> > Just so I understand correctly, do you suggest adding a metric for > delta(Time) / delta(Count) since the last reporting ? > <Collector>.TimePerGc or <Collector>.AverageTime would make sense. > AverageTime may be a bit nicer :) >
Yes, that's what I mean. My only concern is how useful this will be in reality. If there are only > (or several) long pauses then the msPerSec metrics will show it already, > and if there is a single long pause that may not be shown at all if there > are several shorter pauses as well with this metric. Let's say we measure this for every minute and see a 900 msPerSec (which means 54s within the minute are spent on GC). This may come from a single GC that lasts for 54s, or 2 GCs each lasting for ~27s, or more GCs with less time each. As the default heartbeat timeout is 50s, the former means there's likely a heartbeat timeout due to the GC pause, while the latter means otherwise. Mathematically, it is possible that there's 1 long pause together with several short pauses within the same measurement period, making the long pause not observable with AverageTime. However, from my experience, such a pattern is not normal in reality. My observation is that GCs happen at a similar time usually take a similar length of time. Admittedly, this is not a hard guarantee. Best, Xintong On Wed, Sep 6, 2023 at 3:59 PM Gyula Fóra <gyula.f...@gmail.com> wrote: > Matt Wang, > > I think the currently exposed info is all that is available through > GarbageCollectorMXBean. This FLIP does not aim to introduce a new more > granular way of reporting the per collector metrics, that would require a > new mechanism and may be a breaking change. > > We basically want to simply extend the current reporting here with the rate > metrics and the total metrics. > > Gyula > > On Wed, Sep 6, 2023 at 9:24 AM Matt Wang <wang...@163.com> wrote: > > > Hi Gyula, > > > > +1 for this proposal. > > > > Do we need to add a metric to record the count of different > > collectors? Now there is only a total count. For example, > > for G1, there is no way to distinguish whether it is the > > young generation or the old generation. > > > > > > > > -- > > > > Best, > > Matt Wang > > > > > > ---- Replied Message ---- > > | From | Gyula Fóra<gyula.f...@gmail.com> | > > | Date | 09/6/2023 15:03 | > > | To | <dev@flink.apache.org> | > > | Subject | Re: [DISCUSS] FLIP-361: Improve GC Metrics | > > Thanks Xintong! > > > > Just so I understand correctly, do you suggest adding a metric for > > delta(Time) / delta(Count) since the last reporting ? > > <Collector>.TimePerGc or <Collector>.AverageTime would make sense. > > AverageTime may be a bit nicer :) > > > > My only concern is how useful this will be in reality. If there are only > > (or several) long pauses then the msPerSec metrics will show it already, > > and if there is a single long pause that may not be shown at all if there > > are several shorter pauses as well with this metric. > > > > Gyula > > > > On Wed, Sep 6, 2023 at 8:46 AM Xintong Song <tonysong...@gmail.com> > wrote: > > > > Thanks for bringing this up, Gyula. > > > > The proposed changes make sense to me. +1 for them. > > > > In addition to the proposed changes, I wonder if we should also add > > something like timePerGc? This would help understand whether there are > long > > pauses, due to GC STW, that may lead to rpc unresponsiveness and > heartbeat > > timeouts. Ideally, we'd like to understand the max pause time per STW in > a > > recent time window. However, I don't see an easy way to separate the > pause > > time of each STW. Deriving the overall time per GC from the existing > > metrics (time-increment / count-increment) seems to be a good > alternative. > > WDYT? > > > > Best, > > > > Xintong > > > > > > > > On Wed, Sep 6, 2023 at 2:16 PM Rui Fan <1996fan...@gmail.com> wrote: > > > > Thanks for the clarification! > > > > By default the meterview measures for 1 minute sounds good to me! > > > > +1 for this proposal. > > > > Best, > > Rui > > > > On Wed, Sep 6, 2023 at 1:27 PM Gyula Fóra <gyula.f...@gmail.com> wrote: > > > > Thanks for the feedback Rui, > > > > The rates would be computed using the MeterView class (like for any > > other > > rate metric), just because we report the value per second it doesn't > > mean > > that we measure in a second granularity. > > By default the meterview measures for 1 minute and then we calculate > > the > > per second rates, but we can increase the timespan if necessary. > > > > So I don't think we run into this problem in practice and we can keep > > the > > metric aligned with other time rate metrics like busyTimeMsPerSec etc. > > > > Cheers, > > Gyula > > > > On Wed, Sep 6, 2023 at 4:55 AM Rui Fan <1996fan...@gmail.com> wrote: > > > > Hi Gyula, > > > > +1 for this proposal. The current GC metric is really unfriendly. > > > > I have a concern with your proposed rate metric: the rate is > > perSecond > > instead of per minute. I'm unsure whether it's suitable for GC > > metric. > > > > There are two reasons why I suspect perSecond may not be well > > compatible with GC metric: > > > > 1. GCs are usually infrequent and may only occur for a small number > > of time periods within a minute. > > > > Metrics are collected periodically, for example, reported every > > minute. > > If the result reported by the GC metric is 1s/perSecond, it does not > > mean that the GC of the TM is serious, because there may be no GC > > in the remaining 59s. > > > > On the contrary, the GC metric reports 0s/perSecond, which does not > > mean that the GC of the TM is not serious, and the GC may be very > > serious in the remaining 59s. > > > > 2. Stop-the-world may cause the metric to fail(delay) to report > > > > The TM will stop the world during GC, especially full GC. It means > > the metric cannot be collected or reported during full GC. > > > > So the collected GC metric may never be 1s/perSecond. This metric > > may always be good because the metric will only be reported when > > the GC is not severe. > > > > > > If these concerns make sense, how about updating the GC rate > > at minute level? > > > > We can define the type to Gauge for TimeMsPerMiunte, and updating > > this Gauge every second, it is: > > GC Total.Time of current time - GC total time of one miunte ago. > > > > Best, > > Rui > > > > On Tue, Sep 5, 2023 at 11:05 PM Maximilian Michels <m...@apache.org> > > wrote: > > > > Hi Gyula, > > > > +1 The proposed changes make sense and are in line with what is > > available for other metrics, e.g. number of records processed. > > > > -Max > > > > On Tue, Sep 5, 2023 at 2:43 PM Gyula Fóra <gyula.f...@gmail.com> > > wrote: > > > > Hi Devs, > > > > I would like to start a discussion on FLIP-361: Improve GC > > Metrics > > [1]. > > > > The current Flink GC metrics [2] are not very useful for > > monitoring > > purposes as they require post processing logic that is also > > dependent > > on > > the current runtime environment. > > > > Problems: > > - Total time is not very relevant for long running applications, > > only > > the > > rate of change (msPerSec) > > - In most cases it's best to simply aggregate the time/count > > across > > the > > different GabrageCollectors, however the specific collectors are > > dependent > > on the current Java runtime > > > > We propose to improve the current situation by: > > - Exposing rate metrics per GarbageCollector > > - Exposing aggregated Total time/count/rate metrics > > > > These new metrics are all derived from the existing ones with > > minimal > > overhead. > > > > Looking forward to your feedback. > > > > Cheers, > > Gyula > > > > [1] > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-361%3A+Improve+GC+Metrics > > [2] > > > > > > > > > > > > > > > https://nightlies.apache.org/flink/flink-docs-master/docs/ops/metrics/#garbagecollection > > > > > > > > > > > > >
