don't see any reason to baulk at it now. I'm happy with Koichi-san's > patch as-is, assuming further extensive testing will be carried out on > it during beta. > -- ------------- Koichi Suzuki
Here's only a part of the reply I should do, but as to I/O error
checking ...
Here's a list of system calls and other external function/library calls
used in pg_lesslog patch series, together with how current patch checks
each errors and how current postgresql source handles the similar calls:
--------------------------------
1. No error check is done
1-1. fileno()
fileno() is called against stdin and stdout from pg_compresslog.c
and pg_decompresslog.c. They are intended to be invoked from a shell
and so stdin and stdout are both available. fileno() error occurs only
if invoker of pg_compresslog or pg_decompresslog closes stdin and/or
stdout before the invoker executes them. I found similar fileno()
usage in pg_dump/pg_backup_archive.c and postmaster/syslogger.c. I
don't think this is an issue.
1-2. fflush()
fflush() is called against stdout within a debug routine, debug.c.
Such usage can also be found in bin/initdb.c, bin/scripts/createdb.c,
bin/psql/common.c and more. I don't think this is an issue either.
1-3. printf() and fprintf()
It is common practice not to check the error. We can find such
calls in many of existing source codes.
1-4. strerror()
It is checked that system call returns error before calling
strerror. Similar code can be found in other PostgreSQL source too.
----------------------------------
2. Error check is done
All the following function calls are associated with return value check.
open(), close(), fstat(), read(), write()
-----------------------------------
3. Functions do not return error
The following functin will not return errors, so no error check is needed.
exit(), memcopy(), memset(), strcmp()
------------------------------------
I hope this helps.
Regards;
Tom Lane wrote:
> "Simon Riggs" <simon@2ndquadrant.com> writes:
>> Writing lots of additional code simply to remove a parameter that
>> *might* be mis-interpreted doesn't sound useful to me, especially when
>> bugs may leak in that way. My take is that this is simple and useful
>> *and* we have it now; other ways don't yet exist, nor will they in time
>> for 8.3.
>
> The potential for misusing the switch is only one small part of the
> argument; the larger part is that this has been done in the wrong way
> and will cost performance unnecessarily. The fact that it's ready
> now is not something that I think should drive our choices.
>
> I believe that it would be possible to make the needed core-server
> changes in time for 8.3, and then to work on compress/decompress
> on its own time scale and publish it on pgfoundry; with the hope
> that it would be merged to contrib or core in 8.4. Frankly the
> compress/decompress code needs work anyway before it could be
> merged (eg, I noted a distinct lack of I/O error checking).
>
> regards, tom lane
>
--
-------------
Koichi Suzuki
On Fri, 2007-04-20 at 10:16 +0200, Zeugswetter Andreas ADI SD wrote: > Your work in this area is extremely valuable and I hope my comments are > not discouraging. I think its too late in the day to make the changes suggested by yourself and Tom. They make the patch more invasive and more likely to error, plus we don't have much time. This really means the patch is likely to be rejected at the 11th hour when what we have essentially works... -- Simon Riggs EnterpriseDB http://www.enterprisedb.com
Hi, I don't insist the name and the default of the GUC parameter. I'm afraid wal_fullpage_optimization = on (default) makes some confusion because the default behavior becomes a bit different on WAL itself. I'd like to have some more opinion on this. Zeugswetter Andreas ADI SD wrote: >> With DBT-2 benchmark, I've already compared the amount of WAL. The >> result was as follows: >> >> Amount of WAL after 60min. run of DBT-2 benchmark >> wal_add_optimization_info = off (default) 3.13GB > > how about wal_fullpage_optimization = on (default) > >> wal_add_optimization_info = on (new case) 3.17GB -> can be >> optimized to 0.31GB by pg_compresslog. >> >> So the difference will be around a couple of percents. I think this > is >> very good figure. >> >> For information, >> DB Size: 12.35GB (120WH) >> Checkpoint timeout: 60min. Checkpoint occured only once in the run. > > Unfortunately I think DBT-2 is not a good benchmark to test the disabled > wal optimization. > The test should contain some larger rows (maybe some updates on large > toasted values), and maybe more frequent checkpoints. Actually the poor > ratio between full pages and normal WAL content in this benchmark is > strange to begin with. > Tom fixed a bug recently, and it would be nice to see the new ratio. > > Have you read Tom's comment on not really having to be able to > reconstruct all record types from the full page image ? I think that > sounded very promising (e.g. start out with only heap insert/update). > > Then: > - we would not need the wal optimization switch (the full page flag > would always be added depending only on backup) > - pg_compresslog would only remove such "full page" images where it > knows how to reconstruct a "normal" WAL record from > - with time and effort pg_compresslog would be able to compress [nearly] > all record types's full images (no change in backend) > >> I don't think replacing LSN works fine. For full recovery to >> the current time, we need both archive log and WAL. >> Replacing LSN will make archive log LSN inconsistent with >> WAL's LSN and the recovery will not work. > > WAL recovery would have had to be modified (decouple LSN from WAL > position during recovery). > An "archive log" would have been a valid WAL (with appropriate LSN > advance records). > >> Reconstruction to regular WAL is proposed as >> pg_decompresslog. We should be careful enough not to make >> redo routines confused with the dummy full page writes, as >> Simon suggested. So far, it works fine. > > Yes, Tom didn't like "LSN replacing" eighter. I withdraw my concern > regarding pg_decompresslog. > > Your work in this area is extremely valuable and I hope my comments are > not discouraging. > > Thank you > Andreas > -- ------------- Koichi Suzuki
