Hi David, I don't find it troublesome per se, I was rather trying to understand what should be expected (and documented) for my application. Before I restarted the job and changed some configurations, it ran for around 10 days and ended up with a state size of about 1.8GB, so I'm still not sure what is the upper bound in my scenario, or if that amount of "uncompacted garbage" is normal or not (for our throughput). This is important for us because we need to know how to size (disk space) the infrastructure, although it is also having a big impact on timings because each checkpoint ends up requiring 30+ seconds to complete, and they will eventually time out for sure.
I understand RocksDB has different sophisticated mechanisms, so I certainly don't expect one magic button that does exactly what I want, but ideally there would be a way to tune configuration in a way that a rough upper bound estimate of disk space can be deduced. Having some expired state for a while is expected, what I find odd is that it grows so fast, the size of the state quickly outpaces the size of processed events, even though the state only persists a subset of information (some integer ids, string ids, longs for epochs). At this point I think I can conclude that the "live" state from my operators is not growing indefinitely (based on what I see with the state processor API), so is there a way to get a better estimate of disk utilization other than letting the job run and wait? I've been reading through RocksDB documentation as well, but that might not be enough because I don't know how Flink handles its own framework state internally. Regards, Alexis. ________________________________ From: David Anderson <dander...@apache.org> Sent: Friday, April 22, 2022 9:57 AM To: Alexis Sarda-Espinosa <alexis.sarda-espin...@microfocus.com> Cc: ro...@apache.org <ro...@apache.org>; user@flink.apache.org <user@flink.apache.org> Subject: Re: RocksDB's state size discrepancy with what's seen with state processor API Alexis, Compaction isn't an all-at-once procedure. RocksDB is organized as a series of levels, each 10x larger than the one below. There are a few different compaction algorithms available, and they are tunable, but what's typically happening during compaction is that one SST file at level n is being merged into the relevant SST files at level n+1. During this compaction procedure, obsolete and deleted entries are cleaned up. And several such compactions can be occurring concurrently. (Not to mention that each TM has its own independent RocksDB instance.) It's not unusual for jobs with a small amount of state to end up with checkpoints of a few hundred MBs in size, where a lot of that is uncompacted garbage. If you find this troublesome, you could configure RocksDB to compact more frequently. David On Thu, Apr 21, 2022 at 12:49 PM Alexis Sarda-Espinosa <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> wrote: Hello, I enabled some of the RocksDB metrics and I noticed some additional things. After changing the configuration YAML, I restarted the cluster with a savepoint, and I can see that it only used 5.6MB on disk. Consequently, after the job switched to running state, the new checkpoints were also a few MB in size. After running for 1 day, checkpoint size is now around 100MB. From the metrics I can see with the Prometheus reporter: - All entries for num-live-versions show 1 - All entries for compaction-pending show 0 - Most entries for estimate-num-keys are in the range of 0 to 100, although I see a few with 151 coming from flink_taskmanager_job_task_operator__timer_state_event_window_timers_rocksdb_estimate_num_keys Is compaction expected after only 100MB? I imagine not, but if the savepoint shows that the effective amount of data is so low, size growth still seems far too large. In fact, if I only look at the UI, Bytes Received for the relevant SubTasks is about 14MB, yet the latest checkpoint already shows a Data Size of 75MB for said SubTasks. Regards, Alexis. -----Original Message----- From: Roman Khachatryan <ro...@apache.org<mailto:ro...