Richard Yen wrote:
> the pg_xlog partition filled up (33GB)
> checkpoint_segments=16
> a sudden flurry of write activity
Was this database bulk-loaded in some way (like by restoring the
output of pg_dump, for example)? If so, all rows inserted into all
tables would have the same (or very nearly the same) xmin value. At
some later time, virtually all tuples would need to be rewritten to
freeze them. This would be a logged operation, so WAL files would be
getting created rapidly. If you have WAL archiving turned on, and
you can't copy the files out as fast as they're being created, this
might happen.
To avoid such a crushing mass freeze at an unpredictable time, we
always run VACUUM FREEZE ANALYZE on a bulk-loaded database before
turning on WAL archiving.
Of course, if this database wasn't bulk-loaded, this can't be your
problem....
-Kevin