> I don't insist the name and the default of the GUC parameter. > I'm afraid wal_fullpage_optimization = on (default) makes > some confusion because the default behavior becomes a bit > different on WAL itself. Seems my wal_fullpage_optimization is not a good name if it caused misinterpretation already :-( > >> Amount of WAL after 60min. run of DBT-2 benchmark > >> wal_add_optimization_info = off (default) 3.13GB > > > > how about wal_fullpage_optimization = on (default) The meaning of wal_fullpage_optimization = on (default) would be the same as your wal_add_optimization_info = off (default). (Reversed name, reversed meaning of the boolean value) It would be there to *turn off* the (default) WAL full_page optimization. For your pg_compresslog it would need to be set to off. "add_optimization_info" sounded like added info about/for some optimization which it is not. We turn off an optimization with the flag for the benefit of an easier pg_compresslog implementation. As already said I would decouple this setting from the part that sets the "removeable full page" flag in WAL, and making the recovery able to skip dummy records. This I would do unconditionally. Andreas
Hackers, > Writing lots of additional code simply to remove a parameter that > *might* be mis-interpreted doesn't sound useful to me, especially when > bugs may leak in that way. My take is that this is simple and useful > *and* we have it now; other ways don't yet exist, nor will they in time > for 8.3. How about naming the parameter wal_compressable? That would indicate pretty clearly that the parameter is intended to be used with wal_compress and nothing else. However, I do agree with Andreas that anything which adds to WAL volume, even 3%, seems like going in the wrong direction. We already have higher log output than any comparable database (higher than InnoDB by 3x) and we should be looking for output to trim as well as compression. So the relevant question is whether the patch in its current form provides enough benefit to make it worthwhile for 8.3, or whether we should wait for 8.4. Questions: 1) is there any throughput benefit for platforms with fast CPU but contrained I/O (e.g. 2-drive webservers)? Any penalty for servers with plentiful I/O? 2) Will this patch make attempts to reduce WAL volume in the future significantly harder? 3) How is this better than command-line compression for log-shipping? e.g. why do we need it in the database? -- Josh Berkus PostgreSQL @ Sun San Francisco
Hi,
Sorry, because of so many comments/questions, I'll write inline....
Josh Berkus wrote:
> Hackers,
>
>> Writing lots of additional code simply to remove a parameter that
>> *might* be mis-interpreted doesn't sound useful to me, especially when
>> bugs may leak in that way. My take is that this is simple and useful
>> *and* we have it now; other ways don't yet exist, nor will they in time
>> for 8.3.
>
> How about naming the parameter wal_compressable? That would indicate pretty
> clearly that the parameter is intended to be used with wal_compress and
> nothing else.
Hmm, it sounds nicer.
>
> However, I do agree with Andreas that anything which adds to WAL volume, even
> 3%, seems like going in the wrong direction. We already have higher log
> output than any comparable database (higher than InnoDB by 3x) and we should
> be looking for output to trim as well as compression.
>
> So the relevant question is whether the patch in its current form provides
> enough benefit to make it worthwhile for 8.3, or whether we should wait for
> 8.4. Questions:
>
Before answering questions below, I'd like to say that archive log
optimization has to be address different point of views to the current
(upto 8.2) settings.
1) To deal with partial/inconsisitent write to the data file at crash
recovery, we need full page writes at the first modification to pages
after each checkpoint. It consumes much of WAL space.
2) 1) is not necessary for archive recovery (PITR) and full page writes
can be removed for this purpose. However, we need full page writes
during hot backup to deal with partial writes by backup commands. This
is implemented in 8.2.
3) To maintain crash recovery chance and reduce the amount of archive
log, removal of unnecessary full page writes from archive logs is a
good choice. To do this, we need both logical log and full page writes
in WAL.
I don't think there should be only one setting. It depend on how
database is operated. Leaving wal_add_optiomization_info = off default
does not bring any change in WAL and archive log handling. I
understand some people may not be happy with additional 3% or so
increase in WAL size, especially people who dosn't need archive log at
all. So I prefer to leave the default off.
For users, I think this is simple enough:
1) For people happy with 8.2 settings:
No change is needed to move to 8.3 and there's really no change.
2) For people who need to reduce archive log size but like to leave full
page writes to WAL (to maintain crash recovery chance):
a) Add GUC parameter: wal_add_optiomization_info=on
b) Change archive command from "cp" to "pg_compresslog"
c) Change restore command from "cp" to "pg_decompresslog"
Archive log can be stored and restored as done in older releases.
> 1) is there any throughput benefit for platforms with fast CPU but contrained
> I/O (e.g. 2-drive webservers)? Any penalty for servers with plentiful I/O?
I've only run benchmarks with archive process running, because
wal_add_optimization_info=on does not make sense if we don't archive
WAL. In this situation, total I/O decreases because writes to archive
log decreases. Because of 3% or so increase in WAL size, there will be
increase in WAL write, but decrease in archive writes makes it up.
>
> 2) Will this patch make attempts to reduce WAL volume in the future
> significantly harder?
Yes, I'd like to continue to work to reduce the WAL size. It's still
an issue when database size becomes several handreds of gigabytes in
size. Anyway, I think WAL size reduction has to be done in
XLogInsert() or XLogWrite(). We need much more discussion for this.
