On 12/21/2015 01:11 PM, Heikki Linnakangas wrote:
On 21/12/15 13:53, Tomas Vondra wrote:
On 12/21/2015 12:03 PM, Heikki Linnakangas wrote:
On 17/12/15 19:07, Robert Haas wrote:
If it works well empirically, does it really matter that it's
arbitrary? I mean, the entire planner is full of fairly arbitrary
assumptions about which things to consider in the cost model and
which to ignore. The proof that we have made good decisions there
is in the query plans it generates. (The proof that we have made
bad decisions in some cases in the query plans, too.)


What if it only seems to work well because it was tested on cases it was
designed for? What about the workloads that behave differently?

Whenever we do changes to costing and query planning, we carefully
consider counter-examples and cases where it might fail. I see nothing
like that in this thread - all I see is a bunch of pgbench tests, which
seems rather insufficient to me.

Agreed on that too.

I'm ready to spend some time on this, assuming we can agree on what
tests to run. Can we come up with realistic workloads where we expect
the patch might actually work poorly?

I think the worst case scenario would be the case where there is no
FPW-related WAL burst at all, and checkpoints are always triggered by
max_wal_size rather than checkpoint_timeout. In that scenario, the
compensation formula will cause the checkpoint to be too lazy in the
beginning, and it will have to catch up more aggressively towards the
end of the checkpoint cycle.

One such scenario might be to do only COPYs into a table with no
indexes. Or hack pgbench to do concentrate all the updates on only a few
very rows. There will be a FPW on those few pages initially, but the
spike will be much shorter. Or turn full_page_writes=off, and hack the
patch to do compensation even when fullpage_writes=off, and then just
run pgbench.

OK, the COPY scenario works interesting and also realistic because it probably applies to systems doing batch loads.

So that's one test to do, can we come up with some other?

We probably do want to do a bunch of pgbench tests, with various scales and also distributions - the gaussian/exponential distributions seem useful for simulating OLTP systems that usually have just s small active set (instead of touching all the data). This surely affects how much FPW we do and at what point - my expectetion is that the non-uniform distributions will have a long tail of FPW.

So I was thinking about these combinations:

* modes: uniform, gaussian, exponential
* scales: 1000 (15GB), 10000 (150GB)
* clients: 1, 2, 4, 8, 16 (to see impact on scalability, if any)

Each combination needs to run for at least an hour or two, possibly with multiple runs. I'll also try running this both on SSD-based sytem and a system with 10k drives, because those will probably behave differently.

Also, are we tracking the amount of FPW during the checkpoint, somewhere? That'd be useful, at least for this patch. Or do we need to just track the amount of WAL produced?


Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:

Reply via email to