On 30 Srpen 2012, 17:46, Robert Haas wrote:
> On Sun, Aug 26, 2012 at 1:04 PM, Tomas Vondra <tv@fuzzy.cz> wrote:
>> Attached is an improved patch, with a call to rand() replaced with
>> getrand().
>>
>> I was thinking about the counter but I'm not really sure how to handle
>> cases like "39%" - I'm not sure a plain (counter % 100 < 37) is not a
>> good sampling, because it always keeps continuous sequences of
>> transactions. Maybe there's a clever way to use a counter, but let's
>> stick to a getrand() unless we can prove is't causing issues. Especially
>> considering that a lot of data won't be be written at all with low
>> sampling rates.
>
> I like this patch, and I think sticking with a random number is a good
> idea. But I have two suggestions. Number one, I think the sampling
> rate should be stored as a float, not an integer, because I can easily
> imagine wanting a sampling rate that is not an integer percentage -
> especially, one that is less than one percent, like half a percent or
> a tenth of a percent. Also, I suggest that the command-line option
> should be a long option rather than a single character option. That
> will be more mnemonic and avoid using up too many single letter
> options, of which there is a limited supply. So to sample every
> hundredth result, you could do something like this:
>
> pgbench --latency-sample-rate 0.01
Right, I was thinking about that too. I'll do that in the next version of
the patch.
> Another option I personally think would be useful is an option to
> record only those latencies that are above some minimum bound, like
> this:
>
> pgbench --latency-only-if-more-than $MICROSECONDS
>
> The problem with recording all the latencies is that it tends to have
> a material impact on throughput. Your patch should address that for
> the case where you just want to characterize the latency, but it would
> also be nice to have a way of recording the outliers.
That sounds like a pretty trivial patch. I've been thinking about yet
another option - histograms (regular or with exponential bins).
What I'm not sure about is which of these options should be allowed at the
same time - to me, combinations like 'sampling + aggregation' don't make
much sense. Maybe except 'latency-only-if-more-than + aggregation'.
Tomas