The issue will be how to maintain crash recovery chance by inconsistent
writes (by full_page_writes=off, we have to give it up). On the other
hand we have to keep examining each WAL record.
>
> 3) How is this better than command-line compression for log-shipping? e.g.
> why do we need it in the database?
I don't fully understand what command-line compression means. Simon
suggested that this patch can be used with log-shipping and I agree.
If we compare compression with gzip or other general purpose
compression, compression ratio, CPU usage and I/O by pg_compresslog are
all quite better than those in gzip.
Please let me know if you intended defferently.
Regards;
--
-------------
Koichi Suzuki
> 3) To maintain crash recovery chance and reduce the amount of > archive log, removal of unnecessary full page writes from > archive logs is a good choice. Definitely, yes. pg_compresslog could even move the full pages written during backup out of WAL and put them in a different file that needs to be applied before replay of the corresponding WAL after a physical restore. This would further help reduce log shipping volume. > To do this, we need both logical log and full page writes in WAL. This is only true in the sense, that it allows a less complex implementation of pg_compresslog. Basically a WAL record consists of info about what happened and currently eighter per tuple new data or a full page image. The info of "what happened" together with the full page image is sufficient to reconstruct the "per tuple new data". There might be a few WAL record types (e.g. in btree split ?) where this is not so, but we could eighter fix those or not compress those. This is why I don't like Josh's suggested name of wal_compressable eighter. WAL is compressable eighter way, only pg_compresslog would need to be more complex if you don't turn off the full page optimization. I think a good name would tell that you are turning off an optimization. (thus my wal_fullpage_optimization on/off) Andreas
Koichi, Andreas, > 1) To deal with partial/inconsisitent write to the data file at crash > recovery, we need full page writes at the first modification to pages > after each checkpoint. It consumes much of WAL space. We need to find a way around this someday. Other DBs don't do this; it may be becuase they're less durable, or because they fixed the problem. > I don't think there should be only one setting. It depend on how > database is operated. Leaving wal_add_optiomization_info = off default > does not bring any change in WAL and archive log handling. I > understand some people may not be happy with additional 3% or so > increase in WAL size, especially people who dosn't need archive log at > all. So I prefer to leave the default off. Except that, is there any reason to turn this off if we are archiving? Maybe it should just be slaved to archive_command ... if we're not using PITR, it's off, if we are, it's on. > > 1) is there any throughput benefit for platforms with fast CPU but > > contrained I/O (e.g. 2-drive webservers)? Any penalty for servers with > > plentiful I/O? > > I've only run benchmarks with archive process running, because > wal_add_optimization_info=on does not make sense if we don't archive > WAL. In this situation, total I/O decreases because writes to archive > log decreases. Because of 3% or so increase in WAL size, there will be > increase in WAL write, but decrease in archive writes makes it up. Yeah, I was just looking for a way to make this a performance feature. I see now that it can't be. ;-) > > 3) How is this better than command-line compression for log-shipping? > > e.g. why do we need it in the database? > > I don't fully understand what command-line compression means. Simon > suggested that this patch can be used with log-shipping and I agree. > If we compare compression with gzip or other general purpose > compression, compression ratio, CPU usage and I/O by pg_compresslog are > all quite better than those in gzip. OK, that answered my question. > This is why I don't like Josh's suggested name of wal_compressable > eighter. > WAL is compressable eighter way, only pg_compresslog would need to be > more complex if you don't turn off the full page optimization. I think a > good name would tell that you are turning off an optimization. > (thus my wal_fullpage_optimization on/off) Well, as a PG hacker I find the name wal_fullpage_optimization quite baffling and I think our general user base will find it even more so. Now that I have Koichi's explanation of the problem, I vote for simply slaving this to the PITR settings and not having a separate option at all. -- Josh Berkus PostgreSQL @ Sun San Francisco
Josh Berkus <josh@agliodbs.com> writes:
> Well, as a PG hacker I find the name wal_fullpage_optimization quite
> baffling and I think our general user base will find it even more so.
> Now that I have Koichi's explanation of the problem, I vote for simply
> slaving this to the PITR settings and not having a separate option at
> all.
The way to not have a separate option is to not need one, by having the
feature not cost anything extra in the first place. Andreas and I have
made the point repeatedly about how to do that.
