I'm looking at the most recent version of the Hot Standby patch at
Robert Haas' GIT repository. The conflict cache code is broken:
> +void
> +SetDeferredRecoveryConflicts(TransactionId latestRemovedXid, RelFileNode node,
> + XLogRecPtr conflict_lsn)
> +{
> + ProcArrayStruct *arrayP = procArray;
> + int index;
> + Oid dbOid = node.dbNode;
> +
> + Assert(InRecovery);
> +
> + if (!LatestRemovedXidAdvances(latestRemovedXid))
> + return;
> +
The idea of LatestRemoveXidAdvances() is to exit quickly when we're
called with a latestRemovedXid value <= the previous latestRemovedXid
value. However, the conflict caches store information per relation. If
you first call e.g "SetDeferredRecoveryConflicts(1000, 'rel_A', 1234)",
followed by "SetDeferredRecoveryConflicts(1000, 'rel_B', 1234)", the
latter call exits quickly. If a transaction that holds a "too old"
snapshot then accesses rel_B, it won't fail as it should.
Something else must be severly broken in the conflict resolution code as
well: while testing with just one tiny table, I can easily reproduce a
violation of serializable snapshot:
postgres=# begin ISOLATION LEVEL serializable;
BEGIN
postgres=# SELECT * FROM foo;id
-----101102
(2 rows)
(In master: UPDATE foo SET id = id + 10; VACUUM foo; SELECT
pg_xlog_switch())
postgres=# SELECT * FROM foo;id
----
(0 rows)
And it looks like the recovery cache is not reset properly: when I start
a new backend after one that just got a "canceling statement due to
recent buffer changes during recovery" error, and run a query, I get
that error again:
psql (8.5devel)
Type "help" for help.
postgres=# SELECT * FROM foo;
postgres=# begin ISOLATION LEVEL serializable;
BEGIN
postgres=# SELECT * FROM foo;
ERROR: canceling statement due to recent buffer changes during recovery
I haven't dug deeper into those, but before I do, I want to ask if we
really need to bother with a per-relation conflict cache at all? I'd
really like to keep it simple for now, and tracking the conflicts
per-relation only alleviates the situation somewhat. The nature of the
cache is such that it's quite unpredictable to a regular user when it
will save you, so you can't rely on it. You need to set
max_standby_delay and/or other such settings correctly anyway, so it
doesn't really help with usability.
Another thing:
I'm quite surprised to see that the logic in WAL redo to stop the redo
and wait for read-only queries to finish before applying a WAL record
that would cause conflicts, and thus cause a read-only query to be
killed, is only used with a few WAL record types like database or
tablespace creation. Not the usual VACUUM records. I was under the
impression that max_standby_delay option and logic would apply to all
operations.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com