From: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
---|---|
To: | Daniel Westermann <daniel(dot)westermann(at)dbi-services(dot)com> |
Cc: | Merlin Moncure <mmoncure(at)gmail(dot)com>, PostgreSQL General <pgsql-general(at)postgresql(dot)org> |
Subject: | Re: Query never completes with low work_mem (at least not within one hour) |
Date: | 2017-04-05 14:04:50 |
Message-ID: | 6343.1491401090@sss.pgh.pa.us |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-general |
Daniel Westermann <daniel(dot)westermann(at)dbi-services(dot)com> writes:
> Thank you, Merlin. As said I know that "not in" is not a good choice in this case but I still do not get what is going here. Why does the server repeatedly search for NULL values when I decrease work_mem and why not when increasing work_mem?
The core point is that one plan is using a hashed subplan and the other is
not, because the planner estimated that the hashtable wouldn't fit into
work_mem. With a hashtable you'll have one probe into the hashtable per
outer row, and each probe is O(1) unless you are unlucky about data
distributions, so the runtime is more or less linear. Without a
hashtable, the inner table is rescanned for each outer row, so the
runtime is O(N^2) which gets pretty bad pretty fast. "Materializing"
the inner table doesn't really help: it gets rid of per-inner-row
visibility checks and some buffer locking overhead, so it cuts the
constant factor some, but the big-O situation is still disastrous.
regards, tom lane
From | Date | Subject | |
---|---|---|---|
Next Message | vinny | 2017-04-05 14:14:04 | Re: browser interface to forums please? |
Previous Message | Vincent Veyron | 2017-04-05 13:11:36 | Re: browser interface to forums please? |