regards, tom lane
> > 1) To deal with partial/inconsisitent write to the data file at crash > > recovery, we need full page writes at the first modification to pages > > after each checkpoint. It consumes much of WAL space. > > We need to find a way around this someday. Other DBs don't > do this; it may be becuase they're less durable, or because > they fixed the problem. They eighter can only detect a failure later (this may be a very long time depending on access and verify runs) or they also write page images. Those that write page images usually write "before images" to a different area that is cleared periodically (e.g. during checkpoint). Writing to a different area was considered in pg, but there were more negative issues than positive. So imho pg_compresslog is the correct path forward. The current discussion is only about whether we want a more complex pg_compresslog and no change to current WAL, or an increased WAL size for a less complex implementation. Both would be able to compress the WAL to the same "archive log" size. Andreas
On Wed, Apr 25, 2007 at 10:00:16AM +0200, Zeugswetter Andreas ADI SD wrote: > > > > 1) To deal with partial/inconsisitent write to the data file at > crash > > > recovery, we need full page writes at the first modification to > pages > > > after each checkpoint. It consumes much of WAL space. > > > > We need to find a way around this someday. Other DBs don't > > do this; it may be becuase they're less durable, or because > > they fixed the problem. > > They eighter can only detect a failure later (this may be a very long > time depending on access and verify runs) or they also write page > images. Those that write page images usually write "before images" to a > different area that is cleared periodically (e.g. during checkpoint). > > Writing to a different area was considered in pg, but there were more > negative issues than positive. > So imho pg_compresslog is the correct path forward. The current > discussion is only about whether we want a more complex pg_compresslog > and no change to current WAL, or an increased WAL size for a less > complex implementation. > Both would be able to compress the WAL to the same "archive log" size. > > Andreas > I definitely am in the camp of not increasing WAL size at all. If we need a bit more complexity to ensure that, so be it. Any approach that increases WAL volume would need to have an amazing benefit to make it warranted. This certainly does not meet that criteria. Ken
Andreas, > Writing to a different area was considered in pg, but there were more > negative issues than positive. > So imho pg_compresslog is the correct path forward. The current > discussion is only about whether we want a more complex pg_compresslog > and no change to current WAL, or an increased WAL size for a less > complex implementation. > Both would be able to compress the WAL to the same "archive log" size. Huh? As conceived, pg_compresslog does nothing to lower log volume for general purposes, just on-disk storage size for archiving. It doesn't help us at all with the tremendous amount of log we put out for an OLTP server, for example. Not that pg_compresslog isn't useful on its own for improving warm standby managability, but it's completely separate from addressing the "we're logging too much" issue. -- Josh Berkus PostgreSQL @ Sun San Francisco
Josh Berkus <josh@agliodbs.com> writes:
> Andreas,
>> So imho pg_compresslog is the correct path forward. The current
>> discussion is only about whether we want a more complex pg_compresslog
>> and no change to current WAL, or an increased WAL size for a less
>> complex implementation.
>> Both would be able to compress the WAL to the same "archive log" size.
> Huh? As conceived, pg_compresslog does nothing to lower log volume for
> general purposes, just on-disk storage size for archiving. It doesn't help
> us at all with the tremendous amount of log we put out for an OLTP server,
> for example.
I don't see how what you said refutes what he said. The sticking point
here is that the patch as-proposed *increases* the log volume before
compression.
regards, tom lane
Hi, Zeugswetter Andreas ADI SD wrote: >> I don't insist the name and the default of the GUC parameter. >> I'm afraid wal_fullpage_optimization = on (default) makes >> some confusion because the default behavior becomes a bit >> different on WAL itself. > > Seems my wal_fullpage_optimization is not a good name if it caused > misinterpretation already :-( > >>>> Amount of WAL after 60min. run of DBT-2 benchmark >>>> wal_add_optimization_info = off (default) 3.13GB >>> how about wal_fullpage_optimization = on (default) > > The meaning of wal_fullpage_optimization = on (default) > would be the same as your wal_add_optimization_info = off (default). > (Reversed name, reversed meaning of the boolean value) > > It would be there to *turn off* the (default) WAL full_page > optimization. > For your pg_compresslog it would need to be set to off. > "add_optimization_info" sounded like added info about/for some > optimization > which it is not. We turn off an optimization with the flag for the > benefit > of an easier pg_compresslog implementation. For pg_compresslog to remove full page writes, we need wal_add_optimization_info=on. > > As already said I would decouple this setting from the part that sets > the "removeable full page" flag in WAL, and making the recovery able to > skip dummy records. This I would do unconditionally. > > Andreas > > ---------------------------(end of broadcast)--------------------------- > TIP 3: Have you checked our extensive FAQ? > > http://www.postgresql.org/docs/faq > -- ------------- Koichi Suzuki
> > Writing to a different area was considered in pg, but there were more > > negative issues than positive. > > So imho pg_compresslog is the correct path forward. The current > > discussion is only about whether we want a more complex pg_compresslog > > and no change to current WAL, or an increased WAL size for a less > > complex implementation. > > Both would be able to compress the WAL to the same "archive log" size. > > Huh? As conceived, pg_compresslog does nothing to lower log > volume for general purposes, just on-disk storage size for > archiving. It doesn't help us at all with the tremendous > amount of log we put out for an OLTP server, for example. Ok, that is not related to the original discussion though. I have thus changed the subject, and removed [PATCHES]. You cannot directly compare the pg WAL size with other db's since they write parts to other areas (e.g. physical log in Informix). You would need to include those writes in a fair comparison. It is definitely not true, that writing to a different area has only advantages. The consensus was, that writing the page images to the WAL has more pro's. We could revisit the pros and cons though. Other options involve special OS and hardware (we already have that), or accepting a high risc of needing a restore after power outage (we don't have that, because we use no mechanism to detect such a failure). I am not sure that shrinking per WAL record size (other than the full page images), e.g. by only logging changed bytes and not whole tuples, would have a huge impact on OLTP tx/sec, since the limiting factor is IO's per second and not Mb per second. Recent developments like HOT seem a lot more promising in this regard since they avoid IO. Andreas
Josh,
Josh Berkus wrote:
> Koichi, Andreas,
>
>> 1) To deal with partial/inconsisitent write to the data file at crash
>> recovery, we need full page writes at the first modification to pages
>> after each checkpoint. It consumes much of WAL space.
>
> We need to find a way around this someday. Other DBs don't do this; it may be
> becuase they're less durable, or because they fixed the problem.
Maybe both. Fixing the problem may need some means to detect
partial/inconsistent writes to the data files, which may needs
additional CPU resource.
