Re: [HACKERS] Re: [BUGS] General Bug Report: palloc fails with lots of ANDs and ORs - Mailing list pgsql-hackers

From Vadim B. Mikheev
Subject Re: [HACKERS] Re: [BUGS] General Bug Report: palloc fails with lots of ANDs and ORs
Date
Msg-id 34E942DD.5B00A2@sable.krasnoyarsk.su
Whole thread Raw
In response to Re: [BUGS] General Bug Report: palloc fails with lots of ANDs and ORs  (Bruce Momjian <maillist@candle.pha.pa.us>)
Responses Re: [HACKERS] Re: [BUGS] General Bug Report: palloc fails with lots of ANDs and ORs
List pgsql-hackers
> >
> > select count(*) from outlet
> > where
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1) or
> >     (number = 1 and number = 1 and number = 1);
> >
> > Not pretty but it makes the point.   Take out two OR clauses and the query
> > works fine (but a bit slow).
> >
> > The above query is all it takes to use up all 130000 Kbytes of memory.
> > And, since the query takes a long time to finally fail, I was able to
> > observe the memory consumption.

Optimizator tries to transform qual above into AND clause with
3 (# of and-ed clauses) ^ 9 (# of OR-s) = 19683 args (each arg
is OR clause with 9 op. expressions. My estimation for current
cnfify() code is that this will require =~ 500Mb of memory :)
I made little changes - just to free memory when it's possible:

          current code                 with free-ing

6 ORs     14.3 Mb                      4.3 Mb
7 ORs     53 Mb                        10.3 Mb
8 ORs     estimation: ~ 160 Mb         30.6 MB

I'm not sure should I aplly my changes or not - it doesn't fix
problem, just reduces memory impact. It obviously can't help you,
David, in your real example (3 ^ 28 = 22876792454961 clauses - he he :).

Resume: cnfify() makes mathematically strong but in some cases
practically unwise work. I can't fix this for 6.3

Vadim

pgsql-hackers by date:

Previous
From: Brook Milligan
Date:
Subject: results of regression tests: NetBSD/i386 v1.3
Next
From: Zeugswetter Andreas SARZ
Date:
Subject: Re: Subselects open issue Nr. NEW