Обсуждение: vacuumdb vs. max_connections: SELECT waiting
Tonight as part of a scheduled maintenance operation, I was going to perform a VACUUM FULL ANALYZE on a postgres 8.1.3 instance that had just had it's FSM settings increased to account for about 2 years' worth of growth (particularly in number of relations). Shortly after I kicked it off, I watched the number of connections trend upward as a result of the aggressive locking of FULL. I didn't want to let this continue without notifying the developers about a potential downtime for their application, so I killed the vacuumdb process I had started, figuring that the existing connections would go right through. Unfortunately, my expectations were not met, and I wound up with a bunch of connections in a waiting state. It seems as though the connections that were waiting for the VACUUM locks were still waiting even once the VACUUM was killed. Is that expected behavior? Eventually, a timeout threshold must've been hit because everything went back to normal, but it was not immediate. statement_timeout is set to 0, so it wasn't that. This postmaster does have autovacuum enabled. Could that have interfered in any way? On the one hand I'm curious about the behavior of postgres in this scenario. On the other hand, though, I'm wondering if this operation is necessary. The postmaster has been restarted such that the new FSM settings are in effect. Can autovacuum recover for the months where the FSM settings were not sufficient to cover the number of relations in this cluster? Is a vacuumdb/reindexdb cycle necessary to reclaim disk space? -- Thomas F. O'Connell Database Architecture and Programming Sitening, LLC http://www.sitening.com/ 3004 B Poston Avenue Nashville, TN 37203-1314 615-260-0005 (cell) 615-469-5150 (office) 615-469-5151 (fax)
"Thomas F. O'Connell" <tfo@sitening.com> writes: > Shortly after I kicked it off, I watched the number of connections > trend upward as a result of the aggressive locking of FULL. I didn't > want to let this continue without notifying the developers about a > potential downtime for their application, so I killed the vacuumdb > process I had started, figuring that the existing connections would > go right through. Killing the vacuumdb client process wouldn't automatically abort the current operation of the connected backend. If that's all you did, it would've finished out the VACUUM command, ie, essentially no effect. Sending a SIGINT to the backend to cancel the current command would've been the right thing to do. > Can autovacuum recover for the months where > the FSM settings were not sufficient to cover the number of relations > in this cluster? Probably not. But VACUUM FULL isn't really a great choice either; it will certainly do nothing at all for index bloat. You might try CLUSTER instead if you need to aggressively recover space. regards, tom lane
On May 13, 2006, at 12:35 AM, Tom Lane wrote: > VACUUM FULL does all right at packing the table (except in > pathological > cases, eg a very large tuple near the end of the table). It mostly > bites as far as shrinking indexes goes, however. If you've got a > serious index bloat problem then REINDEX is the only solution. > CLUSTER > does an automatic REINDEX after compacting the table --- it doesn't > have > any special properties as far as the index space goes. However, if > you've got serious table bloat then CLUSTER will probably be quicker > than VACUUM FULL. So my reading of VACUUM VERBOSE output leads me to believe that I could actually interpret both table and index bloat for the entire database from it (and suggests a useful reporting tool...). This is from the docs: INFO: index "onek_unique1" now contains 1000 tuples in 14 pages DETAIL: 3000 index tuples were removed. 