VERSION 2, with all changed made as requested to date.
As discussed on -hackers, its possible to avoid writing any WAL at all
for COPY in these circumstances:
http://archives.postgresql.org/pgsql-hackers/2006-10/msg01172.php
and again recently.
BEGIN;
CREATE TABLE foo..
COPY foo...
COMMIT;
BEGIN;
TRUNCATE foo..
COPY foo...
COMMIT;
The enclosed patch implements this, as discussed. There is no user
interface to enable/disable, just as with CTAS and CREATE INDEX; no
docs, just code comments.
This plays nicely with the --single-transaction option in psql to allow
fast restores/upgrades.
YMMV but disk bound COPY will benefit greatly from this patch, some
tests showing 100% gain. COPY is still *very* CPU intensive, so some
tests have shown negligible benefit, fyi, but that isn't the typical
case.
While testing this, I realised something: small COPY commands get no
benefit at all, but larger ones do. When we do a small normal COPY the
data stays in cache, but the WAL is written to disk and fsynced. When we
do a small fast COPY, no WAL is written, but the data is written to disk
and fsynced. With COPY, WAL and data are roughly same size, hence no I/O
benefit. With larger COPY statements, benefit is very substantial.
Applies cleanly to CVS HEAD, passes make check.
I enclose a test case that shows whether the test has succeeded by
reading the WAL Insert pointer before/after each COPY. This has been
written in such a way that we could, if we wanted to, include a new
regression test for this. There is a function that returns an immutable
value if the test passes, rather than simply showing the WAL insert
pointer which would obviously vary between tests. The tests enclosed
here *also* include the WAL insert pointer so you can manually/visibly
see that the enclosed patch writes no WAL at appropriate times.
psql -f copy_nowal_prep.sql postgres
psql -f copy_nowal_test.sql postgres
Do we want an additional test case along these lines?
Agreed doc changes for Performance Tips forthcoming.
--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com