>
>> I don't think there should be only one setting. It depend on how
>> database is operated. Leaving wal_add_optiomization_info = off default
>> does not bring any change in WAL and archive log handling. I
>> understand some people may not be happy with additional 3% or so
>> increase in WAL size, especially people who dosn't need archive log at
>> all. So I prefer to leave the default off.
>
> Except that, is there any reason to turn this off if we are archiving? Maybe
> it should just be slaved to archive_command ... if we're not using PITR, it's
> off, if we are, it's on.
Hmm, this sounds to work. On the other hand, existing users, who are
happy with the current archiving and would not like to change current
archiving command to pg_compresslog or archive log size will increase a
bit. I'd like to hear some more on this.
>
>>> 1) is there any throughput benefit for platforms with fast CPU but
>>> contrained I/O (e.g. 2-drive webservers)? Any penalty for servers with
>>> plentiful I/O?
>> I've only run benchmarks with archive process running, because
>> wal_add_optimization_info=on does not make sense if we don't archive
>> WAL. In this situation, total I/O decreases because writes to archive
>> log decreases. Because of 3% or so increase in WAL size, there will be
>> increase in WAL write, but decrease in archive writes makes it up.
>
> Yeah, I was just looking for a way to make this a performance feature. I see
> now that it can't be. ;-)
As to the performance feature, I tested the patch against 8.3HEAD.
With pgbench, throughput was as follows:
Case1. Archiver: cp command, wal_add_optimization_info = off,
full_page_writes=on
Case2. Archiver: pg_compresslog, wal_add_optimization_info = on,
full_page_writes=on
DB Size: 1.65GB, Total transaction:1,000,000
Throughput was:
Case1: 632.69TPS
Case2: 653.10TPS ... 3% gain.
Archive Log Size:
Case1: 1.92GB
Case2: 0.57GB (about 30% of the Case1)... Before compression, the size
was 1.92GB. Because this is based on the number of WAL segment file
size, there will be at most 16MB error in the measurement. If we count
this, the increase in WAL I/O will be less than 1%.
>
>>> 3) How is this better than command-line compression for log-shipping?
>>> e.g. why do we need it in the database?
>> I don't fully understand what command-line compression means. Simon
>> suggested that this patch can be used with log-shipping and I agree.
>> If we compare compression with gzip or other general purpose
>> compression, compression ratio, CPU usage and I/O by pg_compresslog are
>> all quite better than those in gzip.
>
> OK, that answered my question.
>
>> This is why I don't like Josh's suggested name of wal_compressable
>> eighter.
>> WAL is compressable eighter way, only pg_compresslog would need to be
>> more complex if you don't turn off the full page optimization. I think a
>> good name would tell that you are turning off an optimization.
>> (thus my wal_fullpage_optimization on/off)
>
> Well, as a PG hacker I find the name wal_fullpage_optimization quite baffling
> and I think our general user base will find it even more so. Now that I have
> Koichi's explanation of the problem, I vote for simply slaving this to the
> PITR settings and not having a separate option at all.
Could I have more specific suggestion on this?
Regards;
--
-------------
Koichi Suzuki
On Thu, 26 Apr 2007, Zeugswetter Andreas ADI SD wrote: > I am not sure that shrinking per WAL record size (other than the full > page images), e.g. by only logging changed bytes and not whole tuples, > would have a huge impact on OLTP tx/sec, since the limiting factor is > IO's per second and not Mb per second. With the kind of caching controller that's necessary for any serious OLTP work with Postgres, number of I/Os per second isn't really an important number. Total volume of writes to the WAL volume can be though. It's difficult but not impossible to encounter a workload that becomes bottlenecked by WAL volume on a good OLTP server, particularly because that's often going to a single or RAID-1 disk. Whether those workloads also have the appropriate properties such that their WAL could be shrunk usefully in real-time is a good question. -- * Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD
On Apr 27, 2007, at 4:58 AM, Greg Smith wrote: > On Thu, 26 Apr 2007, Zeugswetter Andreas ADI SD wrote: >> I am not sure that shrinking per WAL record size (other than the full >> page images), e.g. by only logging changed bytes and not whole >> tuples, >> would have a huge impact on OLTP tx/sec, since the limiting factor is >> IO's per second and not Mb per second. > > With the kind of caching controller that's necessary for any > serious OLTP work with Postgres, number of I/Os per second isn't > really an important number. Total volume of writes to the WAL > volume can be though. It's difficult but not impossible to > encounter a workload that becomes bottlenecked by WAL volume on a > good OLTP server, particularly because that's often going to a > single or RAID-1 disk. Whether those workloads also have the > appropriate properties such that their WAL could be shrunk usefully > in real-time is a good question. Yes, but how many data drives would you need to have to bottleneck on WAL? Even if the entire database is memory resident you'd still have to write all the pages out at some point, and it seems to me that you'd need a fair amount of disk capacity the data directory before you got pegged by WAL. When I did some DBT2 testing a bit over a year ago I had a 20 drive RAID10 for data and a mirror for WAL and was nowhere close to pegged on WAL (this was on a Sun V40 connected to one of their storage arrays). -- Jim Nasby jim@nasby.net EnterpriseDB http://enterprisedb.com 512.569.9461 (cell)
On Fri, 27 Apr 2007, Jim Nasby wrote: > Yes, but how many data drives would you need to have to bottleneck on WAL? > Even if the entire database is memory resident you'd still have to write all > the pages out at some point, and it seems to me that you'd need a fair amount > of disk capacity the data directory before you got pegged by WAL. Depends on the type of transactions. If you're doing something with lots of INSERT and perhaps some UPDATE volume that doesn't need to read heavily from the database to complete, because most of the important stuff is already in memory, you might run into the WAL limit without too much on the database disk side. I did say it was difficult... > When I did some DBT2 testing a bit over a year ago I had a 20 drive RAID10 > for data and a mirror for WAL and was nowhere close to pegged on WAL (this > was on a Sun V40 connected to one of their storage arrays). No doubt, the main reason I haven't used DBT2 more is because the WAL volume produced before you run into database limited bottlenecks isn't large, and certainly not in the same ratio as some of the apps I'm modeling. Mine lean more toward straightforward transaction logging in parts. I'm running on similar hardware (V40 is very close, I think the EMC array I test against is a bit better than the most of the Sun models) and I've seen some scenarios that produce 40MB/s average - 60MB/s peak of WAL volume. Sure seems like I'm rate limited by the RAID-1 WAL disk. As you say, eventually all the data has to make it to disk, but since it's not too expensive nowadays to have gigabytes worth of memory and disk array cache you can put off database writes for a surprisingly long period of time with the right system design. It's harder to buffer those pesky O_DIRECT WAL writes when they blow right though at least one level of cache. -- * Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD
Hi,
As replied to "Patch queue triage" by Tom, here's simplified patch to
mark WAL record as "compressable", with no increase in WAL itself.
