Thread: Autovacuum integration patch
Hackers, Here is a first cut at autovacuum integration. Please have a look at it. Note that this patch automatically creates three new files: src/backend/postmaster/autovacuum.c src/include/catalog/pg_autovacuum.h src/include/postmaster/autovacuum.h Note that the daemon is not activated by default. There are several things that are painfully evident with this thing on: - TRUNCATE does not update stats. It should send a stat message to which we can react. - If you empty a whole table using DELETE just after an automatically-issued VACUUM takes place, the new threshold may not be enough to trigger a new VACUUM. Thus you end up with a bloated table, and it won't get vacuumed until it grows again. This may be a problem with the cost equations, but those are AFAICT identical to those of pg_autovacuum, so we may need to rethink the equations. - The default value of on for reset stats on server start is going to be painful with autovacuum, because it reacts badly to losing the info. - We should make VACUUM and ANALYZE update the pg_autovacuum relation, in order to make the autovacuum daemon behave sanely with manually issued VACUUM/ANALYZE. - Having an autovacuum process running on a database can be surprising if you want to drop a database, or create a new one using it as a template. This happenned to me several times. - The shutdown sequence is not debugged nor very well tested. It may be all wrong. - The startup sequence is a mixture from pgarch, normal backend and pgstat. I find it relatively clean but I can't swear it's bug-free. - There are no docs - There are no ALTER TABLE commands to change the pg_autovacuum attributes for a table. (Enable/disable, set thresholds and scaling factor) - I compiled with -DEXEC_BACKEND, but I didn't look to see if it actually worked on that case. Apart from all these issues, it is completely functional :-) It can survive several "make installcheck" runs without problem, and the regression database is vacuumed/analyzed as it runs. Some of these issues are trivial to handle. However I'd like to release this right now, so I can go back to "shared dependencies" now that role support is in. Barring any objections I think this should be integrated, so these issues can be tackled by interested parties. -- Alvaro Herrera (<alvherre[a]surnet.cl>) "World domination is proceeding according to plan" (Andrew Morton)
Attachment
Alvaro Herrera wrote: >Hackers, > >Here is a first cut at autovacuum integration. Please have a look at >it. Note that this patch automatically creates three new files: > > Couple more things that I didn't think about while we were talking about this the other day. XID wraparound: The patch as submitted doesn't handle XID wraparound issues. The old contrib autovacuum would do an XID wraparound check as it's 1st operation upon connecting to a database. If XID wraparound was looks like it's going to be a problem soon, then the whole database would be vacuumed, eliminating the need to check specific tables. Better logging of autovacuum activity: I think the we could use some more detail in the debug elog statements. For example showing exactly what autovacuum believes the threshold and current count is. How to deal with shared relations: As an optimization, the contrib version of autovacuum treated shared relations different than it treated the rest when connected to any database that is not template1. That is, when connected to a DB other than template1, autovacuum would not issue vacuum commands. rather it would only issue analyze commands. When autovacuum got around to connecting to template1, it would then issue the vacuum command. The hope was that this would reducing a shared relation from getting vacuumed n times (where n is the number of databases in a cluster) whenever it crossed over it's threshold. I'm not sure if this optimizaion is really important, or even exactly correct.
> > >>>> XID wraparound: The patch as submitted doesn't handle XID wraparound >>>> issues. The old contrib autovacuum would do an XID wraparound check as >>>> it's 1st operation upon connecting to a database. If XID wraparound was >>>> looks like it's going to be a problem soon, then the whole database >>>> would be vacuumed, eliminating the need to check specific tables. >> >> >> >> Hmm. Yes, this patch doesn't handle Xid wraparound. This should be >> easy to add though. Anyway, I was thinking that we could add a "last >> vacuum Xid" to pg_autovacuum, and handle Xid wraparound for each table >> separately -- this means you don't have to issue huge whole-database >> VACUUMs, because it will be handled nicely for each table. Storing the >> last vacuum Xid in pg_database would have to be rethought. > > The current implementation of XID wraparound requires that the vacuum command be run against the entire database, you can not run it on a per table basis and have it work. At least that is my understanding, it would require some reworking of the vacuum system and I have no idea what is involved in that. For now, we should just do it the simple way. BTW, I think this is a candidate for only being done during the maintenance window. >> Maybe what we could do is have a separate pg_vacuum table to hold >> constantly-moving information about the tables: last vacuum Xid, count >> of tuples at last vacuum/analyze, etc; so pg_autovacuum would only hold >> the constants for autovacuum equations. This pg_vacuum table would be >> updated by VACUUM, not autovacuum, so it would be always correct and up- >> to-date. > > I'm not sure I see the value in a new pg_vacuum table. reltuples