Re: pgbench--new transaction type

On 06/30/2011 12:13 AM, Jeff Janes wrote:
> One more thought I had, would it make sense to change this from the
> creation of a PL/pgSQL permanent function to instead use the recently
> added DO anonymous block syntax? I think that would be somewhat
> cleaner about leaving cruft behind in the database. But it would
> increase the overhead of each outer execution, and would also mean
> that it would not be backwards compatible to run against servers
> before 9.0
>

I think some measurement of the overhead difference would be needed to
know for sure about the first part. I suspect that given the block size
of 512 now being targeted, that would end up not mattering very much.

pgbench's job is to generate a whole database full of cruft, so I can't
say I'd find an argument from either side of that to be very
compelling. I'm not real busy anymore testing performance of PostgreSQL
instances from before 9.0 anymore either, so whether this mode was
compatible with them or not isn't very compelling either. Just a mixed
bag all around in those areas.

I would say take a look at what the performance change looks like, and
see if it turns out to make the patch to pgbench less obtrusive. The
main objection against committing this code I can see is that it will
further complicate pgbench for a purpose not many people care about. So
if the DO version ends up with a smaller diff and less impact on the
codebase, that would likely be a more substantial tie-breaker in its
favor than any of these other arguments.

--
Greg Smith 2ndQuadrant US greg(at)2ndQuadrant(dot)com Baltimore, MD