> On May 13, 2015, 5:14 p.m., Jay Kreps wrote:
> > clients/src/main/java/org/apache/kafka/common/metrics/stats/Rate.java, line 
> > 62
> > <https://reviews.apache.org/r/34170/diff/1/?file=958215#file958215line62>
> >
> >     Is this actually right? I agree you'll get discontinuities as the 
> > measured time shrinks to zero but why is giving back 0 the right answer? In 
> > the case you guys were testing 0 was safe, but imagine a case where the 
> > monitoring was checking that the value didn't fall below some threshold.
> 
> Dong Lin wrote:
>     The question is, when Rate.measure() is called right after 
> Rate.record(n), what should be the return value? I think there are two 
> possibilities: 0 and n/config.timeWindowMs(). I didn't find any use case 
> where these two values make a difference. 
>     
>     Which value do you think is the best?
>     
>     Thank you.
> 
> Aditya Auradkar wrote:
>     I think returning 0 is reasonable if no time has elapsed (technically).
>     As an alternate solution, what if we assumed that "elapsed" is always 1 
> (at least). For example:
>     
>     double elapsed = convert(now - stat.oldest(now).lastWindowMs) + 1
>     
>     In case of seconds, this basically means that you assume the current 
> second is always complete. This is only a problem (for a couple of seconds) 
> when all previous samples have zero activity or when the server is just 
> starting up.
> 
> Jay Kreps wrote:
>     Actually there is a fundamental issue with the computation that patching 
> around the 0 case doesn't fix. That is the instability of the estimate. This 
> is the "taking a poll with sample size one" problem.
>     
>     Even if you patch the 0 case you still get a bad answer 1 ms later. That 
> is, let's say you get a single 50k request and your quota is 1MB/sec. 
> Currently at 0ms we estimate infinity which is in fact the measured rate but 
> obviously not a good estimate. But even 1 ms later the estimate is bad. 
> 50k*1000ms = ~50MB/sec.
>     
>     This is somewhat rare because it only happens when there is just one 
> sample.
>     
>     They key observation is that if a sample is missing, nothing happened in 
> that time period. But the calculation should still use that time period.
>     
>     So the right way to compute it, I think, is actually
>       ellapsed = (num_samples-1)*window_size + (now - current_sample.begin)
>       
>     For safety I think we should also require the number of samples to be >= 
> 2 and default it to 3.
> 
> Dong Lin wrote:
>     Hi Jay: thanks for comments!
>     
>     I think the problem here is that, when event number is very small (e.g. 
> 0, 1), what value should rate.measure() return, right? If I understand your 
> solution formula right, when there is only one sample, your measured rate is 
> n/ellapsed = n/(now - current_sample.begin), which is exactly same as the 
> current code. I agree that requiring number of samples to be >= 2 solve the 
> problem. But what happens when user call rate.measure() when there is only 1 
> sample?
>     
>     I agree with you that we still get a bad answer if we patch the 0 case. 
> How about we patch the 0 to timeWindowMs case: if (ellapsed < timeWindowMs) 
> then ellapsed = timeWindowMs. Does this solve the problem here?
> 
> Dong Lin wrote:
>     I think we all agree that there is problem with rate measure when 
> ellapsed time is too small. What we need to properly define the rate measure 
> to handle such case, which is currently missing.
>     
>     If we want to require number of samples to be >= 2 as Jay suggested, we 
> can also let ellapsed = max(ellapsed, 2*timeWindowMs) and keep the rest of 
> the code, right?

If we require num samples greater than 2, we shoud simply change this condition 
in MetricConfig:
        if (samples < 1)
            throw new IllegalArgumentException("The number of samples must be 
at least 1.");

Jay's equation is now similar to Dongs.

If num samples = 3,
ellapsed = (num_samples-1)*window_size + (now - current_sample.begin)
ellapsed = 2*window_size + (now - current_sample.begin)


- Aditya


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/34170/#review83629
-----------------------------------------------------------


On May 14, 2015, 7:34 a.m., Dong Lin wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/34170/
> -----------------------------------------------------------
> 
> (Updated May 14, 2015, 7:34 a.m.)
> 
> 
> Review request for kafka.
> 
> 
> Bugs: KAFKA-2191
>     https://issues.apache.org/jira/browse/KAFKA-2191
> 
> 
> Repository: kafka
> 
> 
> Description
> -------
> 
> KAFKA-2191; Measured rate should not be infinite
> 
> 
> Diffs
> -----
> 
>   clients/src/main/java/org/apache/kafka/common/metrics/stats/Rate.java 
> 98429da34418f7f1deba1b5e44e2e6025212edb3 
>   
> clients/src/main/java/org/apache/kafka/common/metrics/stats/SampledStat.java 
> b341b7daaa10204906d78b812fb05fd27bc69373 
> 
> Diff: https://reviews.apache.org/r/34170/diff/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Dong Lin
> 
>

Reply via email to