Re: extend pgbench expressions with functions - Mailing list pgsql-hackers

From Fabien COELHO
Subject Re: extend pgbench expressions with functions
Date
Msg-id alpine.DEB.2.10.1507261101310.25082@sto
Whole thread Raw
In response to Re: extend pgbench expressions with functions  (Fabien COELHO <coelho@cri.ensmp.fr>)
Responses Re: extend pgbench expressions with functions
List pgsql-hackers
Hello Heikki,

>> As soon as we add more functions, the way they are documented needs to be 
>> reworked too; we'll need to add a table in the manual to list them.
>
> Here is a v8 with "abs", "min", "max", "random", "gaussian" et
> "exponential".
>
> [...]
>
> There is no real doc, WIP...

Here is a v9 with a doc. The implicit typing of expressions is improved.

I also added two debug functions which allow to show intermediate integer 
(idebug) or double (ddebug).
  \set i idebug(random(1, 10))

will print the random value and assign it to i.


I updated the defaut scripts, which seems to speed up meta command 
evaluations. The initial version does less than 2 million evals per 
second:
  sh> cat before.sql  \set naccounts 100000 * :scale  \setrandom aid 1 :naccounts
  sh> ./pgbench -T 3 -P 1 -f before.sql  [...]  tps = 1899004.793098 (excluding connections establishing)


The updated version does more than 3 million evals per second:
  sh> cat after.sql  \set aid random(1, 100000 * :scale)
  sh> ./pgbench -T 3 -P 1 -f after.sql  [...]  tps = 3143168.813690 (excluding connections establishing)


Suggestion:

A possibility would be to remove altogether the \setrandom stuff as the 
functionality is available through \set, maybe with an error message to 
advise about using \set with one of the random functions. That would 
remove about 200 ugly locs...

Another option would be to translate the setrandom stuff into a \set 
expression, that would maybe save 100 locs for the eval, but keep and 
expand a little the "manual" parser part.

-- 
Fabien.

pgsql-hackers by date:

Previous
From: Noah Misch
Date:
Subject: Re: Restore-reliability mode
Next
From: Andres Freund
Date:
Subject: Re: Improving replay of XLOG_BTREE_VACUUM records