Hi Piotr, Thank you for the very detailed comments! Please see my reply inline.
On Thu, Jul 20, 2023 at 12:24 AM Piotr Nowojski <piotr.nowoj...@gmail.com> wrote: > Hi Dong, > > I have a couple of follow up questions about switching back and forth > between streaming and batching mode. > Especially around shuffle/watermark strategy, and keyed state backend. > > First of all, it might not always be beneficial to switch into the batch > modes: > - Shuffle strategy > - Is sorting going to be purely in-memory? If not, obviously spilling > to disks might cause larger overheads > compared to not sorting the records. > Sorting might require spilling data to disk depending on the input size. The behavior of sorting w.r.t. memory/disk is expected to be exactly the same as the behavior of input sorting automatically performed by Flink runtime in batch mode for keyed inputs. More specifically, ExternalSorter <https://github.com/apache/flink/blob/master/flink-runtime/src/main/java/org/apache/flink/runtime/operators/sort/ExternalSorter.java> is currently used to sort keyed inputs in batch mode. It is automatically used by Flink runtime in OneInputStreamTask (here <https://github.com/apache/flink/blob/master/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/tasks/OneInputStreamTask.java#L114>) and in MultiInputSortingDataInput (here <https://github.com/apache/flink/blob/master/flink-streaming-java/src/main/java/org/apache/flink/streaming/api/operators/sort/MultiInputSortingDataInput.java#L188>). We plan to re-use the same code/mechanism to do sorting. It is possible that spilling to disks might cause larger overhead. IMO it is an orthogonal issue already existing in Flink. This is because a Flink job running batch mode might also be slower than its throughput in stream mode due to the same reason. However, even though it is possible in theory, I expect that in practice the throughput of using sorting + BatchExecutionKeyedStateBackend should be much higher than using other keyed statebackends when the amount of data is large. As a matter of fact, we have not heard of complaints of such performance regression issues in batch mode. The primary goal of this FLIP is to allow the operator to run at the same throughput (in stream mode when there is backlog) as it can currently do in batch mode. And this goal is not affected by the disk overhead issue mentioned above. I am thinking maybe we can treat it as an orthogonal performance optimization problem instead of solving this problem in this FLIP? - If it will be at least partially in-memory, does Flink have some > mechanism to reserve optional memory that > can be revoked if a new operator starts up? Can this memory be > redistributed? Ideally we should use as > much as possible of the available memory to avoid spilling costs, but > also being able to revoke that memory > This FLIP does not support dynamically revoking/redistribuitng managed memory used by the ExternalSorter. For operators with isInternalSorterSupported = true, we will allocate to this operator execution.sorted-inputs.memory <https://github.com/apache/flink/blob/master/flink-core/src/main/java/org/apache/flink/configuration/ExecutionOptions.java#L144> amount of managed memory. This is the same as how Flink allocates managed memory to an operator when this operator has keyed inputs in batch mode. Note that this FLIP intends to support operators to sort inputs whenever there is backlog. And there is currently no way for an operator to know in advance whether there will be no backlog after a given time. So it seems simpler to just keep managed memory for such an operator throughout the lifecycle of this operator, for now. Besides, it seems that the lack of ability to dynamically revoke/redistribute un-used managed memory is an existing issue in Flink. For example, we might have two operators sharing the same slot and these two operators both use managed memory (e.g. to sort inputs). There is currently no way for one operator to re-use the memory not used by the other operator. Therefore, I think we can treat this as an orthogonal performance optimization problem which can be addressed separately. What do you think? > - Sometimes sorting, even if we have memory to do that, might be an > unnecessary overhead. > - Watermarks > - Is holding back watermarks always good? If we have tons of data > buffered/sorted and waiting to be processed > with multiple windows per key and many different keys. When we > switch back to `isBacklog=false` we > first process all of that data before processing watermarks, for > operators that are not using sorted input the > state size can explode significantly causing lots of problems. Even > for those that can use sorting, switching to > sorting or BatchExecutionKeyedStateBackend is not always a good > idea, but keeping RocksDB also can be > risky. > With the current FLIP, the proposal is to use a sorter only when the inputs have keys. According to this practice, operators which are not using sorting should have un-keyed inputs. I believe such an operator will not even use a keyed state backend. Maybe I missed some use-case. Can you