0 index pages have been deleted, 0 are currently reusable. Does this imply that 75% of the tuples in this index were free space? Even so, since this is an index, the pages aren't shrunk, per your note above, and a REINDEX would be required for reclamation, right? And, then, later: INFO: "onek": removed 3000 tuples in 108 pages DETAIL: CPU 0.01s/0.06u sec elapsed 0.07 sec. INFO: "onek": found 3000 removable, 1000 nonremovable tuples in 143 pages DETAIL: 0 dead tuples cannot be removed yet. Which implies that this table was also 75% bloated? And a VACUUM FULL (or CLUSTER) could improve this even more than the VACUUM VERBOSE ANALYZE from the example in the docs? >> What is the likely performance impact of having database (table or >> index) bloat from several months in a > 100 GB database with tens of >> thousands of relations of wildly varying sizes and insufficient FSM >> settings? If autovacuum is keeping up with statistics for index >> usage, is the only potential impact related to disk usage (in basic >> filesystem terms)? > > Unless you have a lot of seqscan-using queries, there's no particular > reason to panic over file bloat that I can see. It's just a matter of > how close you are to running out of disk space ... Unfortunately, there are several seqscan-using queries and several large tables (with the largest currently approaching 13% of physical memory). I'm actually trying to prioritize administrative operations that could result in noticeable performance gains. If compacting tables and indexes turns out to be low on the pole in terms of performance considerations, then I'm inclined to look elsewhere, especially considering the administrative headache (from the perspective of the related application) required to undertake a database-wide CLUSTER or VACUUM FULL + REINDEX. -- Thomas F. O'Connell Database Architecture and Programming Sitening, LLC http://www.sitening.com/ 3004 B Poston Avenue Nashville, TN 37203-1314 615-260-0005 (cell) 615-469-5150 (office) 615-469-5151 (fax)
On Sat, May 13, 2006 at 03:47:00AM -0500, Thomas F. O'Connell wrote: > > On May 13, 2006, at 12:35 AM, Tom Lane wrote: > > >VACUUM FULL does all right at packing the table (except in > >pathological > >cases, eg a very large tuple near the end of the table). It mostly > >bites as far as shrinking indexes goes, however. If you've got a > >serious index bloat problem then REINDEX is the only solution. > >CLUSTER > >does an automatic REINDEX after compacting the table --- it doesn't > >have > >any special properties as far as the index space goes. However, if > >you've got serious table bloat then CLUSTER will probably be quicker > >than VACUUM FULL. > > So my reading of VACUUM VERBOSE output leads me to believe that I > could actually interpret both table and index bloat for the entire > database from it (and suggests a useful reporting tool...). > > This is from the docs: > > INFO: index "onek_unique1" now contains 1000 tuples in 14 pages > DETAIL: 3000 index tuples were removed. > 0 index pages have been deleted, 0 are currently reusable. > > Does this imply that 75% of the tuples in this index were free space? > Even so, since this is an index, the pages aren't shrunk, per your > note above, and a REINDEX would be required for reclamation, right? Correct. > And, then, later: > > INFO: "onek": removed 3000 tuples in 108 pages > DETAIL: CPU 0.01s/0.06u sec elapsed 0.07 sec. > INFO: "onek": found 3000 removable, 1000 nonremovable tuples in 143 > pages > DETAIL: 0 dead tuples cannot be removed yet. > > Which implies that this table was also 75% bloated? And a VACUUM FULL > (or CLUSTER) could improve this even more than the VACUUM VERBOSE > ANALYZE from the example in the docs? Also correct. There are some tricks you can do to shrink the table without resorting to a vacuum full or a cluster though, but it's a real PITA. Also, remember that if your database is growing it will eventually end up reusing all that free space. > Unfortunately, there are several seqscan-using queries and several > large tables (with the largest currently approaching 13% of physical > memory). So are you saying that the entire table fits in memory? Does the entire database fit in memory? If so, you're unlikely to find huge gains from vacuuming, unless you're really pushing the system hard. > I'm actually trying to prioritize administrative operations that > could result in noticeable performance gains. If compacting tables > and indexes turns out to be low on the pole in terms of performance > considerations, then I'm inclined to look elsewhere, especially > considering the administrative headache (from the perspective of the > related application) required to undertake a database-wide CLUSTER or > VACUUM FULL + REINDEX. It depends. On badly bloated systems I've seen 20-50% improvements from setting up an adequate vacuum plan. -- Jim C. Nasby, Sr. Engineering Consultant jnasby@pervasive.com Pervasive Software http://pervasive.com work: 512-231-6117 vcard: http://jim.nasby.net/pervasive.vcf cell: 512-569-9461