Re: [HACKERS] auto_explain : log queries with wrong estimation

On 24.08.2017 14:56, Adrien Nayrat wrote:

Hi hackers,

Hi,

I try to made a patch to auto_explain in order to log queries with wrong estimation.

I compare planned row id : queryDesc->planstate->plan->plan_rows

Vs ntuples : queryDesc->planstate->instrument->ntuples;

AFAICS you want to introduce two additional per-node variables:
 - auto_explain_log_estimate_ratio that denotes minimum ratio (>= 1) between real value and planned one. I would add 'min' prefix before 'ratio'.
 - auto_explain_log_estimate_min_rows - minimum absolute difference between those two values. IMHO this name is somewhat poor, the suffix 'min_diff_rows' looks better.
If real expressions (ratio and diff) exceed these threshold values both, you log this situation. I'm right?

If I understand, instrumentation is used only with explain. So my patch works
only with explain (and segfault without).

Instrumentation is initialized only with analyze (log_analyze is true)[1]

Is there a simple way to get ntuples?

It's interesting question. In one's time I didn't find any way to get the amount of tuples emitted from a node.


1. contrib/auto_explain/auto_explain.c:221

-- 
Regards,
Maksim Milyutin