fix(expr): tolerate table function errors on streams

[wangrunji0408]

Signed-off-by: Runji Wang [email protected]I hereby agree to the terms of the RisingWave Labs, Inc. Contributor License Agreement.

What's changed and what's your intention?

partially fix_ #11915

Currently, when an error occurs in a table function call, the returned stream stops and the streaming actor will exit. This PR makes streaming tolerant to table function errors. When an error occurs, an error column will be appended to the returned chunk. (see the example in arrow-udf) The batch side will raise error immediately when seeing this column, while the streaming side will report errors and keep going. If some rows have been output when the error occurs, these rows will not be reverted.

Checklist

• I have written necessary rustdoc comments
• I have added necessary unit tests and integration tests
• I have added test labels as necessary. See details.
• I have added fuzzing tests or opened an issue to track them. (Optional, recommended for new SQL features Sqlsmith: Sql feature generation #7934).
• My PR contains breaking changes. (If it deprecates some features, please create a tracking issue to remove them in the future).
• All checks passed in ./risedev check (or alias, ./risedev c)
• My PR changes performance-critical code. (Please run macro/micro-benchmarks and show the results.)

• My PR contains critical fixes that are necessary to be merged into the latest release. (Please check out the details)

Documentation

• My PR needs documentation updates. (Please use the Release note section below to summarize the impact on users)

Release note

If this PR includes changes that directly affect users or other significant modifications relevant to the community, kindly draft a release note to provide a concise summary of these changes. Please prioritize highlighting the impact these changes will have on users.

[BugenZhao]

There was an discussion on the behavior in this case. #12474 (comment) cc @fuyufjh

Not sure which one we should adopt.

[wangrunji0408]

The default behavior for a table function not being called (e.g. with null input) is returning no rows. So I vote for this for the sake of consistency.

[fuyufjh]

The default behavior for a table function not being called (e.g. with null input) is returning no rows. So I vote for this for the sake of consistency.

I agree that these 2 cases should be consistent

[xiangjinwu]

This is Option 1 in the original issue. And I agree it is consistent with null input. In other words, the table function returns null (which is similar to {} but not {null}) on error.

Related edges cases on the differences among null, {} and {null}:

• array_agg function on zero input rows return no rows with group by, but a null row with scalar agg. fix(optimizer): decorrelate SimpleAgg with array_agg/jsonb_agg/jsonb_object_agg #15590
• Scalar subquery returning zero rows result in a null row. Array constructing subquery returning zero rows result in a {}. fix(planner): array-construction subquery shall return {} rather than null #15593

^ Just mentioning these cases. Not suggesting any value is always preferred than the other 2.

[BugenZhao]

LGTM!

[xxchan]

Thanks. Cool idea 👍

What I don't fully understand is where is the line leading to panic before? 🤣

When an error occurs, an error column will be appended to the returned chunk. (see the example in arrow-udf)

To clarify, arrow_udf::function and risingwave_expr::function don't have strict relationship, but share similar ideas. Do I understand it correctly?

[BugenZhao]

What I don't fully understand is where is the line leading to panic before? 🤣

I think there won't be any panic. We'll just enter a recovery loop where the expression error gets thrown out of the actor.

[wangrunji0408]

When an error occurs, an error column will be appended to the returned chunk. (see the example in arrow-udf)

To clarify, arrow_udf::function and risingwave_expr::function don't have strict relationship, but share similar ideas. Do I understand it correctly?

Yes. They share the same behavior behind the interface so that we don't need to write additional adapters for UDFs.