Consider removing StorageEmbeddedRocksDB.

[alexey-milovidov]

Bugs have been found:

facebook/rocksdb#7711
facebook/rocksdb#7821
facebook/rocksdb#7778

Are there use cases that will motivate against removal?

[chenziliang]

We are using rocksdb table engine as a kv store for now since we don't wanna introduce another kv store dependency. If we get this removed, may i ask if there is any other alternative in CH (ReplacingMergeTree etc may not be as convenient as kv store) ?

[alexey-milovidov]

@chenziliang Thank you! Now the chance it will be removed is low.

[alexey-milovidov]

But we have to fix bugs, otherwise it will make constant noise in our CI system.

[alexey-milovidov]

@sundy-li Maybe you want to help fixing the remaining use-after-free error? facebook/rocksdb#7821

[sundy-li]

@sundy-li Maybe you want to help fixing the remaining use-after-free error? facebook/rocksdb#7821

Sure.

BTW I was not aware that rocksdb may introduce lots of bugs to broke ClickHouse. Because I thought there were many newsql database built base on rocksdb (cockroachdb, tidb, yugabytedb), it proves that ClickHouse has better tests systems than other databases . I am not rocksdb expert, but I'm willing to help it work better in ClickHouse.

[alexey-milovidov]

cockroachdb, tidb, yugabytedb

Maybe none of them have CI testing with fuzzing and all four sanitizers?
Or they test only frontend part of the codebase but don't test RocksDB itself?

[alexey-milovidov]

All bugs has been fixed.

[toktarev]

Not sure if it is correct to write here, but curiosity and this is Open Source :).

I have a question to ClickHouse team regarding usage of RocksDB.

Guys, have you faced with RocksDB's memory problems ?

I mean unlimited memory growth during long period of insertion ?

I receive more and more complains during last 3 years in this issue: facebook/rocksdb#4112.

[alexey-milovidov]

@toktarev It's not actively used yet and we have no reports about this issue.

There are no tests for long period of insertion - maximum are fuzz and stress tests that last only one hour (for every commit).
And that tests are not specific - they run queries with RocksDB only occasionally (by random choice among other tests).
If memory won't grow for several tens of GB in short period of time - these tests will not notice.

There is still a chance that we will notice something if it exists...

[toktarev]

@alexey-milovidov , I see , thanks.

[alexey-milovidov]

#21479