Hey Shawn, Rahil, Thanks for raising this issue. These are good suggestions; I would recommend simplifying the code structure of Hudi Spark incrementally and gradually making the code less coupled with Spark engine.
Identify breaking changes introduced by the new Spark version and patch > affected Hudi classes. This is important. No matter how we organize the code structure, we need to understand breaking changes from Spark that can affect Hudi. Right now, only a handful of Spark experts in our community have such knowledge and how Hudi integrates with Spark at the implementation level. We should document the integration, e.g., in an RFC, to avoid knowledge gaps. Based on the discussion, we should also write down the formal process of supporting a new Spark version in Hudi, with clear testing and certification criteria. The current structure involves common code shared by several Spark > versions, such as hudi-spark-common, hudi-spark3-common, > hudi-spark3.2plus-common, etc. (a detailed description can be found in the > readme here: > https://github.com/apache/hudi/blob/master/hudi-spark-datasource/README.md). > This setup aims to minimize duplicate code in Hudi. Hudi currently utilizes > the SparkAdapter to invoke specific code based on the Spark version, > allowing different Spark versions to trigger different logic. We took the current approach of having hudi-spark3.2plus-common module to deduplicate the code between Spark 3.2 and Spark 3.3 integrations ( https://issues.apache.org/jira/browse/HUDI-4691), because we anticipated that this should reduce the code duplication going forward. Now it proved inflexible given Spark 3.4 changes some APIs and hudi-spark3.2plus-common is no longer "common" any more. It makes sense to have Spark version-specific module to contain classes for specific integration logic. I would still keep common modules for classes that are more Hudi-relevant and applicable to all Spark or Spark 2/3 versions. How we draw the line depends on the implementation, which we can gradually make the decision. This should solve the problem of applying a general fix for multiple Spark versions. For Spark 3.4 integration, I suggest duplicating the code for now in a separate module, while we figure out a better code structure for Hudi Spark integration. Currently I think our integration with Spark is too tight, and brings up > serious issues when upgrading. > I will describe one example(however there are many more) but one area is > we extend Spark's *ParquetFileFormat* in the following classes: > buildReaderWithPartitionValues method I agree that Hudi integration with Spark should not be tightly coupled. Yet we should also honor the fact that some are for functionality and performance reasons (which we should clearly document). We should definitely revisit such coupling points and see if they can be relaxed. The Hoodie Parquet format classes are introduced mainly for supporting full schema evaluation in Spark (https://github.com/apache/hudi/pull/4910). AFAIK, in Spark32HoodieParquetFileFormat class, we did use default ParquetFileFormat logic when schema evaluation is not enabled ( https://github.com/apache/hudi/pull/4910/files#diff-fc11bc10091e5312b58068f263960ef9459b1d01cf08512f33362b76f5554416R61), meaning that it did not change any parquet reading logic. However, such fallback/default behavior is removed later on. I think we need to revisit such a decision and bring back the default behavior if OK. In the future, we should understand the implications and review such code changes carefully. Best, - Ethan On Fri, Jun 2, 2023 at 12:28 AM Vinoth Chandar <vin...@apache.org> wrote: > This is a good topic, thanks for raising this. Overall our reliance on > spark classes/APIs that are declared experimental is an issue on paper. But > there is few other ways to get right performance without relying on these. > This has been the tricky issue IMO. Thoughts? > > I ll review the code organization more carefully and report back. > > On Fri, Jun 2, 2023 at 04:23 Rahil C <rchert...@gmail.com> wrote: > > > Thanks Shawn for writing this, I would like to also add on to the Spark > > Discussion. > > > > Currently I think our integration with Spark is too tight, and brings up > > serious issues when upgrading. > > > > I will describe one example(however there are many more) but one area is > we > > extend Spark's *ParquetFileFormat* in the following classes. > > > > > > > https://github.com/apache/hudi/blob/master/hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/HoodieParquetFileFormat.scala > > > > > https://github.com/apache/hudi/blob/master/hudi-spark-datasource/hudi-spark3.2plus-common/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/Spark32PlusHoodieParquetFileFormat.scala > > > > and specifically the main logic changes is we override > > *buildReaderWithPartitionValues > > *method > > *.