Thread: files under pg_clog directories are missing
Hello Everybody - Due to hardware crash we ran into issues where some blocks were corrupted and some files were missing. I was able to get over the corrupted blocks ( errmsg - "invalid page header in block 12345 of realtion x") by setting zero_damaged_pages=0 and running vacuum afterwards. Now I am running into situations where pg_clog files are missing (errmsg - "could not open pg_clog/0D0D). I have a backup which is quite old ( considering this as a last resort). Is there any other way to fix this problem? I also created empty blocks to fool postgres, but there are so many of these file missing I was wondering if there a better/faster way to fix this problem. Thanks. Dinesh
Dinesh Bhandary wrote: > Hello Everybody - > > Due to hardware crash we ran into issues where some blocks were > corrupted and some files were missing. > I was able to get over the corrupted blocks ( errmsg - "invalid page > header in block 12345 of realtion x") by setting > zero_damaged_pages=0 and running vacuum afterwards. Now I am running > into situations where pg_clog files are missing (errmsg - "could not > open pg_clog/0D0D). I have a backup which is quite old ( considering > this as a last resort). Is there any other way to fix this problem? If those clog numbers are not in your usual range, it usually means the row is corrupt and is referring to an corrupt xid number, so the row is bad, not the clog files are missing. > > I also created empty blocks to fool postgres, but there are so many of > these file missing I was wondering if there a better/faster way to fix > this problem. Seems you will have to clear those pages too somehow. I have seen it done with dd. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
Dinesh Bhandary wrote: > Due to hardware crash we ran into issues where some blocks were > corrupted and some files were missing. > I was able to get over the corrupted blocks ( errmsg - "invalid page > header in block 12345 of realtion x") by setting > zero_damaged_pages=0 and running vacuum afterwards. Now I am running > into situations where pg_clog files are missing (errmsg - "could not > open pg_clog/0D0D). I have a backup which is quite old ( considering > this as a last resort). Is there any other way to fix this problem? > > I also created empty blocks to fool postgres, but there are so many of > these file missing I was wondering if there a better/faster way to fix > this problem. I hope you made a whole backup of the files before you started trying to fix the damage too. It's possible to try and fix this using tricks like zero_damaged_pages and dummy clog files, only to make things worse. We do data recovery services here, and I've had to roll back to the original copy of the data multiple times before in order to try different things before getting a get good copy of someone's data back again. If you don't have a copy of the database yet, do that before you do any more experimenting with the clog files. I wrote a summary of links to past work like this you may find useful, and a little program to create missing pg_clog files that all say "the transaction you're asking about committed", available at: http://archives.postgresql.org/pgsql-general/2009-07/msg00985.php You might further script that to speed up how fast you can fix these as they pop up, which makes the test/correct cycle time go down. You might even write a script that loops over starting the database, looks at the end of the log file, and if it's yet another one of these missing just extract its number, recreate it, and start again. Unfortunately, doing better than that is tricky. We had to modify the PostgreSQL source code to automatically create them in order to handle this safely last time I ran into one of these that was badly corrupted and missing a whole lot of real clog files, not just ones that were unlikely to exist. You should be staring at the numbers of each one of these as they're requested. If the range is way outside of the active clog files you have, that's probably one you can create safely because it's garbage data anyway. But if it starts asking for clog files that are in the middle or near the ends of the set you've got, you may have a bigger problem on your hands. P.S. Make sure you dump a whole copy of the database the minute you get it started again and reload that before you start using it. You have no idea what state all of the tables are really in after a crash like this without such an exercise. -- Greg Smith, 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services and Support www.2ndQuadrant.us Author, "PostgreSQL 9.0 High Performance" Pre-ordering at: https://www.packtpub.com/postgresql-9-0-high-performance/book
As a followup to all this, the OP also needs to figure out why a crashed server could cause so much corruption. This is usually the result of hardware lying about fsync status, i.e. consumer grade hardware pressed into service as a server.
