Under what conditions does EXPLAIN ANALYZE report spilling work_mem to disk? When does it not report work_mem or other overflow to disk? I know that a planned disk-sort shows up. I have also seen it report a hash-agg on disk, but this was a while ago and rather difficult to reproduce and I'm somewhat confident I have seen it spill to temp disk without reporting it in EXPLAIN ANALYZE, but I could be wrong.
On Thu, Sep 18, 2008 at 11:13 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> It's probably writing hint bits to improve performance of subsequent > access to the table. The issue is discussed here: > http://wiki.postgresql.org/wiki/Hint_Bits
Hint-bit updates wouldn't be WAL-logged. If the table has been around a long time, it might be freezing old tuples, which *would* be WAL-logged (since 8.2 or so) --- but that would be a one-time, non-repeatable behavior. How sure are you that there was WAL output?
What I was thinking was more likely was that the hash table for the hash join was spilling out to temp files. That wouldn't be WAL-logged either, but depending on your tablespace setup it might result in I/O on some other disk than the table proper.