provide a use-case where we will have an operator with un-keyed inputs whose state size can explode due to we holding back watermarks? For operators with keyed inputs that use sorting, I suppose it is possible that sorting + BatchExecutionKeyedStateBackend can be worse than using RocksDB. But I believe this is very very rare (if possible) in almost practical usage of Flink. Take one step back, if this indeed cause regression for a real use-case, user can set execution.checkpointing.interval-during-backlog to anything other than 0 so that this FLIP will not use sorter + BatchExecutionKeyedStateBackend even even when there is backlog. I would hope we can find a way to automatically determine whether using sorting + BatchExecutionKeyedStateBackend can be better or worse than using RocksDB alone. But I could not find a good and reliable way to do this. Maybe we can update Flink to do this when we find a good way to do this in the future? > - Keyed state backend > - I think you haven't described what happens during switching from > streaming to backlog processing. > Good point. This indeed needs to be described. I added a TODO in the "Behavior changes ..." section to describe what happens when isBacklog switches from false to true, for all watermark/checkpoint/statebackend etc. Let me explain this for the state backend here for now. I will update FLIP later. When isBacklog switches from false to true, operator with keyed inputs can optionally (as determined by its implementation) starts to use internal sorter to sort inputs by key, without processing inputs or updating statebackend, until it receives end-of-inputs or isBacklog is switched to false again. > - Switch can be an unnecessary overhead. I agree it can cause unnecessary overhead, particularly when isBacklog switches back and forth frequently. Whether or not this is unnecessary likely depends on the duration/throughput of the backlog phase as well as the specific computation logic of the operator. I am not sure there is a good way for Flink to determine in advance whether switching is unnecessary. Note that for the existing use-case where we expect to change isBacklog to true (e.g. MySQL CDC snapshot phase, Kafka source watermark lag being too high), we don't expect the watermark to switch back and force frequently. And user can disable this switch by setting execution.checkpointing.interval-during-backlog to anything other than 0. Therefore, I am wondering if we can also view this as a performance optimization opportunity for extra use-cases in the future, rather than a blocking issue of this FLIP for the MVP use-case (e.g. snapshot phase for any CDC source, Kafka watermark lag). > At the same time, in your current proposal, for > `execution.checkpointing.interval-during-backlog > 0` we won't > switch to "batch" mode at all. That's a bit of shame, I don't understand > why those two things should be coupled > together? > We can in general classify optimizations as those that are compatible with checkpointing, and those that are not compatible with checkpointing. For example, input sorting is currently not compatible with checkpointing. And buffering input records to reduce state backend overhead (and probably columnar processing for mini-batch in the future) is compatible with checkpointing. The primary of FLIP-327 is to support optimizations not compatible with checkpointing. If execution.checkpointing.interval-during-backlog > 0, which means that user intends to still do checkpointing even when there is backog, then we will not be able to support such optimizations. For optimizations that are compatible with checkpointing, we can do this even when the operator does not run in "batch mode". There are extra problems to solve in order to achieve this optimization, such as supporting unaligned checkpointing without prolonging its sync phase. I plan to explain how this can be done in FLIP-325. > All in all, shouldn't we aim for some more clever process of switching back > and forth between streaming/batch modes > for watermark strategy/state backend/sorting based on some metrics? Trying > to either predict if switching might help, > or trying to estimate if the last switch was beneficial? Maybe something > along the lines: > - sort only in memory and during sorting count the number of distinct keys > (NDK) > - maybe allow for spilling if so far in memory we have NDK * 5 >= > #records > - do not allow to buffer records above a certain threshold, as otherwise > checkpointing can explode > - switch to `BatchExecutionKeyedStateBackend` only if NDK * 2 >= #records > - do not sort if last NDKs (or EMA of NDK?) 1.5 <= #records > > Or even maybe for starters something even simpler and then test out > something more fancy as a follow up? > I agree it is worth investigating these ideas to further optimize the performance during backlog. I just think these can be done independently after this FLIP. The focus of this FLIP is to re-use in stream mode the same optimization which we already use in batch mode, rather than inventing or improving the performance of these existing optimizations. Given that there are already