Scott Marlowe wrote: > As a followup to all this, the OP also needs to figure out why a > crashed server could cause so much corruption. This is usually the > result of hardware lying about fsync status, i.e. consumer grade > hardware pressed into service as a server. > See http://www.postgresql.org/docs/9.0/static/wal-reliability.html for more information. Nothing there has changed in 9.0 relative to the version being run here, but the docs for that version got a big upgrade over the earlier ones. Bad memory is always a potential source for issues in this area too. Mostly if you don't know why the crash happened, either. If you know for sure how it crashed, but only then found corruption, Scott's right that it's more likely to be a lying disk drive instead. -- Greg Smith, 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services and Support www.2ndQuadrant.us Author, "PostgreSQL 9.0 High Performance" Pre-ordering at: https://www.packtpub.com/postgresql-9-0-high-performance/book
Good reminder to backup before you start. Yes, I've made the backup of the db in its current state. After reading your posts and doing a slew of research I am inclining more towards restoring from the last successful dump and reload data from applications (it seems like it is going to be partial recovery). I am concerned that even if we are able to clear all those messages by whatever means the state of the database will be corrupted. Luckily this happens to be ready only db and we can live with it. As for the nature of the corruption I still do know know what kind of hardware problems led to this; it happened at one of our clients site and we are still waiting to find out what caused it. One piece of info we got was postgres data directory turned into read only partition. Thanks everyone. Dinesh ps@ the ids of the missing clog_files are out of range. On 9/17/2010 8:32 PM, Greg Smith wrote: > Dinesh Bhandary wrote: >> Due to hardware crash we ran into issues where some blocks were >> corrupted and some files were missing. >> I was able to get over the corrupted blocks ( errmsg - "invalid page >> header in block 12345 of realtion x") by setting >> zero_damaged_pages=0 and running vacuum afterwards. Now I am running >> into situations where pg_clog files are missing (errmsg - "could not >> open pg_clog/0D0D). I have a backup which is quite old ( considering >> this as a last resort). Is there any other way to fix this problem? >> >> I also created empty blocks to fool postgres, but there are so many >> of these file missing I was wondering if there a better/faster way to >> fix this problem. > > I hope you made a whole backup of the files before you started trying > to fix the damage too. It's possible to try and fix this using tricks > like zero_damaged_pages and dummy clog files, only to make things > worse. We do data recovery services here, and I've had to roll back > to the original copy of the data multiple times before in order to try > different things before getting a get good copy of someone's data back > again. If you don't have a copy of the database yet, do that before > you do any more experimenting with the clog files. > > I wrote a summary of links to past work like this you may find useful, > and a little program to create missing pg_clog files that all say "the > transaction you're asking about committed", available at: > http://archives.postgresql.org/pgsql-general/2009-07/msg00985.php > > You might further script that to speed up how fast you can fix these > as they pop up, which makes the test/correct cycle time go down. You > might even write a script that loops over starting the database, looks > at the end of the log file, and if it's yet another one of these > missing just extract its number, recreate it, and start again. > > Unfortunately, doing better than that is tricky. We had to modify the > PostgreSQL source code to automatically create them in order to handle > this safely last time I ran into one of these that was badly corrupted > and missing a whole lot of real clog files, not just ones that were > unlikely to exist. You should be staring at the numbers of each one > of these as they're requested. If the range is way outside of the > active clog files you have, that's probably one you can create safely > because it's garbage data anyway. But if it starts asking for clog > files that are in the middle or near the ends of the set you've got, > you may have a bigger problem on your hands. > > P.S. Make sure you dump a whole copy of the database the minute you > get it started again and reload that before you start using it. You > have no idea what state all of the tables are really in after a crash > like this without such an exercise. >
Dinesh Bhandary wrote: > Good reminder to backup before you start. Yes, I've made the backup of > the db in its current state. > > After reading your posts and doing a slew of research I am inclining > more towards restoring from the last successful dump and reload data > from applications (it seems like it is going to be partial recovery). I > am concerned that even if we are able to clear all those messages by > whatever means the state of the database will be corrupted. Luckily this True. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
Dinesh Bhandary wrote: > As for the nature of the corruption I still do know know what kind of > hardware problems led to this; it happened at one of our clients site > and we are still waiting to find out what caused it. One piece of > info we got was postgres data directory turned into read only partition. That can happen when a major disk-level problem occurs; system remounts as read-only because it doesn't think it can safely write to it anymore. I'd check into the system kernel logs as soon as possible, to try and find a disk error that kicked the whole thing off. -- Greg Smith, 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services and Support www.2ndQuadrant.us Author, "PostgreSQL 9.0 High Performance" Pre-ordering at: https://www.packtpub.com/postgresql-9-0-high-performance/book