already has the tuple count from the last vacuum and I don't think last XID on a per table basis is helpful. >>>> Better logging of autovacuum activity: I think the we could use some >>>> more detail in the debug elog statements. For example showing exactly >>>> what autovacuum believes the threshold and current count is. >> >> >> >> Ok. I actually had lots more logging in the original patch, but I >> removed it because it was too verbose. Again, it's easy to add. > > Well, I don't know what is best, but it would be nice to be able to get at the information that tells you why autovacuum did or did not take action. Perhaps put back what you had in, but move it up to a higher debug level. FWIW, I think the debug info from the contribversion was sufficient. >>>> How to deal with shared relations: As an optimization, the contrib >>>> version of autovacuum treated shared relations different than it treated >>>> the rest when connected to any database that is not template1. >> >> >> >> Ah, interesting. Yes, I think that could be done too. Very easy to do. >> Anyway, the shared relations are not that big usually, so this shouldn't >> be an issue. > > Agreed this is not a big issue, it's a bit of a micro optimization. >>>> Couple of other thoughts: >>>> Do the vacuum commands respect the GUC vacuum delay settings? >> >> >> >> Huh, I don't know. I just issue a vacuum() call. That function sets >> the delay settings AFAICS, so I think it should be working. > > Can someone confirm this? >>>> Should we be able to set per table vacuum delay settings? >> >> >> We could set that in the hypotetical pg_vacuum relation. > > Again, I don't think this would be good for the pg_vacuum table, I think it should be in the autovacuum table, because what a user wants autovacuum to do might be different than what he wants a manually run vacuum to do. >>>> This patch doesn't have the "maintenance window" that was discussed a >>>> while ago. >> >> >> True. I have several questions about it. Where would that information >> be stored, in another system catalog? Would it be per-database or >> per-table? What happens if I'm not able to do all work inside the >> maintenance window, is it left for the next one? If the maintenance >> window ends and there is a vacuum running, is it terminated or is it >> allowed to continue? > > One could argue that it should be per database, but I think per cluster should be sufficient. I think it could be handledas few GUC settings, such as: autovac_maint_begin = "1AM" autovac_maint_duration = 4 (measured in hours) autovac_maint_factor = .5 (reduce the thresholds by half during the maintenance window, this option might be good to haveon a per table basis, if so, then add it to the pg_autovacuum table) If there is still work to do after the maint window expires, then it's left for next time or when the regular threshold isexceeded which ever happens first. I wouldn't terminate an in progress vacuum. >> There is a very important issue I forgot to mention. This autovacuum >> process only handles databases that exist in the Stats hash table. >> However, the stat hash table only has information about databases and >> tables that have been used in the current postmaster run. So if you >> don't connect to a database regularly, that database won't get >> "autovacuumed" after a postmaster restart. I think (but IMBFOS) that >> this is also true for individual tables, i.e. a table that doesn't get >> used won't be in the stat hash table and thus won't be processed. This >> is a very important "gotcha." > > Is it? If no one has connected to that database, it can't need to be vacuumed. Especially in a production situation, I could imagine this being an issue in development but... >> This is, of course, not important in the normal case, because a table >> that isn't used does not need vacuum. But in the Xid-wraparound case >> it may be fatal. > > hmmm... didn't think about that... Perhaps a better way to get our database list is to manually create one each time we connect to the postgres database, or the template1 database if the postgres database doesn't exist. Thoughts? >> I think the stat collector should be told about every existant database >> and every table therein, so that autovacuum can do its work as the user >> would expect. > > Seems wrong to me to rework the stats system to accommodate autovacuum. Perhaps there is an easier way such as what I mentionedabove or something else. Matt
"Matthew T. O'Connor" <matthew@zeut.net> writes:
>>> Hmm. Yes, this patch doesn't handle Xid wraparound. This should be
>>> easy to add though. Anyway, I was thinking that we could add a "last
>>> vacuum Xid" to pg_autovacuum, and handle Xid wraparound for each table
>>> separately -- this means you don't have to issue huge whole-database
>>> VACUUMs, because it will be handled nicely for each table. Storing the
>>> last vacuum Xid in pg_database would have to be rethought.
> The current implementation of XID wraparound requires that the vacuum
> command be run against the entire database, you can not run it on a per
> table basis and have it work. At least that is my understanding,
No, you're wrong. VACUUMing of individual tables is perfectly good
enough as far as XID wrap protection goes, it's just that we chose to
track whether it had been done at the database level. If we tracked it
in, say, a new pg_class column then in principle you could protect
against XID wrap with only table-at-a-time VACUUMs. (I think you'd
still want the pg_database column, but you'd update it to be the minimum
of the per-table values at the completion of any VACUUM.)
