On Sat, Jan 14, 2012 at 11:17 AM, Noah Misch <noah@leadboat.com> wrote:
> On Sat, Jan 14, 2012 at 08:08:29AM +0000, Simon Riggs wrote:
>> On Sat, Jan 14, 2012 at 1:02 AM, Noah Misch <noah@leadboat.com> wrote:
>> > However, CLUSTER/VACUUM FULL already
>> > remove tuples still-visible to standby snapshots without provoking a recovery
>> > conflict. ?(Again only with hot_standby_feedback=off.)
>>
>> If that were the case it would be a bug.
>>
>> CLUSTER/VACUUM FULL emit an AccessExclusiveLock record that would
>> conflict with any current lock holders, so should be fine on that.
>
> I speak of this sequence (M = master connection, S = standby connection):
>
> M: CREATE TABLE t AS SELECT * FROM generate_series(1,1000) t(n);
> S: BEGIN ISOLATION LEVEL REPEATABLE READ; SELECT 0;
> M: DELETE FROM t WHERE n <= 10;
> M: VACUUM FULL t;
> S: SELECT count(*) FROM t; -- 990, should be 1000
OK, so we need to emit a heap_xlog_cleanup_info() record at the end of
cluster to conflict with anybody that doesn't yet have a lock but has
a snapshot that can see tuples the cluster implicitly removed. Will
do.
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