* > > I understand the pro of reusability of spark's code, but the con is that > we > > dont then get the latest changes from the latest implementation of these > > methods. This gets more complex as we then need to understand which spark > > changes are required to cherry pick over as spark upgrades, such as these > > issues. > > > > For spark 3.3.2 we faced several issues documented here > > https://github.com/apache/hudi/pull/8082, > > and for spark 3.4.0 we have encountered several issues as well. > > https://github.com/apache/hudi/pull/8682 > > > > We also are not keeping up to date with certain spark features as a > result > > of the integration we have made. I have created a JIRA that goes more > into > > this in-depth to this. > > https://issues.apache.org/jira/browse/HUDI-6262 > > > > Would be happy to sync with other hudi spark committers/experts, or > anyone > > interested in revisiting this integration so that future spark work will > be > > more achievable. > > > > Regards, > > Rahil Chertara > > > > On Tue, May 23, 2023 at 8:16 PM Shawn Chang <yc2...@cornell.edu> wrote: > > > > > Hi Hudi developers, > > > > > > I am writing to discuss the current code structure of the existing > > > hudi-spark-datasource and propose a more scalable approach for > supporting > > > multiple Spark versions. The current structure involves common code > > shared > > > by several Spark versions, such as hudi-spark-common, > hudi-spark3-common, > > > hudi-spark3.2plus-common, etc. (a detailed description can be found in > > the > > > readme here: > > > > > > https://github.com/apache/hudi/blob/master/hudi-spark-datasource/README.md > > > ). > > > This setup aims to minimize duplicate code in Hudi. Hudi currently > > utilizes > > > the SparkAdapter to invoke specific code based on the Spark version, > > > allowing different Spark versions to trigger different logic. > > > > > > However, this code structure proves to be complex and hampers the > process > > > of adding support for newer Spark versions. The current approach > involves > > > the following steps: > > > 1) Identify breaking changes introduced by the new Spark version and > > patch > > > affected Hudi classes. > > > 2) Separate affected Hudi classes into different folders so that older > > > Spark versions can continue using the existing logic, while the new > Spark > > > version can work with the updated Hudi classes. > > > 3) Connect SparkAdapter to these Hudi classes, enabling Hudi to utilize > > the > > > correct code based on the Spark version. > > > 4) Collect common code and place it in a new folder, such as > > > hudi-spark3.2plus-common, to reduce duplicate code. > > > > > > This convoluted process has significantly slowed down the pace of > adding > > > support for newer Spark versions in Hudi. Fortunately, there is a > simpler > > > alternative that can streamline the process. I propose removing the > > common > > > modules and having only one folder for each Spark version. For example: > > > > > > > > > > > > > > > > > > > > > *hudi-spark-datasource/---hudi-spark2.4.0/---hudi-spark3.2.0/---hudi-spark3.3.0/...* > > > > > > With this revised code structure, each Spark version will have its own > > > corresponding Hudi module. The process of adding Spark support will be > > > simplified as follows: > > > 1) Copy the latest existing hudi-spark module to a new module, > > > hudi-spark<new_Spark_version>. > > > 2) Identify breaking changes introduced by the new Spark version and > > patch > > > affected Hudi classes. > > > > > > Let's consider some pros and cons of this new code structure: > > > *Pros:* > > > -A more readable codebase, with each Spark version having its > individual > > > module. > > > -Easier addition of support for new Spark versions by duplicating the > > most > > > recent module and making necessary modifications. > > > -Simpler implementation of improvements specific to a particular Spark > > > version. > > > *Cons:* > > > -Increased duplicate code (though this shouldn't impact the Hudi jar > size > > > during runtime, as the jar will still only contain support for one > Spark > > > version). > > > -When applying a general fix for multiple Spark versions, the fix needs > > to > > > be applied to different Spark modules instead of a common codebase. > > > > > > Please feel free to share your opinion, any feedback would be welcome! > > > > > > Thank you. > > > > > > Best, > > > Shawn > > > > > >
