> If you assume the 1st checkpoint needs to be "full" you know you are not > allowed to use any shared files. > It's true you should know about the shared files of the previous checkpoint, > but e.g. RocksDB already tracks that.
I mean that the design described by FLIP implies the following (PCIIW): 1. treat SST files from the initial checkpoint specially: re-upload or send placeholder - depending on those attributes in state handle 2. (SST files from newer checkpoints are re-uploaded depending on confirmation currently; so yes there is tracking, but it's different) 3. SharedStateRegistry must allow replacing state under the existing key; otherwise, if a new key is used then other parallel subtasks should learn somehow this key and use it; However, allowing replacement must be limited to this scenario, otherwise it can lead to previous checkpoint corruption in normal cases Forcing a full checkpoint after completing N checkpoints instead of immediately would only require enabling (1) after N checkpoints. And with the "poll API until checkpoint released" approach, those changes aren't necessary. > There is one more fundamental issue with either of your two proposals that've > just came to my mind. > What happens if you have externalized checkpoints and the job fails before > the initial checkpoint can be safely removed? You start the job from the latest created checkpoint and wait for it to be allowed for deletion. Then you can delete it, and all previous checkpoints (or am I missing something?) > With tracking the shared files on JM you can not say if you can clear the > files after couple of checkpoints or 10s, 100s or 1000s, > which translates into minutes/hours/days/weeks of processing. This doesn't necessarily translate into higher cost (because of saved RPC etc., as I mentioned above). However, I do agree that an infinite or arbitrary high delay is unacceptable. The added complexity above doesn't seem negligible to me (especially in SharedStateHandle); and should therefore be weighted against those operational disadvantages (given that the number of checkpoints to wait is bounded in practice). Regards, Roman On Mon, Nov 22, 2021 at 5:05 PM Dawid Wysakowicz <dwysakow...@apache.org> wrote: > > There is one more fundamental issue with either of your two proposals > that've just came to my mind. What happens if you have externalized > checkpoints and the job fails before the initial checkpoint can be > safely removed? You have a situation where you have a retained > checkpoint that was built on top of the original one. Basically ending > in a situation we have right now that you never know when it is safe to > delete a retained checkpoint. > > BTW, the intention for the "claim" mode was to support cases when users > are concerned with the performance of the first checkpoint. In those > cases they can claim the checkpoint on don't pay the additional cost of > the first checkpoint. > > Best, > > Dawid > > On 22/11/2021 14:09, Roman Khachatryan wrote: > > Thanks Dawid, > > > > Regarding clarity, > > I think that all proposals require waiting for some event: re-upload / > > checkpoint completion / api response. > > But with the current one, there is an assumption: "initial checkpoint > > can be deleted once a new one completes" (instead of just "initial > > checkpoint can be deleted once the API says it can be deleted"). > > So I think it's actually more clear to offer this explicit API and rely on > > it. > > > > Regarding delaying the deletion, > > I agree that it can delay deletion, but how important is it? > > Checkpoints are usually stored on relatively cheap storage like S3, so > > some delay shouldn't be an issue (especially taking rounding into > > account); it can even be cheaper or comparable to paying for > > re-upload/duplicate calls. > > > > Infinite delay can be an issue though, I agree. > > Maybe @Yun can clarify the likelihood of never deleting some SST files > > by RocksDB? > > For the changelog backend, old files won't be used once > > materialization succeeds. > > > > Yes, my concern is checkpointing time, but also added complexity: > >> It would be a bit invasive though, as we would have to somehow keep track > >> which files should not be reused on TMs. > > I think we need this anyway if we choose to re-upload files once the > > job is running. > > The new checkpoint must be formed by re-uploaded old artifacts AND > > uploaded new artifacts. > > > > > > Regards, > > Roman > > > > > > On Mon, Nov 22, 2021 at 12:42 PM Dawid Wysakowicz > > <dwysakow...@apache.org> wrote: > >> @Yun > >> > >> I think it is a good comment with I agree in principal. However, we use > >> --fromSavepoint (cli), savepointPath (REST API), SavepointRestoreSettings > >> for both restoring from a savepoint and an externalized checkpoint > >> already. I wanted to voice that concern. Nevertheless I am fine with > >> changing it to execution.restore-mode, if there are no other comments on > >> that matter, I will change it. > >> > >> @Roman: > >> > >> Re 1. Correct, stop-with-savepoint should commit side-effects. Will add > >> that to the doc. > >> > >> Re.2 What I don't like about this counter proposal is that it still has no > >> clearly defined point in time when it is safe to delete the original > >> checkpoint. Users would have a hard time reasoning about it and debugging. > >> Even worse, I think worst case it might never happen that all the original > >> files are no longer in use (I am not too familiar with RocksDB compaction, > >> but what happens if there are key ranges that are never accessed again?) I > >> agree it is unlikely, but possible, isn't it? Definitely it can take a > >> significant time and many checkpoints to do so. > >> > >> Re. 3 I believe where you are coming from is that you'd like to keep the > >> checkpointing time minimal and reuploading files may increase it. The > >> proposal