Re: [COMMITTERS] pgsql: Write a WAL record whenever we perform an operation without

Heikki Linnakangas wrote:
> Simon Riggs wrote:
>> On Mon, 2010-02-01 at 12:11 +0200, Heikki Linnakangas wrote:
>>> Simon Riggs wrote:
>>>> On Mon, 2010-02-01 at 11:33 +0200, Heikki Linnakangas wrote:
>>>>> Can you reproduce that?
>>>> Every time
>>> Ok, good :-). What I really meant was: How?
>> As mentioned in first post: VACUUM FULL and CLUSTER.
>
> I'm asking because I couldn't reproduce it. I added an elog(LOG) to
> XLogReportUnloggedStatement() so that I can see when an unlogged-record
> is written. Then I ran:
>
> CREATE TEMPORARY TABLE foo (id int4);
> VACUUM FULL foo;
>
> But XLogReportUnloggedStatement() was not executed. When I removed the
> "TEMPORARY" and tried again, it was executed.

No, hang on. When I run this:

postgres=# CREATE TABLE foo (id int4);
CREATE TABLE
postgres=# vacuum full foo;
VACUUM

I get this in the log:

LOG: Writing record: CLUSTER on "pg_temp_24677"
STATEMENT: vacuum full foo;
LOG: Writing record: heap inserts on "pg_temp_24677"
STATEMENT: vacuum full foo;

So you get those messages when the table is *not* a temporary table. I
can see now what Fujii was trying to say. His patch seems Ok, though
perhaps it would be better to move the responsibility of calling
XLogReportUnloggedStatement() to the callers of heap_sync(). When I put
it in heap_sync(), I didn't take into account that it's sometimes called
just to flush buffers from buffer cache, not to fsync() non-WAL-logged
operations.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com