Re: EXPLAIN BUFFERS

Robert Haas <robertmhaas@gmail.com> wrote:

> Well, I think we need to do something.  I don't really want to tack
> another 5-6% overhead onto EXPLAIN ANALYZE.  Maybe we could recast the
> doInstrument argument as a set of OR'd flags?

I'm thinking the same thing (OR'd flags) right now.

The attached patch adds INSTRUMENT_TIMER and INSTRUMENT_BUFFERS flags.
The types of QueryDesc.doInstrument (renamed to instrument_options) and
EState.es_instrument are changed from bool to int, and they store
OR of InstrumentOption flags. INSTRUMENT_TIMER is always enabled when
instrumetations are initialized, but INSTRUMENT_BUFFERS is enabled only if
we use EXPLAIN BUFFERS. I think the flag options are not so bad idea because
of extensibility. For example, we could support EXPLAIN CPU_USAGE someday.

One issue is in the top-level instrumentation (queryDesc->totaltime).
Since the field might be used by multiple plugins, the first initializer
need to initialize the counter with all options. I used INSTRUMENT_ALL
for it in the patch.

=# EXPLAIN (ANALYZE) SELECT * FROM pgbench_accounts;
                                                           QUERY PLAN

--------------------------------------------------------------------------------------------------------------------------------
 Seq Scan on pgbench_accounts  (cost=0.00..263935.00 rows=10000000 width=97) (actual time=0.003..572.126 rows=10000000
loops=1)
 Total runtime: 897.729 ms

=# EXPLAIN (ANALYZE, BUFFERS) SELECT * FROM pgbench_accounts;
                                                           QUERY PLAN

--------------------------------------------------------------------------------------------------------------------------------
 Seq Scan on pgbench_accounts  (cost=0.00..263935.00 rows=10000000 width=97) (actual time=0.002..580.642 rows=10000000
loops=1)
   Buffers: shared hit=163935
 Total runtime: 955.744 ms

Regards,
---
Takahiro Itagaki
NTT Open Source Software Center