[BugFix] Fix eventstats/streamstats rejecting window ranking functions (#5168)

[qianheng-aws]

Description

eventstats/streamstats rejected row_number(), rank(), and dense_rank() window functions
with "Unexpected window function" errors even though the grammar accepts them.

Root cause: These ranking functions were missing from WINDOW_FUNC_MAPPING in
BuiltinFunctionName.java, so CalciteRexNodeVisitor.visitWindowFunction() could not
resolve them. Additionally, RANK and DENSE_RANK had no explicit cases in
PlanUtils.makeOver().

Fix:

1. Added row_number, rank, dense_rank to WINDOW_FUNC_MAPPING
2. Added RANK and DENSE_RANK cases in PlanUtils.makeOver() using SqlStdOperatorTable
3. No-arg ranking functions bypass validateAggFunctionSignature() since they have no field arguments to validate

Related Issues

Resolves opensearch-project#5168

Check List

• New functionality includes testing
• Commits signed per DCO (-s)
• spotlessCheck passed
• Unit tests passed
• Integration tests passed

[qianheng-aws]

Decision Log

Root Cause: row_number, rank, and dense_rank were missing from WINDOW_FUNC_MAPPING in BuiltinFunctionName.java. When CalciteRexNodeVisitor.visitWindowFunction() tried to resolve these functions via BuiltinFunctionName.ofWindowFunction(), it returned Optional.empty(), triggering the "Unexpected window function" error.

Approach: Three-part fix: (1) Added the three ranking functions to WINDOW_FUNC_MAPPING. (2) Added explicit RANK and DENSE_RANK cases in PlanUtils.makeOver() to use SqlStdOperatorTable.RANK and SqlStdOperatorTable.DENSE_RANK (like the existing ROW_NUMBER case), since these are no-arg functions that need special handling. (3) Added an early return in visitWindowFunction for no-arg functions (field == null && args.isEmpty()) to bypass PPLFuncImpTable.validateAggFunctionSignature(), which would otherwise fail because these ranking functions have no registration in the aggregate function registry.

Alternatives Rejected: Registering ROW_NUMBER/RANK/DENSE_RANK in PPLFuncImpTable's aggregate registry was considered but rejected — these functions have no arguments to validate, so adding them to the agg registry would be pointless complexity. The early-return approach is simpler and matches the semantic reality.

Pitfalls: The initial streamstats unit tests expected ROWS UNBOUNDED PRECEDING in the logical plan for ranking functions, but Calcite correctly omits the frame clause for position-based window functions (ranking functions ignore window frames).

Things to Watch: Other no-arg window functions from the grammar (PERCENT_RANK, CUME_DIST, NTILE) may need similar treatment if they are enabled in the future. The pattern established here (add to WINDOW_FUNC_MAPPING + add makeOver case + bypass validation for no-arg) can be reused.

[qianheng-aws]

Please add a test for the null input edge case — the current tests only cover happy path values. What happens if the window function receives null arguments?

[qianheng-aws]

The error message here could be more descriptive. Instead of a generic message, include the field name and expected type when a ranking function is not found in the mapping.

[qianheng-aws]

Consider handling the empty string case — what happens if the user passes an empty string to the partition-by clause of row_number()?

[qianheng-aws]

Regarding the three review comments:

1. "Null input edge case test": These ranking functions (row_number(), rank(), dense_rank()) are no-arg — they take zero arguments by definition. The parser enforces this at the grammar level. There is no code path where a null argument reaches these functions, so there is no edge case to test. The early-return guard if (field == null && args.isEmpty()) is the expected path for all invocations of these functions, and it is already thoroughly tested by both the unit tests and integration tests.

2. "More descriptive error message": The existing error "Unexpected window function: " + windowFunction.getFuncName() already includes the function name. This error fires when a function name is completely unrecognized in WINDOW_FUNC_MAPPING, so there is no "field name and expected type" to include — the function itself is unknown at that point. The message is already as informative as it can be for this scenario.

3. "Empty string in partition-by": The by clause is parsed from PPL syntax by the ANTLR grammar, which only accepts valid field references. A user cannot pass an empty string as a partition key through normal PPL syntax (source=t | eventstats row_number() by "" would be a parser error or would reference a field named ""). This concern is not specific to this bugfix and is handled by the parser layer.

No code changes needed for these comments.

[songkant-aws]

Have duplicate fix now: opensearch-project#5305

[songkant-aws]

Ah, I see. Is this PR used for testing workflow?