If you suspect corruption of an index on a user table, you can simply rebuild that index, or all indexes on the table, using REINDEX INDEX or REINDEX TABLE.
Things are more difficult if you need to recover from corruption of an index on a system table. In this case it's important for the system to not have used any of the suspect indexes itself. (Indeed, in this sort of scenario you may find that server processes are crashing immediately at start-up, due to reliance on the corrupted indexes.) To recover safely, the server must be started with the -P option, which prevents it from using indexes for system catalog lookups.
One way to do this is to shut down the postmaster and start a stand-alone PostgreSQL server with the -P option included on its command line. Then, REINDEX DATABASE, REINDEX SYSTEM, REINDEX TABLE, or REINDEX INDEX can be issued, depending on how much you want to reconstruct. If in doubt, use REINDEX SYSTEM to select reconstruction of all system indexes in the database. Then quit the standalone server session and restart the regular server. See the
reference page for more information about how to interact with the stand-alone server interface.
Alternatively, a regular server session can be started with -P included in its command line options. The method for doing this varies across clients, but in all libpq-based clients, it is possible to set the PGOPTIONS environment variable to -P before starting the client. Note that while this method does not require locking out other clients, it may still be wise to prevent other users from connecting to the damaged database until repairs have been completed.
If corruption is suspected in the indexes of any of the shared system catalogs (which are pg_authid, pg_auth_members, pg_database, pg_pltemplate, pg_shdepend, and pg_tablespace), then a standalone server must be used to repair it. REINDEX will not process shared catalogs in multiuser mode.
For all indexes except the shared system catalogs, REINDEX is crash-safe and transaction-safe. REINDEX is not crash-safe for shared indexes, which is why this case is disallowed during normal operation. If a failure occurs while reindexing one of these catalogs in standalone mode, it will not be possible to restart the regular server until the problem is rectified. (The typical symptom of a partially rebuilt shared index is "index is not a btree" errors.)
REINDEX is similar to a drop and recreate of the index in that the index contents are rebuilt from scratch. However, the locking considerations are rather different. REINDEX locks out writes but not reads of the index's parent table. It also takes an exclusive lock on the specific index being processed, which will block reads that attempt to use that index. In contrast, DROP INDEX momentarily takes exclusive lock on the parent table, blocking both writes and reads. The subsequent CREATE INDEX locks out writes but not reads; since the index is not there, no read will attempt to use it, meaning that there will be no blocking but reads may be forced into expensive sequential scans. Another important point is that the drop/create approach invalidates any cached query plans that use the index, while REINDEX does not.
Reindexing a single index or table requires being the owner of that index or table. Reindexing a database requires being the owner of the database (note that the owner can therefore rebuild indexes of tables owned by other users). Of course, superusers can always reindex anything.
Prior to PostgreSQL 8.1, REINDEX DATABASE processed only system indexes, not all indexes as one would expect from the name. This has been changed to reduce the surprise factor. The old behavior is available as REINDEX SYSTEM.
Prior to PostgreSQL 7.4, REINDEX TABLE did not automatically process TOAST tables, and so those had to be reindexed by separate commands. This is still possible, but redundant.