@apache.org>> Sent: Mittwoch, 20. April 2022 10:37 To: Alexis Sarda-Espinosa <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> Cc: user@flink.apache.org<mailto:user@flink.apache.org> Subject: Re: RocksDB's state size discrepancy with what's seen with state processor API State Processor API works on a higher level and is not aware of any RocksDB specifics (in fact, it can be used with any backend). Regards, Roman On Tue, Apr 19, 2022 at 10:52 PM Alexis Sarda-Espinosa <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> wrote: > > I can look into RocksDB metrics, I need to configure Prometheus at some point > anyway. However, going back to the original question, is there no way to gain > more insight into this with the state processor API? You've mentioned > potential issues (too many states, missing compaction) but, with my > admittedly limited understanding of the way RocksDB is used in Flink, I would > have thought that such things would be visible when using the state > processor. Is there no way for me to "parse" those MANIFEST files with some > of Flink's classes and get some more hints? > > Regards, > Alexis. > > ________________________________ > From: Roman Khachatryan <ro...@apache.org<mailto:ro...@apache.org>> > Sent: Tuesday, April 19, 2022 5:51 PM > To: Alexis Sarda-Espinosa > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > Cc: user@flink.apache.org<mailto:user@flink.apache.org> > <user@flink.apache.org<mailto:user@flink.apache.org>> > Subject: Re: RocksDB's state size discrepancy with what's seen with > state processor API > > > I assume that when you say "new states", that is related to new descriptors > > with different names? Because, in the case of windowing for example, each > > window "instance" has its own scoped (non-global and keyed) state, but > > that's not regarded as a separate column family, is it? > Yes, that's what I meant, and that's regarded as the same column family. > > Another possible reason is that SST files aren't being compacted and > that increases the MANIFEST file size. > I'd check the total number of loaded SST files and the creation date > of the oldest one. > > You can also see whether there are any compactions running via RocksDB > metrics [1] [2] (a reporter needs to be configured [3]). > > [1] > https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/c > onfig/#state-backend-rocksdb-metrics-num-running-compactions > [2] > https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/c > onfig/#state-backend-rocksdb-metrics-compaction-pending > [3] > https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/m > etric_reporters/#reporters > > Regards, > Roman > > On Tue, Apr 19, 2022 at 1:38 PM Alexis Sarda-Espinosa > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > wrote: > > > > Hi Roman, > > > > I assume that when you say "new states", that is related to new descriptors > > with different names? Because, in the case of windowing for example, each > > window "instance" has its own scoped (non-global and keyed) state, but > > that's not regarded as a separate column family, is it? > > > > For the 3 descriptors I mentioned before, they are only instantiated once > > and used like this: > > > > - Window list state: each call to process() executes > > context.windowState().getListState(...).get() > > - Global map state: each call to process() executes > > context.globalState().getMapState(...) > > - Global list state: within open(), runtimeContext.getListState(...) is > > executed once and used throughout the life of the operator. > > > > According to [1], the two ways of using global state should be equivalent. > > > > I will say that some of the operators instantiate the state descriptor in > > their constructors, i.e. before they are serialized to the TM, but the > > descriptors are Serializable, so I imagine that's not relevant. > > > > [1] https://stackoverflow.com/a/50510054/5793905 > > > > Regards, > > Alexis. > > > > -----Original Message----- > > From: Roman Khachatryan <ro...@apache.org<mailto:ro...@apache.org>> > > Sent: Dienstag, 19. April 2022 11:48 > > To: Alexis Sarda-Espinosa > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > Cc: user@flink.apache.org<mailto:user@flink.apache.org> > > Subject: Re: RocksDB's state size discrepancy with what's seen with > > state processor API > > > > Hi Alexis, > > > > Thanks a lot for the information, > > > > MANIFEST files list RocksDB column families (among other info); ever > > growing size of these files might indicate that some new states are > > constantly being created. > > Could you please confirm that the number of state names is constant? > > > > > Could you confirm if Flink's own operators could be creating state in > > > RocksDB? I assume the window operators save some information in the state > > > as well. > > That's correct, window operators maintain a list of elements per window and > > a set of timers (timestamps). These states' names should be fixed > > (something like "window-contents" and "window-timers"). > > > > > is that related to managed state used by my functions? Or does that > > > indicate size growth is elsewhere? > > The same mechanism is used for both Flink internal state and operator > > state, so it's hard to say without at least knowing the state names. > > > > > > Regards, > > Roman > > > > > > On Tue, Apr 12, 2022 at 2:06 PM Roman Khachatryan > > <ro...