On Wed, Sep 21, 2016 at 8:33 PM, Peter Eisentraut
<peter.eisentraut@2ndquadrant.com> wrote:
> There is apparently some misbehavior if max_wal_size is less than 5 *
> wal_segment_size.
>
> For example, if you build with --with-wal-segsize=64, then the recovery
> test fails unless you set max_wal_size to at least 320MB in
> PostgresNode.pm. The issue is that pg_basebackup fails with:
>
In recovery tests, max_wal_size is set to 128MB. Now, when you build
with --with-wal-segsize=64,
max_wal_size is calculated as follows:
max_wal_size = 128 / (64 * 1024 * 1024) / (1024 * 1024) = 2.
and CheckPointSegments is calculated as follows:
CheckPointSegments = 2 / (2 + 0.5) = 0.8 rounded to 1. (Default is 3)
Hence, checkpoints occurs very frequently at master.
> pg_basebackup: could not get transaction log end position from server:
> ERROR: could not find any WAL files
This error occurs when the recovery test tries to take backup from the
standby using the
above settings. pg_basebackup scans pg_xlog and include all WAL files
in the range
between 'startptr' and 'endptr', regardless of the timeline the file
is stamped with.
'startptr' is initialized to ControlFile->checkPointCopy.redo and
'endptr' is initialized to
ControlFile->minRecoveryPoint.
Now, whenever we redo a CHECKPOINT_ONLINE log, we update checkPointCopy.redo and
whenever we flush logs, we update minRecoveryPoint.
In this case, we are having frequent checkpoints at master which in
turn updates checkPointCopy.redo
in standy frequently. Sometimes, it even goes ahead of
minRecoveryPoint. At this point, if you call
pg_basebackup, it will throw the aforesaid error.
> This should probably be made friendlier in some way. But it also shows
> that bigger WAL segment sizes are apparently not well-chartered
> territory lately.
>
Well, there can be multiple solutions to this problem.
1. If somebody intends to increase wal segment size, he should
increase max_wal_size accordingly.
2. In recovery test, we can add some delay before taking backup so
that the pending logs in the buffer
gets flushed. (Not a good solution)
3. In CreateRestartPoint() method, we can force a XLogFlush to update
minRecoveryPoint.
Thoughts?
--
Thanks & Regards,
Kuntal Ghosh
EnterpriseDB: http://www.enterprisedb.com