Hi Luke, Thanks for the KIP. I think the idea makes sense and would provide useful observability of log recovery. I have a few comments.
1. There's not a JIRA for this KIP (or the JIRA link needs updating). 2. Similarly the link to this discussion thread needs updating. 3. I wonder whether we need to keep these metrics (with value 0) once the broker enters the running state. Do you see it as valuable? A benefit of removing the metrics would be a reduction on storage required for metric stores which are recording these metrics. 4. I think the KIP's public interfaces section could be a bit clearer. Previous KIPs which added metrics usually used a table, with the MBean name, metric type and description. SeeKIP-551 for example (or KIP-748, KIP-608). Similarly you could use a table in the proposed changes section rather than describing the tree you'd see in an MBean console. Kind regards, Tom On Wed, 11 May 2022 at 09:08, Luke Chen <show...@gmail.com> wrote: > > And if people start using RemainingLogs and RemainingSegments and then > REALLY FEEL like they need RemainingBytes, then we can always add it in the > future. > > +1 > > Thanks James! > Luke > > On Wed, May 11, 2022 at 3:57 PM James Cheng <wushuja...@gmail.com> wrote: > > > Hi Luke, > > > > Thanks for the detailed explanation. I agree that the current proposal of > > RemainingLogs and RemainingSegments will greatly improve the situation, > and > > that we can go ahead with the KIP as is. > > > > If RemainingBytes were straight-forward to implement, then I’d like to > > have it. But we can live without it for now. And if people start using > > RemainingLogs and RemainingSegments and then REALLY FEEL like they need > > RemainingBytes, then we can always add it in the future. > > > > Thanks Luke, for the detailed explanation, and for responding to my > > feedback! > > > > -James > > > > Sent from my iPhone > > > > > On May 10, 2022, at 6:48 AM, Luke Chen <show...@gmail.com> wrote: > > > > > > Hi James and all, > > > > > > I checked again and I can see when creating UnifiedLog, we expected the > > > logs/indexes/snapshots are in good state. > > > So, I don't think we should break the current design to expose the > > > `RemainingBytesToRecovery` > > > metric. > > > > > > If there is no other comments, I'll start a vote within this week. > > > > > > Thank you. > > > Luke > > > > > >> On Fri, May 6, 2022 at 6:00 PM Luke Chen <show...@gmail.com> wrote: > > >> > > >> Hi James, > > >> > > >> Thanks for your input. > > >> > > >> For the `RemainingBytesToRecovery` metric proposal, I think there's > one > > >> thing I didn't make it clear. > > >> Currently, when log manager start up, we'll try to load all logs > > >> (segments), and during the log loading, we'll try to recover logs if > > >> necessary. > > >> And the logs loading is using "thread pool" as you thought. > > >> > > >> So, here's the problem: > > >> All segments in each log folder (partition) will be loaded in each log > > >> recovery thread, and until it's loaded, we can know how many segments > > (or > > >> how many Bytes) needed to recover. > > >> That means, if we have 10 partition logs in one broker, and we have 2 > > log > > >> recovery threads (num.recovery.threads.per.data.dir=2), before the > > >> threads load the segments in each log, we only know how many logs > > >> (partitions) we have in the broker (i.e. RemainingLogsToRecover > metric). > > >> We cannot know how many segments/Bytes needed to recover until each > > thread > > >> starts to load the segments under one log (partition). > > >> > > >> So, the example in the KIP, it shows: > > >> Currently, there are still 5 logs (partitions) needed to recover under > > >> /tmp/log1 dir. And there are 2 threads doing the jobs, where one > thread > > has > > >> 10000 segments needed to recover, and the other one has 3 segments > > needed > > >> to recover. > > >> > > >> - kafka.log > > >> - LogManager > > >> - RemainingLogsToRecover > > >> - /tmp/log1 => 5 ← there are 5 logs under > > >> /tmp/log1 needed to be recovered > > >> - /tmp/log2 => 0 > > >> - RemainingSegmentsToRecover > > >> - /tmp/log1 ← 2 threads are doing log > > >> recovery for /tmp/log1 > > >> - 0 => 10000 ← there are 10000 segments needed to > be > > >> recovered for thread 0 > > >> - 1 => 3 > > >> - /tmp/log2 > > >> - 0 => 0 > > >> - 1 => 0 > > >> > > >> > > >> So, after a while, the metrics might look like this: > > >> It said, now, there are only 4 logs needed to recover in /tmp/log1, > and > > >> the thread 0 has 9000 segments left, and thread 1 has 5 segments left > > >> (which should imply the thread already completed 2 logs recovery in > the > > >> period) > > >> > > >> - kafka.log > > >> - LogManager > > >> - RemainingLogsToRecover > > >> - /tmp/log1 => 3 ← there are 3 logs under > > >> /tmp/log1 needed to be recovered > > >> - /tmp/log2 => 0 > > >> - RemainingSegmentsToRecover > > >> - /tmp/log1 ← 2 threads are doing log > > >> recovery for /tmp/log1 > > >> - 0 => 9000 ← there are 9000 segments needed to be > > >> recovered for thread 0 > > >> - 1 => 5 > > >> - /tmp/log2 > > >> - 0 => 0 > > >> - 1 => 0 > > >> > > >> > > >> That said, the `RemainingBytesToRecovery` metric is difficult to > achieve > > >> as you expected. I think the current proposal with > > `RemainingLogsToRecover` > > >> and `RemainingSegmentsToRecover` should already provide enough info > for > > >> the log recovery progress. > > >> > > >> I've also updated the KIP example to make it clear. > > >> > > >> > > >> Thank you. > > >> Luke > > >> > > >> > > >>> On Thu, May 5, 2022 at 3:31 AM James Cheng <wushuja...