[jira] [Commented] (KAFKA-9148) Consider forking RocksDB for Streams

["Sophie Blee-Goldman (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-9148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17121356#comment-17121356 ] 

Sophie Blee-Goldman commented on KAFKA-9148:
--------------------------------------------

Unfortunately we still haven't been able to upgrade rocks to 6.x. But it seems like we may be bumping the major version of Kafka soon, possibly even for the next release, and we'll be able to upgrade to the new versions. Looking forward to doing so since it looks like there have been a number of improvements to the Java layer recently

> Consider forking RocksDB for Streams 
> -------------------------------------
>
>                 Key: KAFKA-9148
>                 URL: https://issues.apache.org/jira/browse/KAFKA-9148
>             Project: Kafka
>          Issue Type: Improvement
>          Components: streams
>            Reporter: Sophie Blee-Goldman
>            Priority: Major
>
> We recently upgraded our RocksDB dependency to 5.18 for its memory-management abilities (namely the WriteBufferManager, see KAFKA-8215). Unfortunately, someone from Flink recently discovered a ~8% [performance regression|https://github.com/facebook/rocksdb/issues/5774] that exists in all versions 5.18+ (up through the current newest version, 6.2.2). Flink was able to react to this by downgrading to 5.17 and [picking the WriteBufferManage|https://github.com/dataArtisans/frocksdb/pull/4]r to their fork (fRocksDB).
> Due to this and other reasons enumerated below, we should consider also forking our own RocksDB for Streams.
> Pros:
>  * We can avoid passing sudden breaking changes on to our users, such removal of methods with no deprecation period (see discussion on KAFKA-8897)
>  * We can pick whichever version has the best performance for our needs, and pick over any new features, metrics, etc that we need to use rather than being forced to upgrade (and breaking user code, introducing regression, etc)
>  * Support for some architectures does not exist in all RocksDB versions, making Streams completely unusable for some users until we can upgrade the rocksdb dependency to one that supports their specific case. It's worth noting that we've only had [one user|https://issues.apache.org/jira/browse/KAFKA-9225] hit this so far (that we know of), and some workarounds have been discussed on the ticket.
>  * The Java API seems to be a very low priority to the rocksdb folks.
>  ** They leave out critical functionality, features, and configuration options that have been in the c++ API for a very long time
>  ** Those that do make it over often have random gaps in the API such as setters but no getters (see [rocksdb PR #5186|https://github.com/facebook/rocksdb/pull/5186])
>  ** Others are poorly designed and require too many trips across the JNI, making otherwise incredibly useful features prohibitively expensive.
>  *** [|#issuecomment-83145980] [Custom Comparator|https://github.com/facebook/rocksdb/issues/538#issuecomment-83145980]: a custom comparator could significantly improve the performance of session windows. This is trivial to do but given the high performance cost of crossing the jni, it is currently only practical to use a c++ comparator
>  *** [Prefix Seek|https://github.com/facebook/rocksdb/issues/6004]: not currently used by Streams but a commonly requested feature, and may also allow improved range queries
>  ** Even when an external contributor develops a solution for poorly performing Java functionality and helpfully tries to contribute their patch back to rocksdb, it gets ignored by the rocksdb people ([rocksdb PR #2283|https://github.com/facebook/rocksdb/pull/2283])
> Cons:
>  * More work (not to be trivialized, the truth is we don't and can't know how much extra work this will ultimately be)
> Given that we rarely upgrade the Rocks dependency, use only some fraction of its features, and would need or want to make only minimal changes ourselves, it seems like we could actually get away with very little extra work by forking rocksdb. Note that as of this writing the frocksdb repo has only needed to open 5 PRs on top of the actual rocksdb (two of them trivial). Of course, the LOE to maintain this will only grow over time, so we should think carefully about whether and when to start taking on this potential burden.
>  



--
This message was sent by Atlassian Jira
(v8.3.4#803005)