Обсуждение: Re: [HACKERS] Performance testing of COPY (SELECT) TO

> Böszörményi Zoltán <zboszor@dunaweb.hu> writes:
>> With PostgreSQL 8.1.4, I used this:
>
>> begin;
>> select ... into temp myquery1;
>> copy myquery1 to stdout csv delimiter '|';
>> rollback;
>
> The performance of this would doubtless vary a lot with the temp_buffers
> setting.  Did you try different values?

Yes, I did, but now checked back with 8.2CVS.
The previously quoted result was achieved with
temp_buffers = 1000 on both 8.1.4 and 8.2CVS.
On 8.2CVS with temp_buffers = 4096, the 10 client case kills
the machine with swapping, but the 3 client runtime with
COPY(SELECT) went down to 2:41. The SELECT INTO TEMP
case went down to 3:36.

> It'd also be interesting to time the same way (with a temp table) in
> devel.  I don't remember whether we did any performance work on the
> COPY CSV data path in this cycle, or whether that was all present in
> 8.1.  In any case it'd be worth proving that the COPY SELECT patch isn't
> degrading performance of the copy-a-relation case.

I will report back with that, say on Monday.

In the meantime, I documented the COPY (SELECT) case
and modified parser/analyze.c and tcop/utility.c so neither of them
calls anything from under another directory. I think it's cleaner now.
Also, I tried to implement more closely what printtup() does.
Please, review.

Best regards,
Zoltán Böszörményi