Обсуждение: Autovacuum potential bug
We have been seeing an issue where reads on a table seem to slowly degrade over time. When we run VACUUM ANALYZE on the table the operations perform much better. It usually takes about a week or two for the reads to degrade to a point where we have to manually run VACUUM ANALYZE. This has been happening to a handful of our tables. The mutation operations ran against these tables are mainly update operations. Very few insert and delete operations are ran against these tables. Also a jsonb object stores most of the data within each record. The size of the jsonb object can vary widely between records. We thought the autovacuum facility would have prevented us from having to periodically run VACUUM ANALYZE. Is this a bug or is it a configuration issue on our tables?
Hi, On Fri, Jan 10, 2020 at 02:13:47PM -0700, maschka, sam wrote: >We have been seeing an issue where reads on a table seem to slowly degrade >over time. When we run VACUUM ANALYZE on the table the operations perform >much better. It usually takes about a week or two for the reads to degrade >to a point where we have to manually run VACUUM ANALYZE. This has been >happening to a handful of our tables. The mutation operations ran against >these tables are mainly update operations. Very few insert and delete >operations are ran against these tables. Also a jsonb object stores most of >the data within each record. The size of the jsonb object can vary widely >between records. We thought the autovacuum facility would have prevented us >from having to periodically run VACUUM ANALYZE. Is this a bug or is it a >configuration issue on our tables? > Unfortunately that's impossible to answer without providing much more detailed information about the issue. This could easily be just a matter of autovacuum not being aggressive enough on those tables, for example. I don't think this qualifies as a bug, certainly not in the current shape. Perhaps we'll discover something, but if I had to guess at this point, I'd say it's likely a configuration issue. I suggest you post a more detailed report to pgsql-performance [1] with much more details. Right of the bat we need to know - PostgreSQL version - size of the database - non-default configuration - what type of queries you're running - how much slower it gets (2x, 10x, 100x?) - does autovacuum run at all on the tables? - does the execution plan change? - if the execution plan does not change, maybe try some CPU profiling There's also [2] which is primarily focused on slow query analysis, but maybe it'll give you ideas what other information to provide. regards [1] https://www.postgresql.org/list/pgsql-performance/ [2] https://wiki.postgresql.org/wiki/Slow_Query_Questions -- Tomas Vondra http://www.2ndQuadrant.com PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services