a lot of new mechanism/features to discuss and address in this FLIP, I am hoping we can limit the scope of this FLIP to re-use the existing optimization, and do these extra optimization opportunities as future work. What do you think? > > At the same time, `execution.checkpointing.interval-during-backlog=0` seems > a weird setting to me, that I would > not feel safe recommending to anyone. If processing of a backlog takes a > long time, a job might stop making > any progress due to some random failures. Especially dangerous if a job switches from streaming mode back to > backlog processing due to some reasons, as that could happen months after > someone started a job with this > strange setting. So should we even have it? I would simply disallow it. I > Good point. I do agree we need to further work to improve the failover performance in case any task fails. As of the current FLIP, if any task fails during backlog and execution.checkpointing.interval-during-backlog = 0, we will need to restart all operators to the last checkpointed state and continue processing backlog. And this can be a lot of rollback since there is no checkpoint during backlog. And this can also be worse than batch since this FLIP currently does not support exporting/saving records to local disk (or shuffle service) so that a failed task can re-consume the records from the upstream task (or shuffle service) in the same way as how Flink failover a task in batch mode. I think we can extend this FLIP to solve this problem so that it can have at least the same behavior/performance as batch-mode job. The idea is to also follow what batch mode does. For example, we can trigger a checkpoint when isBacklog switches to true, and every operator should buffer its output in the TM local disk (or remote shuffle service). Therefore, after a task fails, it can restart from the last checkpoint and re-consume data buffered in the upstream task. I will update FLIP as described above. Would this address your concern? > could see a power setting like: > `execution.backlog.use-full-batch-mode-on-start (default false)` > I am not sure I fully understand this config or its motivation. Can you help explain the exact semantics of this config? > that would override any heuristic of switching to backlog if someone is > submitting a new job that starts with > `isBacklog=true`. > > Or we could limit the scope of this FLIP to only support starting with > batch mode and switching only once to > streaming, and design a follow up with switching back and forth? > Sure, that sounds good to me. I am happy to split this FLIP into two FLIPs so that we can make incremental progress. Best, Dong > I'm looking forwards to hearing/reading out your thoughts. > > Best, > Piotrek > > > śr., 12 lip 2023 o 12:38 Jing Ge <j...@ververica.com.invalid> napisał(a): > > > Hi Dong, > > > > Thanks for your reply! > > > > Best regards, > > Jing > > > > On Wed, Jul 12, 2023 at 3:25 AM Dong Lin <lindon...@gmail.com> wrote: > > > > > Hi Jing, > > > > > > Thanks for the comments. Please see my reply inline. > > > > > > On Wed, Jul 12, 2023 at 5:04 AM Jing Ge <j...@ververica.com.invalid> > > > wrote: > > > > > > > Hi Dong, > > > > > > > > Thanks for the clarification. Now it is clear for me. I got > additional > > > noob > > > > questions wrt the internal sorter. > > > > > > > > 1. when to call setter to set the internalSorterSupported to be true? > > > > > > > > > > Developer of the operator class (i.e. those classes which implements > > > `StreamOperator`) should override the `#getOperatorAttributes()` API to > > set > > > internalSorterSupported to true, if he/she decides to sort records > > > internally in the operator. > > > > > > > > > > 2 > > > > *"For those operators whose throughput can be considerably improved > > with > > > an > > > > internal sorter, update it to take advantage of the internal sorter > > when > > > > its input has isBacklog=true.* > > > > *Typically, operators that involve aggregation operation (e.g. join, > > > > cogroup, aggregate) on keyed inputs can benefit from using an > internal > > > > sorter."* > > > > > > > > *"The operator that performs CoGroup operation will instantiate two > > > > internal sorter to sorts records from its two inputs separately. Then > > it > > > > can pull the sorted records from these two sorters. This can be done > > > > without wrapping input records with TaggedUnion<...>. In comparison, > > the > > > > existing DataStream#coGroup needs to wrap input records with > > > > TaggedUnion<...> before sorting them using one external sorter, which > > > > introduces higher overhead."