Compression/decompression commands will be posted separately to PG
Foundary for further review.
-------------------------------
As suggested by Tom, I agree that WAL should not include "both" full
page write and incremental (logical) log. I began to examine WAL
record format to see if incremental log can be made from full page
writes. It will be okay even before 8.4, if simplified patch to the
core is accepted. I will post simplified patch to the core as follows:
1. Mark the flag to indicate that the WAL record is compressable from
full page writes to incremental log. This flag will be set if
a) It is not written during the hot backup, and
b) Archive command is active, and
c) WAL contains full page writes, and
d) full_page_writes=on.
No logical log will be written to WAL in this case, and
2. During recovery, xl_tot_len check will be skipped for compressed WAL
records.
Please note that new GUC is not needed in this patch.
With this patch, compress/decompress can be developped outside the core.
I'd be very grateful if this patch can be considered again.
Best Regards;
--
-------------
Koichi Suzuki
diff -cr pgsql_org/src/backend/access/transam/xlog.c pgsql/src/backend/access/transam/xlog.c
*** pgsql_org/src/backend/access/transam/xlog.c 2007-05-02 15:56:38.000000000 +0900
--- pgsql/src/backend/access/transam/xlog.c 2007-05-07 16:30:38.000000000 +0900
***************
*** 837,842 ****
--- 837,854 ----
return RecPtr;
}
+ /*
+ * If online backup is not in progress and WAL archiving is active, mark
+ * backup blocks removable if any.
+ * This mark will be referenced during archiving to remove needless backup
+ * blocks in the record and compress WAL segment files.
+ */
+ if (XLogArchivingActive() && fullPageWrites &&
+ (info & XLR_BKP_BLOCK_MASK) && !Insert->forcePageWrites)
+ {
+ info |= XLR_BKP_REMOVABLE;
+ }
+
/* Insert record header */
record = (XLogRecord *) Insert->currpos;
***************
*** 2738,2750 ****
blk += blen;
}
! /* Check that xl_tot_len agrees with our calculation */
! if (blk != (char *) record + record->xl_tot_len)
{
! ereport(emode,
! (errmsg("incorrect total length in record at %X/%X",
! recptr.xlogid, recptr.xrecoff)));
! return false;
}
/* Finally include the record header */
--- 2750,2778 ----
blk += blen;
}
! /*
! * If physical log has not been removed, check the length to see
! * the following.
! * - No physical log existed originally,
! * - WAL record was not removable because it is generated during
! * the online backup,
! * - Cannot be removed because the physical log spanned in
! * two segments.
! * The reason why we skip the length check on the physical log removal is
! * that the flag XLR_SET_BKB_BLOCK(0..2) is reset to zero and it prevents
! * the above loop to proceed blk to the end of the record.
! */
! if (!(record->xl_info & XLR_BKP_REMOVABLE) ||
! record->xl_info & XLR_BKP_BLOCK_MASK)
{
! /* Check that xl_tot_len agrees with our calculation */
! if (blk != (char *) record + record->xl_tot_len)
! {
! ereport(emode,
! (errmsg("incorrect total length in record at %X/%X",
! recptr.xlogid, recptr.xrecoff)));
! return false;
! }
}
/* Finally include the record header */
pgsql/src/backend/access/transamだけに発見: xlog.c.orig
diff -cr pgsql_org/src/include/access/xlog.h pgsql/src/include/access/xlog.h
*** pgsql_org/src/include/access/xlog.h 2007-01-06 07:19:51.000000000 +0900
--- pgsql/src/include/access/xlog.h 2007-05-07 16:30:38.000000000 +0900
***************
*** 66,73 ****
/*
* If we backed up any disk blocks with the XLOG record, we use flag bits in
* xl_info to signal it. We support backup of up to 3 disk blocks per XLOG
! * record. (Could support 4 if we cared to dedicate all the xl_info bits for
! * this purpose; currently bit 0 of xl_info is unused and available.)
*/
#define XLR_BKP_BLOCK_MASK 0x0E /* all info bits used for bkp blocks */
#define XLR_MAX_BKP_BLOCKS 3
--- 66,74 ----
/*
* If we backed up any disk blocks with the XLOG record, we use flag bits in
* xl_info to signal it. We support backup of up to 3 disk blocks per XLOG
! * record.