@gmail.com> > > wrote: > > >>> > > >>> Hi Luke, > > >>> > > >>> Thanks for adding RemainingSegmentsToRecovery. > > >>> > > >>> Another thought: different topics can have different segment sizes. I > > >>> don't know how common it is, but it is possible. Some topics might > want > > >>> small segment sizes to more granular expiration of data. > > >>> > > >>> The downside of RemainingLogsToRecovery and > RemainingSegmentsToRecovery > > >>> is that the rate that they will decrement depends on the > configuration > > and > > >>> patterns of the topics and partitions and segment sizes. If someone > is > > >>> monitoring those metrics, they might see times where the metric > > decrements > > >>> slowly, followed by a burst where it decrements quickly. > > >>> > > >>> What about RemainingBytesToRecovery? This would not depend on the > > >>> configuration of the topic or of the data. It would actually be a > > pretty > > >>> good metric, because I think that this metric would change at a > > constant > > >>> rate (based on the disk I/O speed that the broker allocates to > > recovery). > > >>> Because it changes at a constant rate, you would be able to use the > > >>> rate-of-change to predict when it hits zero, which will let you know > > when > > >>> the broker is going to start up. Like, I would imagine if we graphed > > >>> RemainingBytesToRecovery that we'd see a fairly straight line that is > > >>> decrementing at a steady rate towards zero. > > >>> > > >>> What do you think about adding RemainingBytesToRecovery? > > >>> > > >>> Or, what would you think about making the primary metric be > > >>> RemainingBytesToRecovery, and getting rid of the others? > > >>> > > >>> I don't know if I personally would rather have all 3 metrics, or > would > > >>> just use RemainingBytesToRecovery. I'd too would like more community > > input > > >>> on which of those metrics would be useful to people. > > >>> > > >>> About the JMX metrics, you said that if > > >>> num.recovery.threads.per.data.dir=2, that there might be a separate > > >>> RemainingSegmentsToRecovery counter for each thread. Is that actually > > how > > >>> the data is structured within the Kafka recovery threads? Does each > > thread > > >>> get a fixed set of partitions, or is there just one big pool of > > partitions > > >>> that the threads all work on? > > >>> > > >>> As a more concrete example: > > >>> * If I have 9 small partitions and 1 big partition, and > > >>> num.recovery.threads.per.data.dir=2 > > >>> Does each thread get 5 partitions, which means one thread will finish > > >>> much sooner than the other? > > >>> OR > > >>> Do both threads just work on the set of 10 partitions, which means > > likely > > >>> 1 thread will be busy with the big partition, while the other one > ends > > up > > >>> plowing through the 9 small partitions? > > >>> > > >>> If each thread gets assigned 5 partitions, then it would make sense > > that > > >>> each thread has its own counter. > > >>> If the threads works on a single pool of 10 partitions, then it would > > >>> probably mean that the counter is on the pool of partitions itself, > > and not > > >>> on each thread. > > >>> > > >>> -James > > >>> > > >>>> On May 4, 2022, at 5:55 AM, Luke Chen <show...@gmail.com> wrote: > > >>>> > > >>>> Hi devs, > > >>>> > > >>>> If there are no other comments, I'll start a vote tomorrow. > > >>>> > > >>>> Thank you. > > >>>> Luke > > >>>> > > >>>> On Sun, May 1, 2022 at 5:08 PM Luke Chen <show...@gmail.com> wrote: > > >>>> > > >>>>> Hi James, > > >>>>> > > >>>>> Sorry for the late reply. > > >>>>> > > >>>>> Yes, this is a good point, to know how many segments to be > recovered > > if > > >>>>> there are some large partitions. > > >>>>> I've updated the KIP, to add a `*RemainingSegmentsToRecover*` > metric > > >>> for > > >>>>> each log recovery thread, to show the value. > > >>>>> The example in the Proposed section here > > >>>>> < > > >>> > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-831%3A+Add+metric+for+log+recovery+progress#KIP831:Addmetricforlogrecoveryprogress-ProposedChanges > > >>>> > > >>>>> shows what it will look like. > > >>>>> > > >>>>> Thanks for the suggestion. > > >>>>> > > >>>>> Thank you. > > >>>>> Luke > > >>>>> > > >>>>> > > >>>>> > > >>>>> On Sat, Apr 23, 2022 at 8:54 AM James Cheng <wushuja...@gmail.com> > > >>> wrote: > > >>>>> > > >>>>>> The KIP describes RemainingLogsToRecovery, which seems to be the > > >>> number > > >>>>>> of partitions in each log.dir. > > >>>>>> > > >>>>>> We have some partitions which are much much larger than others. > > Those > > >>>>>> large partitions have many many more segments than others. > > >>>>>> > > >>>>>> Is there a way the metric can reflect partition size? Could it be > > >>>>>> RemainingSegmentsToRecover? Or even RemainingBytesToRecover? > > >>>>>> > > >>>>>> -James > > >>>>>> > > >>>>>> Sent from my iPhone > > >>>>>> > > >>>>>>> On Apr 20, 2022, at 2:01 AM, Luke Chen <show...@gmail.com> > wrote: > > >>>>>>> > > >>>>>>> Hi all, > > >>>>>>> > > >>>>>>> I'd like to propose a KIP to expose a metric for log recovery > > >>> progress. > > >>>>>>> This metric would let the admins have a way to monitor the log > > >>> recovery > > >>>>>>> progress. > > >>>>>>> Details can be found here: > > >>>>>>> > > >>>>>> > > >>> > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-831%3A+Add+metric+for+log+recovery+progress > > >>>>>>> > > >>>>>>> Any feedback is appreciated. > > >>>>>>> > > >>>>>>> Thank you. > > >>>>>>> Luke > > >>>>>> > > >>>>> > > >>> > > >>> > > >