* > > > > > > > > According to the performance test, it seems that internal sorter has > > > better > > > > performance than external sorter. Is it possible to make those > > operators > > > > that can benefit from it use internal sorter by default? > > > > > > > > > > Yes, it is possible. After this FLIP is done, users can use > > > DataStream#coGroup with EndOfStreamWindows as the window assigner to > > > co-group two streams in effectively the batch manner. An operator that > > uses > > > an internal sorter will be used to perform the co-group operation. > There > > is > > > no need for users of the DataStream API to explicitly know or set the > > > internal sorter in anyway. > > > > > > In the future, we plan to incrementally optimize other aggregation > > > operation (e.g. aggregate) on the DataStream API when > EndOfStreamWindows > > is > > > used as the window assigner. > > > > > > Best, > > > Dong > > > > > > > > > > > > > > Best regards, > > > > Jing > > > > > > > > > > > > On Tue, Jul 11, 2023 at 2:58 PM Dong Lin <lindon...@gmail.com> > wrote: > > > > > > > > > Hi Jing, > > > > > > > > > > Thank you for the comments! Please see my reply inline. > > > > > > > > > > On Tue, Jul 11, 2023 at 5:41 AM Jing Ge <j...@ververica.com.invalid > > > > > > > wrote: > > > > > > > > > > > Hi Dong, > > > > > > > > > > > > Thanks for the proposal! The FLIP is already in good shape. I got > > > some > > > > > NIT > > > > > > questions. > > > > > > > > > > > > 1. It is a little bit weird to write the hint right after the > > > > motivation > > > > > > that some features have been moved to FLIP-331, because at that > > time, > > > > > > readers don't know the context about what features does it mean. > I > > > > would > > > > > > suggest moving the note to the beginning of "Public interfaces" > > > > sections. > > > > > > > > > > > > > > > > Given that the reviewer who commented on this email thread before I > > > > > refactored the FLIP (i.e. Piotr) has read FLP-331, I think it is > > > simpler > > > > to > > > > > just remove any mention of FLIP-331. I have updated the FLIP > > > accordingly. > > > > > > > > > > > > > > > > 2. It is also a little bit weird to describe all behaviour > changes > > at > > > > > first > > > > > > but only focus on one single feature, i.e. how to implement > > > > > > internalSorterSupported. TBH, I was lost while I was reading the > > > Public > > > > > > interfaces. Maybe change the FLIP title? Another option could be > to > > > > > write a > > > > > > short summary of all features and point out that this FLIP will > > only > > > > > focus > > > > > > on the internalSorterSupported feature. Others could be found in > > > > > FLIP-331. > > > > > > WDYT? > > > > > > > > > > > > > > > > Conceptually, the purpose of this FLIP is to allow a stream mode > job > > to > > > > run > > > > > parts of the topology in batch mode so that it can apply > > > > > optimizations/computations that can not be used together with > > > > checkpointing > > > > > (and thus not usable in stream mode). Although internal sorter is > the > > > > only > > > > > optimization immediately supported in this FLIP, this FLIP lays the > > > > > foundation to support other optimizations in the future, such as > > using > > > > GPU > > > > > to process a bounded stream of records. > > > > > > > > > > Therefore, I find it better to keep the current title rather than > > > > limiting > > > > > the scope to internal sorter. What do you think? > > > > > > > > > > > > > > > > > > > > > 3. There should be a typo at 4) Checkpoint and failover strategy > -> > > > > Mixed > > > > > > mode -> > > > > > > > > > > > > - If any task fails when isBacklog=false true, this task is > > > > restarted > > > > > to > > > > > > re-process its input from the beginning. > > > > > > > > > > > > > > > > > Thank you for catching this issue. It is fixed now. > > > > > > > > > > Best, > > > > > Dong > > > > > > > > > > > > > > > > > > > > > > > > > > > > Best regards > > > > > > Jing > > > > > > > > > > > > > > > > > > On Thu, Jul 6, 2023 at 1:24 PM Dong Lin <lindon...@gmail.com> > > wrote: > > > > > > > > > > > > > Hi Piotr, > > > > > > > > > > > > > > Thanks for your comments! Please see my reply inline. > > > > > > > > > > > > > > On Wed, Jul 5, 2023 at 11:44 PM Piotr Nowojski < > > > > > piotr.nowoj...@gmail.com > > > > > > > > > > > > > > wrote: > > > > > > > > > > > > > > > Hi Dong, > > > > > > > > > > > > > > > > I have a couple of questions. > > > > > > > > > > > > > > > > Could you explain why those properties > > > > > > > > > > > > > > > > @Nullable private Boolean isOutputOnEOF = null; > > > > > > > > @Nullable private Boolean isOutputOnCheckpoint = null; > > > > > > > > @Nullable private Boolean isInternalSorterSupported = > null; > > > > > > > > > > > > > > > > must be `@Nullable`, instead of having the default value set > to > > > > > > `false`? > > > > > > > > > > > > > > > > > > > > > > By initializing these private variables in > > > OperatorAttributesBuilder > > > > as > > > > > > > null, we can implement `OperatorAttributesBuilder#build()` in > > such > > > a > > > > > way > > > > > > > that it can print DEBUG level logging to say > > "isOutputOnCheckpoint > > > is > > > > > not > > > > > > > explicitly set". This can help user/SRE debug performance > issues > > > (or > > > > > lack > > > > > > > of the expected optimization) due to operators not explicitly > > > setting > > > > > the > > > > > > > right operator attribute. > > > > > > > > > > > > > > For example, we might want a job to always use the longer > > > > checkpointing > > > > > > > interval (i.e. execution.checkpointing.interval-during-backlog) > > if > > > > all > > > > > > > running operators have isOutputOnCheckpoint==false, and use the > > > short > > > > > > > checkpointing interval otherwise. If a user has explicitly > > > configured > > > > > the > > > > > > > execution.checkpointing.interval-during-backlog but the > two-phase > > > > > commit > > > > > > > sink library has not been upgraded to set > > > isOutputOnCheckpoint=true, > > > > > then > > > > > > > the job will end up using the long checkpointing interval, and > it > > > > will > > > > > be > > > > > > > useful to figure out what is going wrong in this case by > checking > > > the > > > > > > log. > > > > > > > > > > > > > > Note that the default value of these fields of the > > > OperatorAttributes > > > > > > > instance built by OperatorAttributesBuilder will still be > false. > > > The > > > > > > > following is mentioned in the Java doc of > > > > > > > `OperatorAttributesBuilder#build()`: > > > > > > > > > > > > > > /** > > > > > > > * If any operator attribute is null, we will log it at DEBUG > > > level > > > > > and > > > > > > > use the following > > > > > > > * default values. > > > > > > > * - isOutputOnEOF defaults to false > > > > > > > * - isOutputOnCheckpoint defaults to false > > > > > > > * - isInternalSorterSupported defaults to false > > > > > > > */ > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Second question, have you thought about cases where someone > is > > > > > > > > either bootstrapping from a streaming source like Kafka > > > > > > > > or simply trying to catch up after a long period of downtime > > in a > > > > > > purely > > > > > > > > streaming job? Generally speaking a cases where > > > > > > > > user doesn't care about latency in the catch up phase, > > regardless > > > > if > > > > > > the > > > > > > > > source is bounded or unbounded, but wants to process > > > > > > > > the data as fast as possible, and then switch dynamically to > > real > > > > > time > > > > > > > > processing? > > > > > > > > > > > > > > > > > > > > > > Yes, I have thought about this. We should allow this job to > > > > effectively > > > > > > run > > > > > > > in batch mode when the job is in the catch-up phase. FLIP-327 > is > > > > > actually > > > > > > > an important step toward addressing this use-case. > > > > > > > > > > > > > > In order to address the above use-case, all we need is a way > for > > > > source > > > > > > > operator (e.g. Kafka) to tell Flink runtime (via > > > IsProcessingBacklog) > > > > > > > whether it is in the catch-up phase. > > > > > > > > > > > > > > Since every Kafka message has event-timestamp, we can allow > users > > > to > > > > > > > specify a job-level config such as > > backlog-watermark-lag-threshold, > > > > and > > > > > > > consider a Kafka Source to have IsProcessingBacklog=true if > > > > > system_time - > > > > > > > watermark > backlog-watermark-lag-threshold. This effectively > > > allows > > > > us > > > > > > to > > > > > > > determine whether Kafka is in the catch up phase. > > > > > > > > > > > > > > Once we have this capability (I plan to work on this in > > FLIP-328), > > > we > > > > > can > > > > > > > directly use the features proposed in FLIP-325 and FLIP-327 to > > > > optimize > > > > > > the > > > > > > > above use-case. > > > > > > > > > > > > > > What do you think? > > > > > > > > > > > > > > Best, > > > > > > > Dong > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Best, > > > > > > > > Piotrek > > > > > > > > > > > > > > > > niedz., 2 lip 2023 o 16:15 Dong Lin <lindon...@gmail.com> > > > > > napisał(a): > > > > > > > > > > > > > > > > > Hi all, > > > > > > > > > > > > > > > > > > I am opening this thread to discuss FLIP-327: Support > > > > stream-batch > > > > > > > > unified > > > > > > > > > operator to improve job throughput when processing backlog > > > data. > > > > > The > > > > > > > > design > > > > > > > > > doc can be found at > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-327%3A+Support+stream-batch+unified+operator+to+improve+job+throughput+when+processing+backlog+data > > > > > > > > > . > > > > > > > > > > > > > > > > > > This FLIP enables a Flink job to initially operate in batch > > > mode, > > > > > > > > achieving > > > > > > > > > high throughput while processing records that do not > require > > > low > > > > > > > > processing > > > > > > > > > latency. Subsequently, the job can seamlessly transition to > > > > stream > > > > > > mode > > > > > > > > for > > > > > > > > > processing real-time records with low latency. Importantly, > > the > > > > > same > > > > > > > > state > > > > > > > > > can be utilized before and after this mode switch, making > it > > > > > > > particularly > > > > > > > > > valuable when users wish to bootstrap the job's state using > > > > > > historical > > > > > > > > > data. > > > > > > > > > > > > > > > > > > We would greatly appreciate any comments or feedback you > may > > > have > > > > > on > > > > > > > this > > > > > > > > > proposal. > > > > > > > > > > > > > > > > > > Cheers, > > > > > > > > > Dong > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