! * Bit 0 of xl_info is used to represent that backup blocks are not necessary
! * in archive-log.
*/
#define XLR_BKP_BLOCK_MASK 0x0E /* all info bits used for bkp blocks */
#define XLR_MAX_BKP_BLOCKS 3
***************
*** 75,80 ****
--- 76,82 ----
#define XLR_BKP_BLOCK_1 XLR_SET_BKP_BLOCK(0) /* 0x08 */
#define XLR_BKP_BLOCK_2 XLR_SET_BKP_BLOCK(1) /* 0x04 */
#define XLR_BKP_BLOCK_3 XLR_SET_BKP_BLOCK(2) /* 0x02 */
+ #define XLR_BKP_REMOVABLE XLR_SET_BKP_BLOCK(3) /* 0x01 */
/*
* Sometimes we log records which are out of transaction control.
Your patch has been added to the PostgreSQL unapplied patches list at:
http://momjian.postgresql.org/cgi-bin/pgpatches
It will be applied as soon as one of the PostgreSQL committers reviews
and approves it.
---------------------------------------------------------------------------
Koichi Suzuki wrote:
> Hi,
>
> As replied to "Patch queue triage" by Tom, here's simplified patch to
> mark WAL record as "compressable", with no increase in WAL itself.
> Compression/decompression commands will be posted separately to PG
> Foundary for further review.
>
> -------------------------------
> As suggested by Tom, I agree that WAL should not include "both" full
> page write and incremental (logical) log. I began to examine WAL
> record format to see if incremental log can be made from full page
> writes. It will be okay even before 8.4, if simplified patch to the
> core is accepted. I will post simplified patch to the core as follows:
>
> 1. Mark the flag to indicate that the WAL record is compressable from
> full page writes to incremental log. This flag will be set if
> a) It is not written during the hot backup, and
> b) Archive command is active, and
> c) WAL contains full page writes, and
> d) full_page_writes=on.
> No logical log will be written to WAL in this case, and
> 2. During recovery, xl_tot_len check will be skipped for compressed WAL
> records.
>
> Please note that new GUC is not needed in this patch.
>
> With this patch, compress/decompress can be developped outside the core.
>
> I'd be very grateful if this patch can be considered again.
>
> Best Regards;
>
> --
> -------------
> Koichi Suzuki
> diff -cr pgsql_org/src/backend/access/transam/xlog.c pgsql/src/backend/access/transam/xlog.c
> *** pgsql_org/src/backend/access/transam/xlog.c 2007-05-02 15:56:38.000000000 +0900
> --- pgsql/src/backend/access/transam/xlog.c 2007-05-07 16:30:38.000000000 +0900
> ***************
> *** 837,842 ****
> --- 837,854 ----
> return RecPtr;
> }
>
> + /*
> + * If online backup is not in progress and WAL archiving is active, mark
> + * backup blocks removable if any.
> + * This mark will be referenced during archiving to remove needless backup
> + * blocks in the record and compress WAL segment files.
> + */
> + if (XLogArchivingActive() && fullPageWrites &&
> + (info & XLR_BKP_BLOCK_MASK) && !Insert->forcePageWrites)
> + {
> + info |= XLR_BKP_REMOVABLE;
> + }
> +
> /* Insert record header */
>
> record = (XLogRecord *) Insert->currpos;
> ***************
> *** 2738,2750 ****
> blk += blen;
> }
>
> ! /* Check that xl_tot_len agrees with our calculation */
> ! if (blk != (char *) record + record->xl_tot_len)
> {
> ! ereport(emode,
> ! (errmsg("incorrect total length in record at %X/%X",
> ! recptr.xlogid, recptr.xrecoff)));
> ! return false;
> }
>
> /* Finally include the record header */
> --- 2750,2778 ----
> blk += blen;
> }
>
> ! /*
> ! * If physical log has not been removed, check the length to see
> ! * the following.
> ! * - No physical log existed originally,
> ! * - WAL record was not removable because it is generated during
> ! * the online backup,
> ! * - Cannot be removed because the physical log spanned in
> ! * two segments.
> ! * The reason why we skip the length check on the physical log removal is
> ! * that the flag XLR_SET_BKB_BLOCK(0..2) is reset to zero and it prevents
> ! * the above loop to proceed blk to the end of the record.
> ! */
> ! if (!(record->xl_info & XLR_BKP_REMOVABLE) ||
> ! record->xl_info & XLR_BKP_BLOCK_MASK)
> {
> ! /* Check that xl_tot_len agrees with our calculation */
> ! if (blk != (char *) record + record->xl_tot_len)
> ! {
> ! ereport(emode,
> ! (errmsg("incorrect total length in record at %X/%X",
> ! recptr.xlogid, recptr.xrecoff)));
> ! return false;
> ! }
> }
>
> /* Finally include the record header */
> pgsql/src/backend/access/transamだけに発見: xlog.c.orig
> diff -cr pgsql_org/src/include/access/xlog.h pgsql/src/include/access/xlog.h
> *** pgsql_org/src/include/access/xlog.h 2007-01-06 07:19:51.000000000 +0900
> --- pgsql/src/include/access/xlog.h 2007-05-07 16:30:38.000000000 +0900
> ***************
> *** 66,73 ****
> /*
> * If we backed up any disk blocks with the XLOG record, we use flag bits in
> * xl_info to signal it. We support backup of up to 3 disk blocks per XLOG
> ! * record. (Could support 4 if we cared to dedicate all the xl_info bits for
> ! * this purpose; currently bit 0 of xl_info is unused and available.)
