Fix and test snapshot behavior on standby.
authorAndres Freund <[email protected]>
Thu, 1 Oct 2020 00:28:51 +0000 (17:28 -0700)
committerAndres Freund <[email protected]>
Thu, 1 Oct 2020 00:28:51 +0000 (17:28 -0700)
commit7b28913bcab8d1bf3dbf59c9d8fb4b51cef57664
tree9c3d15b5cef2ba02bdd0d23ccc97d7a10e042ca3
parent9fc21227128980f88b6eb6d5306625af0428ef45
Fix and test snapshot behavior on standby.

I (Andres) broke this in 623a9CA79bx, because I didn't think about the
way snapshots are built on standbys sufficiently. Unfortunately our
existing tests did not catch this, as they are all just querying with
psql (therefore ending up with fresh snapshots).

The fix is trivial, we just need to increment the transaction
completion counter in ExpireTreeKnownAssignedTransactionIds(), which
is the equivalent of ProcArrayEndTransaction() during recovery.

This commit also adds a new test doing some basic testing of the
correctness of snapshots built on standbys. To avoid the
aforementioned issue of one-shot psql's not exercising the snapshot
caching, the test uses a long lived psqls, similar to
013_crash_restart.pl. It'd be good to extend the test further.

Reported-By: Ian Barwick <[email protected]>
Author: Andres Freund <[email protected]>
Author: Ian Barwick <[email protected]>
Discussion: https://postgr.es/m/61291ffe-d611-f889-68b5-c298da9fb18f@2ndquadrant.com
src/backend/storage/ipc/procarray.c
src/test/recovery/t/021_row_visibility.pl[new file with mode: 0644]