@apache.org<mailto:ro...@apache.org>> wrote: > > > > > > /shared folder contains keyed state that is shared among different > > > checkpoints [1]. Most of state should be shared in your case since > > > you're using keyed state and incremental checkpoints. > > > > > > When a checkpoint is loaded, the state that it shares with older > > > checkpoints is loaded as well. I suggested to load different > > > checkpoints (i.e. chk-* folders) and compare the numbers of > > > objects in their states. To prevent the job from discarding the > > > state, it can either be stopped for some time and then restarted > > > from the latest checkpoint; or the number of retained checkpoints > > > can be increased [2]. Copying isn't necessary. > > > > > > Besides that, you can also check state sizes of operator in Flink > > > Web UI (but not the sizes of individual states). If the operators > > > are chained then their combined state size will be shown. To > > > prevent this, you can disable chaining [3] (although this will > > > have performance impact). > > > > > > Individual checkpoint folders should be eventually removed (when > > > the checkpoint is subsumed). However, this is not guaranteed: if > > > there is any problem during deletion, it will be logged, but the > > > job will not fail. > > > > > > [1] > > > https://nightlies.apache.org/flink/flink-docs-master/docs/ops/stat > > > e/ch > > > eckpoints/#directory-structure > > > [2] > > > https://nightlies.apache.org/flink/flink-docs-master/docs/deployme > > > nt/c onfig/#state-checkpoints-num-retained > > > [3] > > > https://nightlies.apache.org/flink/flink-docs-master/docs/dev/data > > > stre am/operators/overview/#disable-chaining > > > > > > Regards, > > > Roman > > > > > > On Tue, Apr 12, 2022 at 12:58 PM Alexis Sarda-Espinosa > > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > > wrote: > > > > > > > > Hi Roman, > > > > > > > > Maybe I'm misunderstanding the structure of the data within the > > > > checkpoint. You suggest comparing counts of objects in different > > > > checkpoints, I assume you mean copying my "checkpoints" folder at > > > > different times and comparing, not comparing different "chk-*" folders > > > > in the same snapshot, right? > > > > > > > > I haven't executed the processor program with a newer checkpoint, but I > > > > did look at the folder in the running system, and I noticed that most > > > > of the chk-* folders have remained unchanged, there's only 1 or 2 new > > > > folders corresponding to newer checkpoints. I would think this makes > > > > sense since the configuration specifies that only 1 completed > > > > checkpoint should be retained, but then why are the older chk-* folders > > > > still there? I did trigger a manual restart of the Flink cluster in the > > > > past (before starting the long-running test), but if my policy is to > > > > CLAIM the checkpoint, Flink's documentation states that it would be > > > > cleaned eventually. > > > > > > > > Moreover, just by looking at folder sizes with "du", I can see that > > > > most of the state is held in the "shared" folder, and that has grown > > > > for sure; I'm not sure what "shared" usually holds, but if that's > > > > what's growing, maybe I can rule out expired state staying around?. My > > > > pipeline doesn't use timers, although I guess Flink itself may use > > > > them. Is there any way I could get some insight into which operator > > > > holds larger states? > > > > > > > > Regards, > > > > Alexis. > > > > > > > > -----Original Message----- > > > > From: Roman Khachatryan <ro...@apache.org<mailto:ro...@apache.org>> > > > > Sent: Dienstag, 12. April 2022 12:37 > > > > To: Alexis Sarda-Espinosa > > > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > > > Cc: user@flink.apache.org<mailto:user@flink.apache.org> > > > > Subject: Re: RocksDB's state size discrepancy with what's seen > > > > with state processor API > > > > > > > > Hi Alexis, > > > > > > > > Thanks a lot for sharing this. I think the program is correct. > > > > Although it doesn't take timers into account; and to estimate the state > > > > size more accurately, you could also use the same serializers used by > > > > the job. > > > > But maybe it makes more sense to compare the counts of objects in > > > > different checkpoints and see which state is growing. > > > > > > > > If the number of keys is small, compaction should eventually clean up > > > > the old values, given that the windows eventually expire. I think it > > > > makes sense to check