[ https://issues.apache.org/jira/browse/CASSANDRA6486?page=com.atlassian.jira.plugin.system.issuetabpanels:commenttabpanel&focusedCommentId=13848498#comment13848498
]
Benedict commented on CASSANDRA6486:

So, after thinking about this a little more, I may be leaning towards a slightly modified
approach, that avoids the perthread allocation and dynamic resizing of ranges in favour of
a single global reservoir that is updated directly by each thread. This has the disadvantage
that the intervals you're timing are more difficult to define, but we really don't need that
kind of paranoia with accuracy for measuring manymicrosecond and above events.
I'm currently thinking of using a rolling collection of samplehistograms (say 10 per timer)
to provide a rolling window on the desired measurement interval, and on retiring the oldest
samplehistogram the result can be merged into the next tier of interval we're measuring.
Alternatively we could take the same approach but with just a regular histogram, but I currently
prefer the sampled approach as, even with larger windows than a histogram, the distribution
for any window is more likely to closely approximate a normal distribution and so should give
a more accurate picture of latencies for the interval even with a very small sample size.
> Latency Measurement
> 
>
> Key: CASSANDRA6486
> URL: https://issues.apache.org/jira/browse/CASSANDRA6486
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Benedict
> Assignee: Benedict
>
> Latency measurement in Cassandra is currently suboptimal. Exactly what the latency measurements
tell you isn't intuitively clear due to their exponentially decaying, but amount to some view
of the latency per (unweighted) operation over the past, approximately, 10 minute period,
with greater weight given to more recent operations. This has some obvious flaws, the most
notable being that due to probabilistic sampling, large outlier events (e.g. GC) can easily
be lost over a multiminute time horizon, and even when caught are unlikely to appear even
in the 99.9th percentile due to accounting for a tiny fraction of events numerically.
> I'm generally thinking about how we might improve on this, and want to dump my ideas
here for discussion. I think the following things should be targeted:
> 1) Ability to see uniform latency measurements for different time horizons stretching
back from the present, e.g. last 1s, 1m, 1hr and 1day
> 2) Ability to bound the error margin of statistics for all of these intervals
> 3) Protect against losing outlier measurements
> 4) Possibly offer the ability to weight statistics, so that longer latencies are not
underplayed even if they are counted
> 5) Preferably nonblocking, memory efficient, and relatively garbagefree
> (3) and (4) are the trickiest, as a theoretically sound and general approach isn't immediately
obvious. There are a number of possibilities that spring to mind:
> 1) ensure that we have enough sample points that we are probabilistically guaranteed
to not lose them, but over large time horizons this is problematic due to memory constraints,
and it doesn't address (4);
> 2) count large events multiple times (or subslices of the events), based on e.g. average
oprate. I am not a fan of this idea because it makes possibly bad assumptions about behaviour
and doesn't seem very theoretically sound;
> 3) weight the probability of retaining an event by its length. the problem with this
approach is that it ties you into (4) without offering the current view of statistics (i.e.
unweighted operations), and it also doesn't lend itself to efficient implementation
> I'm currently leaning towards a fourth approach, which attempts to hybridise uniform
sampling and histogram behaviour, by separating the sample space into ranges, each some multiple
of the last (say 2x the size). Each range has a uniform sample of events that occured in that
range, plus a count of total events. Ideally the size of the sample will be variable based
on the number of events occurring in any range, but that there will be a lowerbound calculated
to ensure we do not lose events.
> This approach lends itself to all 5 goals above:
> 1) by maintaining the same structure for each time horizon, and uniformly sampling from
all of the directly lower order time horizons to maintain it;
> 2) by imposing minimum sample sizes for each range;
> 3) ditto (2);
> 4) by producing time/frequencyweighted statistics using the samples and counts from
each range;
> 5) with threadlocal arraybased timers that are synchronised with the global timer once
every minimum period, by the owning thread
> This also extends reasonably nicely the timers I have already written for CASSANDRA6199,
so some of the work is already done.
> Thoughts / discussion would be welcome, especially if you think I've missed another obvious
approach.

This message was sent by Atlassian JIRA
(v6.1.4#6159)
