Hi everyone, Thanks for the proposal and the discussion, I have some remarks: (I'm not very familiar with the new Sink API but I thought about the same problem in context of the changelog state backend)
1. Merging artifacts from multiple checkpoints would apparently require multiple concurrent checkpoints (otherwise, a new checkpoint won't be started before completing the previous one; and the previous one can't be completed before durably storing the artifacts). However, concurrent checkpoints are currently not supported with Unaligned checkpoints (this is besides increasing e2e-latency). 2. Asynchronous merging in an aggregator would require some resolution logic on recovery, so that a merged artifact can be used if the original one was deleted. Otherwise, wouldn't recovery fail because some artifacts are missing? We could also defer deletion until the "compacted" checkpoint is subsumed - but isn't it too late, as it will be deleted anyways once subsumed? 3. Writing small files, then reading and merging them for *every* checkpoint seems worse than only reading them on recovery. I guess I'm missing some cases of reading, so to me it would make sense to mention these cases explicitly in the FLIP motivation section. 4. One way to avoid write-read-merge is by wrapping SinkWriter with another one, which would buffer input elements in a temporary storage (e.g. local file) until a threshold is reached; after that, it would invoke the original SinkWriter. And if a checkpoint barrier comes in earlier, it would send written data to some aggregator. It will increase checkpoint delay (async phase) compared to the current Flink; but not compared to the write-read-merge solution, IIUC. Then such "BufferingSinkWriters" could aggregate input elements from each other, potentially recursively (I mean something like https://cwiki.apache.org/confluence/download/attachments/173082889/DSTL-DFS-DAG.png ) 5. Reducing the number of files by reducing aggregator parallelism as opposed to merging on reaching size threshold will likely be less optimal and more difficult to configure. OTH, thresholds might be more difficult to implement and (with recursive merging) would incur higher latency. Maybe that's also something to decide explicitly or at least mention in the FLIP. Regards, Roman On Tue, Nov 9, 2021 at 5:23 AM Reo Lei <leinuo...@gmail.com> wrote: > > Hi Fabian, > > Thanks for drafting the FLIP and trying to support small file compaction. I > think this feature is very urgent and valuable for users(at least for me). > > Currently I am trying to support streaming rewrite(compact) for Iceberg on > PR#3323 <https://github.com/apache/iceberg/pull/3323>. As Steven mentioned, > Iceberg sink and compact data through the following steps: > Step-1: Some parallel data writer(sinker) to write streaming data as files. > Step-2: A single parallelism data files committer to commit the completed > files as soon as possible to make them available. > Step-3: Some parallel file rewriter(compactor) to collect committed files > from multiple checkpoints, and rewriter(compact) them together once the > total file size or number of files reach the threshold. > Step-4: A single parallelism rewrite(compact) result committer to commit > the rewritten(compacted) files to replace the old files and make them > available. > > > If Flink want to support small file compaction, some key point I think is > necessary: > > 1, Compact files from multiple checkpoints. > I totally agree with Jingsong, because completed file size usually could > not reach the threshold in a single checkpoint. Especially for partitioned > table, we need to compact the files of each partition, but usually the file > size of each partition will be different and may not reach the merge > threshold. If we compact these files, in a single checkpoint, regardless of > whether the total file size reaches the threshold, then the value of > compacting will be diminished and we will still get small files because > these compacted files are not reach to target size. So we need the > compactor to collect committed files from multiple checkpoints and compact > them until they reach the threshold. > > 2, Separate write phase and compact phase. > Users usually hope the data becomes available as soon as possible, and the > end-to-end latency is very important. I think we need to separate the > write and compact phase. For the write phase, there include the Step-1 > and Step-2, we sink data as file and commit it pre checkpoint and regardless > of whether the file size it is. That could ensure the data will be > available ASAP. For the compact phase, there include the Step-3 > and Step-4, the compactor should collect committed files from multiple > checkpoints and compact them asynchronously once they reach the threshold, > and the compact committer will commit the compaction result in the next > checkpoint. We compact the committed files asynchronously because we don't > want the compaction to affect the data sink or the whole pipeline. > > 3, Exactly once guarantee between write and compact phase. > Once we separate write phase and compact phase, we need to consider > how to guarantee > the exact once semantic between two phases. We should not lose any data or > files on the compactor(Step-3) in any case and cause the compaction result > to be inconsistent with before. I think flink should provide an easy-to-use > interface to make that easier. > > 4, Metadata operation and compaction result validation. > In the compact phase, there may be not only compact files, but also a lot > of metadata operations, such as the iceberg needing to read/write manifest > and do MOR. And we need some interface to support users to do some > validation of the compaction result. I think these points should be > considered when we design the compaction API. > > > Back to FLIP-191, option 1 looks very complicated while option 2 is > relatively simple, but neither of these two solutions separates the write > phase from the compact phase. So I think we should consider the points I > mentioned above. And if you have any other questions you can always feel > free to reach out to me! > > BR, > Reo > > Fabian Paul <fabianp...