feat(stream): normalize unmatched updates

[kwannoel]

I hereby agree to the terms of the RisingWave Labs, Inc. Contributor License Agreement.

What's changed and what's your intention?

Closes: #20342

The bug is that stream_key and storage_key can mismatch for an MV, and can result in UpdateDelete + UpdateInsert pairs being split up.

dev=> explain create materialized view m1 as select * from t order by v2, v3;
                                                      QUERY PLAN                                                       
-----------------------------------------------------------------------------------------------------------------------
 StreamMaterialize { columns: [v1, v2, v3, v4, v5], stream_key: [v1], pk_columns: [v2, v3, v1], pk_conflict: NoCheck }
 └─StreamTableScan { table: t, columns: [v1, v2, v3, v4, v5] }

Details:

1. If upstream updates some record, it will be updated on some stream_key .
2. This propagates to the backfilling MV.
3. When receiving barrier, we will flush buffered upstream chunks. Here we have logic to remove rows which have storage_key > current_pos, where current_pos is the storage_key we have backfilled until.
4. This means that an UpdateDelete + UpdateInsert pair could be split up, because while their stream keys are the same, their storage keys might be different, leading to one being pruned, but not the other. An example is when upstream is an MV with ORDER BY on non-pk columns.
5. Subsequently, UpdateDelete gets propagated to dispatcher, and the dispatcher has some optimization to remove redundant UpdateDelete and UpdateInsert pairs. We may lose an op entry here in this code branch. Then op.len() != col.len(), and we will panic when reconstructing the chunk.

This bug is sort of worked around with https://github.com/risingwavelabs/risingwave/pull/20176/files. Since dispatcher will rewrite U+/U- going to different vnodes into +/-, and dist_key is set to order key. But this PR defensively handles it in backfill as well.

For unmatched update delete, i.e. storage_key <= current_pos, we can rewrite it to a normal delete. Because in the next snapshot read, we should read the new update insert.
For unmatched update insert, we have not backfilled the corresponding update delete key from storage yet. So there won’t be a conflict to just insert it. We can rewrite it to a normal insert.

We also reconstruct the ops lazily, since in most cases it should not be necessary, only when there' a hanging update.

Checklist

• I have written necessary rustdoc comments.
• I have added necessary unit tests and integration tests.
• I have added test labels as necessary.
• I have added fuzzing tests or opened an issue to track them.
• My PR contains breaking changes.
• My PR changes performance-critical code, so I will run (micro) benchmarks and present the results.
• My PR contains critical fixes that are necessary to be merged into the latest release.

Documentation

• My PR needs documentation updates.

Release note

[BugenZhao]

1. If upstream updates some record, it will be updated on some stream_key .

This reminds me (again) of the discussion in (#12539. In my opinion, if maintaining the U- and U+ invariants proves more challenging than beneficial, should we consider only retaining - and + in the system (instead of patching or rewriting everywhere)?

[chenzl25]

The fix LGTM, thanks!

[kwannoel]

1. If upstream updates some record, it will be updated on some stream_key .

This reminds me (again) of the discussion in (#12539. In my opinion, if maintaining the U- and U+ invariants proves more challenging than beneficial, should we consider only retaining - and + in the system (instead of patching or rewriting everywhere)?

Generally agree. We may lose some traceability, but we will gain increased performance and code simplicity. We can now encode ops in a bitmap.

[kwannoel]

• feat(stream): normalize unmatched updates #20350 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.