At the time this didn't seem particularly worth the complication since
no one would be likely to try to do that manually --- but with
autovacuum handling the work, it starts to sound more realistic.
regards, tom lane
Tom Lane wrote: >"Matthew T. O'Connor" <matthew@zeut.net> writes: > > >>The current implementation of XID wraparound requires that the vacuum >>command be run against the entire database, you can not run it on a per >>table basis and have it work. At least that is my understanding, >> >> > >No, you're wrong. VACUUMing of individual tables is perfectly good >enough as far as XID wrap protection goes, it's just that we chose to >track whether it had been done at the database level. If we tracked it >in, say, a new pg_class column then in principle you could protect >against XID wrap with only table-at-a-time VACUUMs. (I think you'd >still want the pg_database column, but you'd update it to be the minimum >of the per-table values at the completion of any VACUUM.) > >At the time this didn't seem particularly worth the complication since >no one would be likely to try to do that manually --- but with >autovacuum handling the work, it starts to sound more realistic. > Good, I'm glad I'm wrong on this. This will be another nice advantage of autovacuum then and should be fairly easy to do. Any thoughts on this being a change we can get in for 8.1? Matt
"Matthew T. O'Connor" <matthew@zeut.net> writes:
> Tom Lane wrote:
>> No, you're wrong. VACUUMing of individual tables is perfectly good
>> enough as far as XID wrap protection goes, it's just that we chose to
>> track whether it had been done at the database level. If we tracked it
>> in, say, a new pg_class column then in principle you could protect
>> against XID wrap with only table-at-a-time VACUUMs.
> Good, I'm glad I'm wrong on this. This will be another nice advantage
> of autovacuum then and should be fairly easy to do. Any thoughts on
> this being a change we can get in for 8.1?
I'd say this is probably a tad too late --- there's a fair amount of
code change that would be needed, none of which has been written, and
we are past the feature-freeze deadline for new code.
regards, tom lane
On Tue, Jul 05, 2005 at 01:00:50PM -0400, Tom Lane wrote: > "Matthew T. O'Connor" <matthew@zeut.net> writes: > > Tom Lane wrote: > >> No, you're wrong. VACUUMing of individual tables is perfectly good > >> enough as far as XID wrap protection goes, it's just that we chose to > >> track whether it had been done at the database level. If we tracked it > >> in, say, a new pg_class column then in principle you could protect > >> against XID wrap with only table-at-a-time VACUUMs. > > > Good, I'm glad I'm wrong on this. This will be another nice advantage > > of autovacuum then and should be fairly easy to do. Any thoughts on > > this being a change we can get in for 8.1? > > I'd say this is probably a tad too late --- there's a fair amount of > code change that would be needed, none of which has been written, and > we are past the feature-freeze deadline for new code. Right. I've written a small, non-intrusive patch that handles the Xid wraparound just as pg_autovacuum used to, checking the Xid from pg_database. -- Alvaro Herrera (<alvherre[a]alvh.no-ip.org>) "Hay quien adquiere la mala costumbre de ser infeliz" (M. A. Evans)
TODO item? --------------------------------------------------------------------------- Alvaro Herrera wrote: > On Tue, Jul 05, 2005 at 01:00:50PM -0400, Tom Lane wrote: > > "Matthew T. O'Connor" <matthew@zeut.net> writes: > > > Tom Lane wrote: > > >> No, you're wrong. VACUUMing of individual tables is perfectly good > > >> enough as far as XID wrap protection goes, it's just that we chose to > > >> track whether it had been done at the database level. If we tracked it > > >> in, say, a new pg_class column then in principle you could protect > > >> against XID wrap with only table-at-a-time VACUUMs. > > > > > Good, I'm glad I'm wrong on this. This will be another nice advantage > > > of autovacuum then and should be fairly easy to do. Any thoughts on > > > this being a change we can get in for 8.1? > > > > I'd say this is probably a tad too late --- there's a fair amount of > > code change that would be needed, none of which has been written, and > > we are past the feature-freeze deadline for new code. > > Right. I've written a small, non-intrusive patch that handles the Xid > wraparound just as pg_autovacuum used to, checking the Xid from > pg_database. > > -- > Alvaro Herrera (<alvherre[a]alvh.no-ip.org>) > "Hay quien adquiere la mala costumbre de ser infeliz" (M. A. Evans) > > ---------------------------(end of broadcast)--------------------------- > TIP 4: Don't 'kill -9' the postmaster > -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073
I think so. Something like: Improve autovacuum xid wraparound detection by moving to a pertable solution rather than per database. Matt Bruce Momjian wrote: >TODO item? > > >
Matthew T. O'Connor wrote: > I think so. Something like: Improve autovacuum xid wraparound detection > by moving to a pertable solution rather than per database. Thanks, added. -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001 + If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania 19073
While we are at it (assuming the autovacuum patch gets included in 8.1) we have a few autovacuum related todo items. These are the ones I can think of right now: * XID Wraparound improvement, moving to per-table vacuuming rather than per database. (8.2) * Alter table commands to set per table autovacuum threshold settings. (8.2) * Incorporate FSM data to improve vacuum decision making (8.2) * Deal with stats reset better, possibly force stats reset to false if autovacuum is enabled. (8.2) * Add the concept of a maintenance window to autovacuum. (maybe 8.1?) * Have the VACUUM and ANALYZE commands update the pg_autovacuum table itself. This will allow autovacuum to work in harmony with manually issued VACUUM's. (I would like to see this done for 8.1, but I will understand if people demand that it wait for 8.2) * Add some regression tests? Not sure what would be appropriate here. (8.1) * Improve autovacuum threshold defaults (8.1) Anyone have anything to add to the list? Matthew Bruce Momjian wrote: >TODO item? >