E.18. Release 9.3.9

E.18.1. Migration to Version 9.3.9

A dump/restore is not required for those running 9.3.X.

However, if you are upgrading an installation that was previously upgraded using a pg_upgrade version between 9.3.0 and 9.3.4 inclusive, see the first changelog entry below.

Also, if you are upgrading from a version earlier than 9.3.7, see Section E.20.

E.18.2. Changes

• Fix possible failure to recover from an inconsistent database state (Robert Haas)

  Recent PostgreSQL releases introduced mechanisms to protect against multixact wraparound, but some of that code did not account for the possibility that it would need to run during crash recovery, when the database may not be in a consistent state. This could result in failure to restart after a crash, or failure to start up a secondary server. The lingering effects of a previously-fixed bug in pg_upgrade could also cause such a failure, in installations that had used pg_upgrade versions between 9.3.0 and 9.3.4.

  The pg_upgrade bug in question was that it would set oldestMultiXid to 1 in pg_control even if the true value should be higher. With the fixes introduced in this release, such a situation will result in immediate emergency autovacuuming until a correct oldestMultiXid value can be determined. If that would pose a hardship, users can avoid it by doing manual vacuuming before upgrading to this release. In detail:

  1. Check whether pg_controldata reports "Latest checkpoint's oldestMultiXid" to be 1. If not, there's nothing to do.

  2. Look in PGDATA/pg_multixact/offsets to see if there's a file named 0000. If there is, there's nothing to do.

  3. Otherwise, for each table that has pg_class.relminmxid equal to 1, VACUUM that table with both vacuum_multixact_freeze_min_age and vacuum_multixact_freeze_table_age set to zero. (You can use the vacuum cost delay parameters described in Section 18.4.4 to reduce the performance consequences for concurrent sessions.) You must use PostgreSQL 9.3.5 or later to perform this step.

• Fix rare failure to invalidate relation cache init file (Tom Lane)

  With just the wrong timing of concurrent activity, a VACUUM FULL on a system catalog might fail to update the "init file" that's used to avoid cache-loading work for new sessions. This would result in later sessions being unable to access that catalog at all. This is a very ancient bug, but it's so hard to trigger that no reproducible case had been seen until recently.

• Avoid deadlock between incoming sessions and CREATE/DROP DATABASE (Tom Lane)

  A new session starting in a database that is the target of a DROP DATABASE command, or is the template for a CREATE DATABASE command, could cause the command to wait for five seconds and then fail, even if the new session would have exited before that.

• Improve planner's cost estimates for semi-joins and anti-joins with inner indexscans (Tom Lane, Tomas Vondra)

  This type of plan is quite cheap when all the join clauses are used as index scan conditions, even if the inner scan would nominally fetch many rows, because the executor will stop after obtaining one row. The planner only partially accounted for that effect, and would therefore overestimate the cost, leading it to possibly choose some other much less efficient plan type.