Simon Riggs <simon@xxxxxxxxxxxxxxx> writes: > You'll find response times spike somewhat during Checkpoints. We've been > working to reduce this effect for some time and there is still more work > to be done. > > Tune your wal settings and bgwriter to improve things, plus upgrade to > 8.1 which has made leaps forward with this issue. If that's the problem presumably moving the WAL segments to a dedicated set of spindles would help? It seems to me it would take a pretty severe case of this to cause 10-20s delays though. You aren't perchance running VACUUM FULL or any DDL like REINDEX or CLUSTER periodically are you? -- greg
