Re: pgbench rate limiting changes transaction latency computation

On 2019-Jun-12, Heikki Linnakangas wrote:

> That was my thought too, when looking at this example. When there is already
> a long queue of transactions, the txtime of individual transactions doesn't
> matter much. The most important thing under that condition is how fast the
> system can dissolve the queue (or how fast it builds up even more). So the
> derivative of the lag or lat seems like the most important figure. We don't
> print exactly that, but it's roughly the same as the TPS. Jitter experienced
> by the user matters too, i.e. stddev of 'lat'.

It's funny that you mention taking the derivative of lat or lag, because
that suggests that these numbers should not be merely printed on the
screen but rather produced in a way that's easy for a database to
consume. Then you can just write the raw numbers and provide a set of
pre-written queries that generate whatever numbers the user desires.
We already have that ... but we don't provide any help on actually using
those log files -- there aren't instructions on how to import that into
a table, or what queries could be useful.

Maybe that's a useful direction to move towards? I think the console
output is good for getting a gut-feeling of the test, but not for actual
data analysis.

--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services