that watermarks in all windows are making progress. > > > > > > > > Setting ExecutionEnvironment#setParallelism(1) shouldn't affect the > > > > results of the State Processor program. > > > > > > > > Regards, > > > > Roman > > > > > > > > On Mon, Apr 11, 2022 at 12:28 PM Alexis Sarda-Espinosa > > > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > > > wrote: > > > > > > > > > > Some additional information that I’ve gathered: > > > > > > > > > > > > > > > > > > > > The number of unique keys in the system is 10, and that is correctly > > > > > reflected in the state. > > > > > TTL for global window state is set to update on read and write, but > > > > > the code has logic to remove old state based on event time. > > > > > Not sure it’s relevant, but the Flink cluster does run with jemalloc > > > > > enabled. > > > > > GitHub gist with the whole processor setup since it’s not too long: > > > > > https://gist.github.com/asardaes/eaf21f18860ec39b325a40acef2db > > > > > 678 > > > > > > > > > > > > > > > > > > > > Relevant configuration entries (explicitly set, others are left with > > > > > defaults): > > > > > > > > > > > > > > > > > > > > state.backend: rocksdb > > > > > > > > > > state.backend.incremental: true > > > > > > > > > > execution.checkpointing.interval: 30 s > > > > > > > > > > execution.checkpointing.min-pause: 25 s > > > > > > > > > > execution.checkpointing.timeout: 5 min > > > > > > > > > > execution.savepoint-restore-mode: CLAIM > > > > > > > > > > execution.checkpointing.externalized-checkpoint-retention: > > > > > RETAIN_ON_CANCELLATION > > > > > > > > > > > > > > > > > > > > Over the weekend, state size has grown to 1.23GB with the operators > > > > > referenced in the processor program taking 849MB, so I’m still pretty > > > > > puzzled. I thought it could be due to expired state being retained, > > > > > but I think that doesn’t make sense if I have finite keys, right? > > > > > > > > > > > > > > > > > > > > Regards, > > > > > > > > > > Alexis. > > > > > > > > > > > > > > > > > > > > From: Alexis Sarda-Espinosa > > > > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > > > > Sent: Samstag, 9. April 2022 01:39 > > > > > To: ro...@apache.org<mailto:ro...@apache.org> > > > > > Cc: user@flink.apache.org<mailto:user@flink.apache.org> > > > > > Subject: Re: RocksDB's state size discrepancy with what's seen > > > > > with state processor API > > > > > > > > > > > > > > > > > > > > Hi Roman, > > > > > > > > > > > > > > > > > > > > Here's an example of a WindowReaderFunction: > > > > > > > > > > > > > > > > > > > > public class StateReaderFunction extends > > > > > WindowReaderFunction<Pojo, Integer, String, TimeWindow> { > > > > > > > > > > private static final ListStateDescriptor<Integer> LSD > > > > > = new ListStateDescriptor<>( > > > > > > > > > > "descriptorId", > > > > > > > > > > Integer.class > > > > > > > > > > ); > > > > > > > > > > > > > > > > > > > > @Override > > > > > > > > > > public void readWindow(String s, Context<TimeWindow> > > > > > context, Iterable<Pojo> elements, Collector<Integer> out) > > > > > throws Exception { > > > > > > > > > > int count = 0; > > > > > > > > > > for (Integer i : > > > > > context.windowState().getListState(LSD).get()) { > > > > > > > > > > count++; > > > > > > > > > > } > > > > > > > > > > out.collect(count); > > > > > > > > > > } > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > That's for the operator that uses window state. The other readers do > > > > > something similar but with context.globalState(). That should provide > > > > > the number of state entries for each key+window combination, no? And > > > > > after collecting all results, I would get the number of state entries > > > > > across all keys+windows for an operator. > > > > > > > > > > > > > > > > > > > > And yes, I do mean ProcessWindowFunction.clear(). Therein I call > > > > > context.windowState().getListState(...).clear(). > > > > > > > > > > > > > > > > > > > > Side note: in the state processor program I call > > > > > ExecutionEnvironment#setParallelism(1) even though my streaming job > > > > > runs with parallelism=4, this doesn't affect the result, does it? > > > > > > > > > > > > > > > > > > > > Regards, > > > > > > > > > > Alexis. > > > > > > > > > > > > > > > > > > > > ________________________________ > > > > > > > > > > From: Roman Khachatryan <ro...@apache.org<mailto:ro...