Regarding splitting, I think SDF is being split on spark runner, but I agree with Jan's comments about split's contract. Specific SDF is also free to make decisions about how big the minimal split will be and the runner should be able to process that with reasonable resources. E.g. ParquetIO is splitting on format's row groups. If the row group is larger and format contains a lot of well compressed column data, it will challenge memory resources.
Jan, as for suggested options to implement it, I have an MR with approach 1) to translate all SDFs to two-threaded executions. I did consider something like option 3) but I was not sure if it makes sense in general for other runners as well for Spark. It begs a question for me if it ever makes sense to create SDF and want it on Spark not to use 2 thread execution and possibly apply memory pressure. On Mon, Jan 2, 2023 at 4:49 PM Jan Lukavský <[email protected]> wrote: > There are different translations of streaming and batch Pipelines in > SparkRunner, this thread was focused on the batch part, if I understand it > correctly. Unbounded PCollections are not supported in batch Spark (by > definition). I agree that fixing the splitting is a valid option, though it > still requires unnecessarily big heap for buffering and/or might induce > some overhead with splitting the restriction. Not to mention, that the > splitting is somewhat optional in the contract of SDF (the DoFn might not > support it, if it is bounded), so it might not solve the issue for all > SDFs. The source might not even be splittable at all (e.g. a completely > compressed blob, without any blocks). > > Jan > On 1/2/23 16:22, Daniel Collins via dev wrote: > > If spark's SDF solution doesn't support splitting, fixing that seems like > the best solution to me. Splitting is the mechanism exposed by the model to > actually limit the amount of data produced in a bundle. If unsupported, > then unbounded-per-element SDFs wouldn't be supported at all. > > -Daniel > > On Mon, Jan 2, 2023 at 7:46 AM Jan Lukavský <[email protected]> wrote: > >> Hi Jozef, >> >> I agree that this issue is most likely related to Spark for the reason >> how Spark uses functional style for doing flatMap(). >> >> It could be fixed with the following two options: >> >> a) SparkRunner's SDF implementation does not use splitting - it could be >> fixed so that the SDF is stopped after N elements buffered via trySplit, >> buffer gets flushed and the restriction is resumed >> >> b) alternatively use two threads and a BlockingQueue between them, which >> is what you propose >> >> The number of output elements per input element is bounded (we are >> talking about batch case anyway), but bounded does not mean it has to fit >> to memory. Furthermore, unnecessary buffering of large number of elements >> is memory-inefficient, which is why I think that the two-thread approach >> (b) should be the most efficient. The option (a) seems orthogonal and might >> be implemented as well. >> >> It rises the question of how to determine if the runner should do some >> special translation of SDF in this case. There are probably only these >> options: >> >> 1) translate all SDFs to two-thread execution >> >> 2) add runtime flag, that will turn the translation on (once turned on, >> it will translate all SDFs) - this is the current proposal >> >> 3) extend @DoFn.BoundedPerElement annotation with some kind of >> (optional) hint - e.g. @DoFn.BoundedPerElement(Bounded.POSSIBLY_HUGE), the >> default would be Bounded.FITS_IN_MEMORY (which is the current approach) >> >> The approach (3) seems to give more information to all runners and might >> result in the ability to apply various optimizations for multiple runners, >> so I'd say that this might be the ideal variant. >> >> Jan >> On 12/29/22 13:07, Jozef Vilcek wrote: >> >> I am surprised to hear that Dataflow runner ( which I never used ) would >> have this kind oflimitation. I see that the `OutputManager` interface is >> implemented to write to `Receiver` [1] which follows the push model. Do you >> have a reference I can take a look to review the must fit memory >> limitation? >> >> In Spark, the problem is that the leaf operator pulls data from previous >> ones by consuming an `Iterator` of values. As per your suggestion, this is >> not a problem with `sources` because they hold e.g. source file and can >> pull data as they are being requested. This gets problematic exactly with >> SDF and flatMaps and not sources. It could be one of the reasons why SDF >> performed badly on Spark where community reported performance degradation >> [2] and increases memory use [3] >> >> My proposed solution is to, similar as Dataflow, use `Receiver`-like >> implementation for DoFns which can output large number of elements. For >> now, this WIP targets SDFs only. >> >> [1] >> https://github.com/apache/beam/blob/v2.43.0/runners/google-cloud-dataflow-java/worker/src/main/java/org/apache/beam/runners/dataflow/worker/SimpleParDoFn.java#L285 >> [2] https://github.com/apache/beam/pull/14755 >> [3] >> https://issues.apache.org/jira/browse/BEAM-10670?focusedCommentId=17332005&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-17332005 >> >> On Wed, Dec 28, 2022 at 8:52 PM Daniel Collins via dev < >> [email protected]> wrote: >> >>> I believe that for dataflow runner, the result of processElement must >>> also fit in memory, so this is not just a constraint for the spark runner. >>> >>> The best approach at present might be to convert the source from a >>> flatMap to an SDF that reads out chunks of the file at a time, and supports >>> runner checkpointing (i.e. with a file seek point to resume from) to chunk >>> your data in a way that doesn't require the runner to support unbounded >>> outputs from any individual @ProcessElements downcall. >>> >>> -Daniel >>> >>> On Wed, Dec 28, 2022 at 1:36 PM Jozef Vilcek <[email protected]> >>> wrote: >>> >>>> Hello, >>>> >>>> I am working on an issue which currently limits spark runner by >>>> requiring the result of processElement to fit the memory [1]. This is >>>> problematic e.g for flatMap where the input element is file split and >>>> generates possibly large output. >>>> >>>> The intended fix is to add an option to have dofn processing over input >>>> in one thread and consumption of outputs and forwarding them to downstream >>>> operators in another thread. One challenge for me is to identify which DoFn >>>> should be using this async approach. >>>> >>>> Here [2] is a commit which is WIP and use async processing only for SDF >>>> naive expansion. I would like to get feedback on: >>>> >>>> 1) does the approach make sense overall >>>> >>>> 2) to target DoFn which needs an async processing __ generates possibly >>>> large output __ I am currently just checking if it is DoFn of SDF naive >>>> expansion type [3]. I failed to find a better / more systematic approach >>>> for identifying which DoFn should benefit from that. I would appreciate any >>>> thoughts how to make this better. >>>> >>>> 3) Config option and validatesRunner tests - do we want to make it >>>> possible to turn async DoFn off? If yes, do we want to run validatesRunner >>>> tests for borth options? How do I make sure of that? >>>> >>>> Looking forward to the feedback. >>>> Best, >>>> Jozef >>>> >>>> [1] https://github.com/apache/beam/issues/23852 >>>> [2] >>>> https://github.com/JozoVilcek/beam/commit/895c4973fe9adc6225fcf35d039e3eb1a81ffcff >>>> [3] >>>> https://github.com/JozoVilcek/beam/commit/895c4973fe9adc6225fcf35d039e3eb1a81ffcff#diff-bd72087119a098aa8c947d0989083ec9a6f2b54ef18da57d50e0978799c79191R362 >>>> >>>>
