Re: Weird CASE WHEN behaviour causing query to be suddenly very slow

Touche ! Thanks a lot.

Looking more at the data yes it goes very often to ELSE Clause.  And
therefore reaching  the MAX_CACHED_RES.

In there anyway to increase that value  ?

Basically, I have several tables containing millions of rows and let say 5
columns. Those five columns, depending of their  combination give me a 6th
value.
We have complex patterns to match and using simple LIKE / EQUAL and so on
wouldn't be enough. This can be applied to N number of table so we
refactored this process into a function that we can use in the SELECT
statement, by giving only the 5 values each time.

I wouldn't mind using a table and mapping it through a join  if it were for
my own use.
But the final query has to be readable and usable for almost-non-initiated
SQL user... So using a function with encapsulated case when seemed to be a
good idea and so far worked nicely.

But we might consider changing it if we have no other choice...

Regards,

Kevin

-----Original Message-----
From: Tom Lane [mailto:tgl@sss.pgh.pa.us]
Sent: Dienstag, 31. März 2015 15:59
To: Kevin Viraud
Cc: pgsql-performance@postgresql.org
Subject: Re: [PERFORM] Weird CASE WHEN behaviour causing query to be
suddenly very slow

"Kevin Viraud" <kevin.viraud@rocket-internet.de> writes:
> I have an issue with a rather large CASE WHEN and I cannot figure out
> why it is so slow...

Do all the arms of the CASE usually fail, leaving you at the ELSE?

I suspect what's happening is that you're running into the MAX_CACHED_RES
limit in src/backend/utils/adt/regexp.c, so that instead of just compiling
each regexp once and then re-using 'em, the regexps are constantly falling
out of cache and then having to be recompiled.  They'd have to be used in a
nearly perfect round robin in order for the behavior to have such a big
cliff as you describe, though.  In this CASE structure, that suggests that
you're nearly always testing every regexp because they're all failing.

I have to think there's probably a better way to do whatever you're trying
to do, but there's not enough info here about your underlying goal to
suggest a better approach.  At the very least, if you need a many-armed
CASE, it behooves you to make sure the common cases appear early.

            regards, tom lane