so far builds on the assumption we could in most cases use a > >> cheap duplicate API instead of re-upload. I could see this as a follow-up > >> if it becomes a bottleneck. It would be a bit invasive though, as we would > >> have to somehow keep track which files should not be reused on TMs. > >> > >> Re. 2 & 3 Neither of the counter proposals work well for taking > >> incremental savepoints. We were thinking of building incremental > >> savepoints on the same concept. I think delaying the completion of an > >> independent savepoint to a closer undefined future is not a nice property > >> of savepoints. > >> > >> Re 4. Good point. We should make sure the first completed checkpoint has > >> the independent/full checkpoint property rather than just the first > >> triggered. > >> > >> Re. 5 & 6 I need a bit more time to look into it. > >> > >> Best, > >> > >> Dawid > >> > >> On 22/11/2021 11:40, Roman Khachatryan wrote: > >> > >> Hi, > >> > >> Thanks for the proposal Dawid, I have some questions and remarks: > >> > >> 1. How will stop-with-savepoint be handled? > >> Shouldn't side effects be enforced in this case? (i.e. send > >> notifyCheckpointComplete) > >> > >> 2. Instead of forcing re-upload, can we "inverse control" in no-claim mode? > >> Anyways, any external tool will have to poll Flink API waiting for the > >> next (full) checkpoint, before deleting the retained checkpoint, > >> right? > >> Instead, we can provide an API which tells whether the 1st checkpoint > >> is still in use (and not force re-upload it). > >> > >> Under the hood, it can work like this: > >> - for the checkpoint Flink recovers from, remember all shared state > >> handles it is adding > >> - when unregistering shared state handles, remove them from the set above > >> - when the set becomes empty the 1st checkpoint can be deleted externally > >> > >> Besides not requiring re-upload, it seems much simpler and less invasive. > >> On the downside, state deletion can be delayed; but I think this is a > >> reasonable trade-off. > >> > >> 3. Alternatively, re-upload not necessarily on 1st checkpoint, but > >> after a configured number of checkpoints? > >> There is a high chance that after some more checkpoints, initial state > >> will not be used (because of compaction), > >> so backends won't have to re-upload anything (or small part). > >> > >> 4. Re-uploaded artifacts must not be deleted on checkpoin abortion > >> This should be addressed in > >> https://issues.apache.org/jira/browse/FLINK-24611. > >> If not, I think the FLIP should consider this case. > >> > >> 5. Enforcing re-upload by a single task and Changelog state backend > >> With Changelog state backend, a file can be shared by multiple operators. > >> Therefore, getIntersection() is irrelevant here, because operators > >> might not be sharing any key groups. > >> (so we'll have to analyze "raw" file usage I think). > >> > >> 6. Enforcing re-upload by a single task and skew > >> If we use some greedy logic like subtask 0 always re-uploads then it > >> might be overloaded. > >> So we'll have to obtain a full list of subtasks first (then probably > >> choose randomly or round-robin). > >> However, that requires rebuilding Task snapshot, which is doable but > >> not trivial (which I think supports "reverse API option"). > >> > >> 7. I think it would be helpful to list file systems / object stores > >> that support "fast" copy (ideally with latency numbers). > >> > >> Regards, > >> Roman > >> > >> On Mon, Nov 22, 2021 at 9:24 AM Yun Gao <yungao...@aliyun.com.invalid> > >> wrote: > >> > >> Hi, > >> > >> Very thanks Dawid for proposing the FLIP to clarify the ownership for the > >> states. +1 for the overall changes since it makes the behavior clear and > >> provide users a determined method to finally cleanup savepoints / retained > >> checkpoints. > >> > >> Regarding the changes to the public interface, it seems currently the > >> changes are all bound > >> to the savepoint, but from the FLIP it seems perhaps we might also need to > >> support the claim declaration > >> for retained checkpoints like in the cli side[1] ? If so, then might it be > >> better to change the option name > >> from `execution.savepoint.restore-mode` to something like > >> `execution.restore-mode`? > >> > >> Best, > >> Yun > >> > >> > >> [1] > >> https://nightlies.apache.org/flink/flink-docs-master/docs/ops/state/checkpoints/#resuming-from-a-retained-checkpoint > >> > >> > >> ------------------------------------------------------------------ > >> From:Konstantin Knauf <kna...@apache.org> > >> Send Time:2021 Nov. 19 (Fri.) 16:00 > >> To:dev <dev@flink.apache.org> > >> Subject:Re: [DISCUSS] FLIP-193: Snapshots ownership > >> > >> Hi Dawid, > >> > >> Thanks for working on this FLIP. Clarifying the differences and > >> guarantees around savepoints and checkpoints will make it easier and safer > >> for users and downstream projects and platforms to work with them. > >> > >> +1 to the changing the current (undefined) behavior when recovering from > >> retained checkpoints. Users can now choose between claiming and not > >> claiming, which I think will make the current mixed behavior obsolete. > >> > >> Cheers, > >> > >> Konstantin > >> > >> On Fri, Nov 19, 2021 at 8:19 AM Dawid Wysakowicz <dwysakow...@apache.org> > >> wrote: > >> > >> Hi devs, > >> > >> I'd like to bring up for a discussion a proposal to clean up ownership > >> of snapshots, both checkpoints and savepoints. > >> > >> The goal here is to make it clear who is responsible for deleting > >> checkpoints/savepoints files and when can that be done in a safe manner. > >> > >> Looking forward for your feedback! > >> > >> Best, > >> > >> Dawid > >> > >> [1] https://cwiki.apache.org/confluence/x/bIyqCw > >> > >> > >> > >> -- > >> > >> Konstantin Knauf > >> > >> https://twitter.com/snntrable > >> > >> https://github.com/knaufk > >> >
