Postgres: “vacuum” command does not clean up dead

问题:

We have a postgres database in Amazon RDS. Initially, we needed to load large amount of data quickly, so autovacuum was turned off according to the best practice suggestion from Amazon. Recently I noticed some performance issue when running queries. Then I realized it has not been vacuumed for a long time. As it turns out many tables have lots of dead tuples.

Surprisingly, even after I manually ran vacuum commands on some of the tables, it did not seem to remove these dead tuples at all. vacuum full takes too long to finish which usually ends up timed out after a whole night.

Why does vacuum command not work? What are my other options, restart the instance?

回答1:

Use VACUUM (VERBOSE) to get detailed statistics of what it is doing and why.

There are three reasons why dead tuples cannot be removed:

1. There is a long running transaction that has not been closed. You can find the bad boys with

   SELECT pid, datname, usename, state, backend_xmin
   FROM pg_stat_activity
   WHERE backend_xmin IS NOT NULL
   ORDER BY age(backend_xmin) DESC;
   

   You can get rid of a transaction with pg_cacnel_backend() or pg_terminate_backend().

2. There are prepared transactions which have not been commited. You can find them with

   SELECT gid, prepared, owner, database, transaction
   FROM pg_prepared_xacts
   ORDER BY age(transaction) DESC;
   

   User COMMIT PREPARED or ROLLBACK PREPARED to close them.

3. There are replication slots which are not used. Find them with

   SELECT slot_name, slot_type, database, xmin
   FROM pg_replication_slots
   ORDER BY age(xmin) DESC;
   

   Use pg_drop_replication_slot() to delete an unused replication slot.

回答2:

https://dba.stackexchange.com/a/77587/30035 explains why not all dead tuples are removed.

for vacuum full not to time out, set statement_timeout = 0

http://docs.aws.amazon.com/AmazonRDS/latest/UserGuide/CHAP_BestPractices.html#CHAP_BestPractices.PostgreSQL recommends disabling autovacuum for the time of database restore, further they definetely recommend using it:

Important

Not running autovacuum can result in an eventual required outage to perform a much more intrusive vacuum operation.

Canceling all sessions and vacuuming table should help with previous dead tuples (regarding your suggestion to restart cluster). But what I suggest you to do in first place - switch autovacuum on. And better probably control vacuum on table, not on the whole cluster with autovacuum_vacuum_threshold, (ALTER TABLE) reference here: https://www.postgresql.org/docs/current/static/sql-createtable.html#SQL-CREATETABLE-STORAGE-PARAMETERS