@apache.org>> > > > > > Sent: Friday, April 8, 2022 11:06 PM > > > > > To: Alexis Sarda-Espinosa > > > > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > > > > Cc: user@flink.apache.org<mailto:user@flink.apache.org> > > > > > <user@flink.apache.org<mailto:user@flink.apache.org>> > > > > > Subject: Re: RocksDB's state size discrepancy with what's seen > > > > > with state processor API > > > > > > > > > > > > > > > > > > > > Hi Alexis, > > > > > > > > > > If I understand correctly, the provided StateProcessor program > > > > > gives you the number of stream elements per operator. However, > > > > > you mentioned that these operators have collection-type states > > > > > (ListState and MapState). That means that per one entry there > > > > > can be an arbitrary number of state elements. > > > > > > > > > > Have you tried estimating the state sizes directly via > > > > > readKeyedState[1]? > > > > > > > > > > > The other operator does override and call clear() > > > > > Just to make sure, you mean ProcessWindowFunction.clear() [2], right? > > > > > > > > > > [1] > > > > > https://nightlies.apache.org/flink/flink-docs-release-1.14/api > > > > > /jav > > > > > a/or > > > > > g/apache/flink/state/api/ExistingSavepoint.html#readKeyedState > > > > > -jav > > > > > a.la<http://a.la> > > > > > ng.String-org.apache.flink.state.api.functions.KeyedStateReade > > > > > rFun > > > > > ctio > > > > > n- > > > > > > > > > > [2] > > > > > https://nightlies.apache.org/flink/flink-docs-release-1.4/api/ > > > > > java > > > > > /org > > > > > /apache/flink/streaming/api/functions/windowing/ProcessWindowFunction. > > > > > html#clear-org.apache.flink.streaming.api.functions.windowing. > > > > > Proc > > > > > essW > > > > > indowFunction.Context- > > > > > > > > > > Regards, > > > > > Roman > > > > > > > > > > > > > > > On Fri, Apr 8, 2022 at 4:19 PM Alexis Sarda-Espinosa > > > > > <alexis.sarda-espin...@microfocus.com<mailto:alexis.sarda-espin...@microfocus.com>> > > > > > wrote: > > > > > > > > > > > > Hello, > > > > > > > > > > > > > > > > > > > > > > > > I have a streaming job running on Flink 1.14.4 that uses managed > > > > > > state with RocksDB with incremental checkpoints as backend. I’ve > > > > > > been monitoring a dev environment that has been running for the > > > > > > last week and I noticed that state size and end-to-end duration > > > > > > have been increasing steadily. Currently, duration is 11 seconds > > > > > > and size is 917MB (as shown in the UI). The tasks with the largest > > > > > > state (614MB) come from keyed sliding windows. Some attributes of > > > > > > this job’s setup: > > > > > > > > > > > > > > > > > > > > > > > > Windows are 11 minutes in size. > > > > > > Slide time is 1 minute. > > > > > > Throughput is approximately 20 events per minute. > > > > > > > > > > > > > > > > > > > > > > > > I have 3 operators with these states: > > > > > > > > > > > > > > > > > > > > > > > > Window state with ListState<Integer> and no TTL. > > > > > > Global window state with MapState<Long, List<String>> and a TTL of > > > > > > 1 hour (with cleanupInRocksdbCompactFilter(1000L)). > > > > > > Global window state with ListState<Pojo> where the Pojo has an int > > > > > > and a long, a TTL of 1 hour, and configured with > > > > > > cleanupInRocksdbCompactFilter(1000L) as well. > > > > > > > > > > > > > > > > > > > > > > > > Both operators with global window state have logic to manually > > > > > > remove old state in addition to configured TTL. The other operator > > > > > > does override and call clear(). > > > > > > > > > > > > > > > > > > > > > > > > I have now analyzed the checkpoint folder with the state processor > > > > > > API, and I’ll note here that I see 50 folders named chk-*** even > > > > > > though I don’t set state.checkpoints.num-retained and the default > > > > > > should be 1. I loaded the data from the folder with the highest chk > > > > > > number and I see that my operators have these amounts respectively: > > > > > > > > > > > > > > > > > > > > > > > > 10 entries > > > > > > 80 entries > > > > > > 200 entries > > > > > > > > > > > > > > > > > > > > > > > > I got those numbers with something like this: > > > > > > > > > > > > > > > > > > > > > > > > savepoint > > > > > > > > > > > > > > > > > > .window(SlidingEventTimeWindows.of(Time.minutes(11L), > > > > > > Time.minutes(1L))) > > > > > > > > > > > > .process(...) > > > > > > > > > > > > .collect() > > > > > > > > > > > > .parallelStream() > > > > > > > > > > > > .reduce(0, Integer::sum); > > > > > > > > > > > > > > > > > > > > > > > > Where my WindowReaderFunction classes just count the number of > > > > > > entries in each call to readWindow. > > > > > > > > > > > > > > > > > > > > > > > > Those amounts cannot possibly account for 614MB, so what am I > > > > > > missing? > > > > > > > > > > > > > > > > > > > > > > > > Regards, > > > > > > > > > > > > Alexis. > > > > > > > > > > > >