@ververica.com> 于2021年11月8日周一 下午7:59写道： > > > Hi all, > > > > Thanks for the lively discussions. I am really excited to see so many > > people > > participating in this thread. It also underlines the need that many people > > would > > like to see a solution soon. > > > > I have updated the FLIP and removed the parallelism configuration because > > it is > > unnecessary since users can configure a constant exchange key to send all > > committables to only one committable aggregator. > > > > > > 1. Burden for developers w.r.t batch stream unification. > > > > @yun @guowei, from a theoretical point you are right about exposing the > > DataStream > > API in the sink users have the full power to write correct batch and > > streaming > > sinks. I think in reality a lot of users still struggle to build pipelines > > with > > i.e. the operator pipeline which works correct in streaming and batch mode. > > Another problem I see is by exposing more deeper concepts is that we > > cannot do > > any optimization because we cannot reason about how sinks are built in the > > future. > > > > We should also try to steer users towards using only `Functions` to give > > us more > > flexibility to swap the internal operator representation. I agree with > > @yun we > > should try to make the `ProcessFunction` more versatile to work on that > > goal but > > I see this as unrelated to the FLIP. > > > > > > 2. Regarding Commit / Global commit > > > > I envision the global committer to be specific depending on the data lake > > solution you want to write to. However, it is entirely orthogonal to the > > compaction. > > Currently, I do not expect any changes w.r.t the Global commit introduces > > by > > this FLIP. > > > > > > 3. Regarding the case of trans-checkpoints merging > > > > @yun, as user, I would expect that if the committer receives in a > > checkpoint files > > to merge/commit that these are also finished when the checkpoint finishes. > > I think all sinks rely on this principle currently i.e., KafkaSink needs to > > commit all open transactions until the next checkpoint can happen. > > > > Maybe in the future, we can somehow move the Committer#commit call to an > > asynchronous execution, but we should discuss it as a separate thread. > > > > > We probably should first describe the different causes of small files and > > > what problems was this proposal trying to solve. I wrote a data shuffling > > > proposal [1] for Flink Iceberg sink (shared with Iceberg community [2]). > > It > > > can address small files problems due to skewed data distribution across > > > Iceberg table partitions. Streaming shuffling before writers (to files) > > is > > > typically more efficient than post-write file compaction (which involves > > > read-merge-write). It is usually cheaper to prevent a problem (small > > files) > > > than fixing it. > > > > > > @steven you are raising a good point, although I think only using a > > customizable > > shuffle won't address the generation of small files. One assumption is that > > at least the sink generates one file per subtask, which can already be too > > many. > > Another problem is that with low checkpointing intervals, the files do not > > meet > > the required size. The latter point is probably addressable by changing the > > checkpoint interval, which might be inconvenient for some users. > > > > > The sink coordinator checkpoint problem (mentioned in option 1) would be > > > great if Flink can address it. In the spirit of source > > (enumerator-reader) > > > and sink (writer-coordinator) duality, sink coordinator checkpoint should > > > happen after the writer operator. This would be a natural fit to support > > > global committer in FLIP-143. It is probably an orthogonal matter to this > > > proposal. > > > > > > To me the question here is what are the benefits of having a coordinator in > > comparison to a global committer/aggregator operator. > > > > > Personally, I am usually in favor of keeping streaming ingestion (to data > > > lake) relatively simple and stable. Also sometimes compaction and sorting > > > are performed together in data rewrite maintenance jobs to improve read > > > performance. In that case, the value of compacting (in Flink streaming > > > ingestion) diminishes. > > > > > > I agree it is always possible to have scheduled maintenance jobs keeping > > care of > > your data i.e., doing compaction. Unfortunately, the downside is that you > > have to your data after it is already available for other downstream > > consumers. > > I guess this can lead to all kinds of visibility problems. I am also > > surprised that > > you personally are a fan of this approach and, on the other hand, are > > developing > > the Iceberg sink, which goes somewhat against your mentioned principle of > > keeping > > the sink simple. > > > > > Currently, it is unclear from the doc and this thread where the > > compaction > > > is actually happening. Jingsong's reply described one model > > > writer (parallel) -> aggregator (single-parallelism compaction planner) > > -> > > > compactor (parallel) -> global committer (single-parallelism) > > > > > > My idea of the topology is very similar to the one outlined by Jinsong. The > > compaction will happen in the committer operator. > > > > > > > > In the Iceberg community, the following model has been discussed. It is > > > better for Iceberg because it won't delay the data availability. > > > writer (parallel) -> global committer for append (single parallelism) -> > > > compactor (parallel) -> global committer for rewrite commit (single > > > parallelism) > > > > > > From a quick glimpse, it seems that the exact same topology is possible to > > express with the committable aggregator, but this definitely depends on > > the exact > > setup. > > > > Best, > > Fabian