> */
> #define XLR_BKP_BLOCK_MASK 0x0E /* all info bits used for bkp blocks */
> #define XLR_MAX_BKP_BLOCKS 3
> --- 66,74 ----
> /*
> * If we backed up any disk blocks with the XLOG record, we use flag bits in
> * xl_info to signal it. We support backup of up to 3 disk blocks per XLOG
> ! * record.
> ! * Bit 0 of xl_info is used to represent that backup blocks are not necessary
> ! * in archive-log.
> */
> #define XLR_BKP_BLOCK_MASK 0x0E /* all info bits used for bkp blocks */
> #define XLR_MAX_BKP_BLOCKS 3
> ***************
> *** 75,80 ****
> --- 76,82 ----
> #define XLR_BKP_BLOCK_1 XLR_SET_BKP_BLOCK(0) /* 0x08 */
> #define XLR_BKP_BLOCK_2 XLR_SET_BKP_BLOCK(1) /* 0x04 */
> #define XLR_BKP_BLOCK_3 XLR_SET_BKP_BLOCK(2) /* 0x02 */
> + #define XLR_BKP_REMOVABLE XLR_SET_BKP_BLOCK(3) /* 0x01 */
>
> /*
> * Sometimes we log records which are out of transaction control.
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 9: In versions below 8.0, the planner will ignore your desire to
> choose an index scan if your joining column's datatypes do not
> match
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com
+ If your life is a hard drive, Christ can be your backup. +
Koichi Suzuki <suzuki.koichi@oss.ntt.co.jp> writes:
> As replied to "Patch queue triage" by Tom, here's simplified patch to
> mark WAL record as "compressable", with no increase in WAL itself.
> Compression/decompression commands will be posted separately to PG
> Foundary for further review.
Applied with some minor modifications. I didn't like the idea of
suppressing the sanity-check on WAL record length; I think that's
fairly important. Instead, I added a provision for an XLOG_NOOP
WAL record type that can be used to fill in the extra space.
The way I envision that working is that the compressor removes
backup blocks and converts each compressible WAL record to have the
same contents and length it would've had if written without backup
blocks. Then, it inserts an XLOG_NOOP record with length set to
indicate the amount of extra space that needs to be chewed up --
but in the compressed version of the WAL file, XLOG_NOOP's "data
area" is not actually stored. The decompressor need only scan
the file looking for XLOG_NOOP and insert the requisite number of
zero bytes (and maybe recompute the XLOG_NOOP's CRC, depending on
whether you want it to be valid for the short-format record in the
compressed file). There will also be some games to be played for
WAL page boundaries, but you had to do that anyway.
regards, tom lane
I really appreciate for the modification. I also believe XLOG_NOOP is cool to maintains XLOG format consistent. I'll continue to write a code to produce incremental log record from the full page writes as well as too maintain CRC, XLOOG_NOOP and other XLOG locations, I also found that you've added information on btree strip log records, which anables to produce corresponding incremental logs from the full page writes. 2007/5/21, Tom Lane <tgl@sss.pgh.pa.us>: > Koichi Suzuki <suzuki.koichi@oss.ntt.co.jp> writes: > > As replied to "Patch queue triage" by Tom, here's simplified patch to > > mark WAL record as "compressable", with no increase in WAL itself. > > Compression/decompression commands will be posted separately to PG > > Foundary for further review. > > Applied with some minor modifications. I didn't like the idea of > suppressing the sanity-check on WAL record length; I think that's > fairly important. Instead, I added a provision for an XLOG_NOOP > WAL record type that can be used to fill in the extra space. > The way I envision that working is that the compressor removes > backup blocks and converts each compressible WAL record to have the > same contents and length it would've had if written without backup > blocks. Then, it inserts an XLOG_NOOP record with length set to > indicate the amount of extra space that needs to be chewed up -- > but in the compressed version of the WAL file, XLOG_NOOP's "data > area" is not actually stored. The decompressor need only scan > the file looking for XLOG_NOOP and insert the requisite number of > zero bytes (and maybe recompute the XLOG_NOOP's CRC, depending on > whether you want it to be valid for the short-format record in the > compressed file). There will also be some games to be played for > WAL page boundaries, but you had to do that anyway. > > regards, tom lane > > ---------------------------(end of broadcast)--------------------------- > TIP 3: Have you checked our extensive FAQ? > > http://www.postgresql.org/docs/faq > -- ------ Koichi Suzuki
On Tue, 2007-04-10 at 16:23 +0900, Koichi Suzuki wrote: > Here're two patches for > > 1) lesslog_core.patch, patch for core, to set a mark to the log record > to be removed in archiving, > > 2) lesslog_contrib.patch, patch for contrib/lesslog, pg_compresslog and > pg_decompresslog, > > respectively, as asked. I hope they work. Koichi-san, Earlier, I offered to document the use of pg_compresslog and pg_decompresslog and would like to do that now. My understanding was that we would make these programs available on pgfoundry.org. Unfortunately, I can't find these files there, so perhaps I misunderstood. Do we have later versions of these programs that work with the changes Tom committed on 20 May? Or is the code posted here the latest version? Many thanks, -- Simon Riggs 2ndQuadrant http://www.2ndQuadrant.com