[Analytics Engine] Skip partial/final split on non-prefix groupSet type mismatch (exact SqlTypeName)

[ahkcs]

Description

OpenSearchAggregateSplitRule.shouldSkipPartialFinalSplit() skips the PARTIAL/FINAL split for non-prefix group sets — those where a group-key ordinal k >= groupCount lands on a PARTIAL agg-output slot. The structural FINAL reuses the original ordinals over the keys-first PARTIAL output, so such a group key would take the agg-output column's type, which must equal the original input-field type or Volcano's Aggregate.typeMatchesInferred / row-type-equivalence check throws.

The collision check compared SqlTypeFamily, which is too coarse: BIGINT and INTEGER both belong to the NUMERIC family, so a span group key (INTEGER) paired with a measure declared BIGINT slipped through, the split was attempted, and planning crashed with either:

AssertionError: type mismatch: aggCall type: BIGINT inferred type: INTEGER

or

IllegalArgumentException: Type mismatch: rel rowtype ... $f2: BIGINT -> INTEGER

Fix

Compare the exact SqlTypeName instead of the type family, so these cases fall back to a SINGLE coordinator aggregate (always correct) rather than an unsound partial/final split. One-line semantics change; no behavior change for prefix group sets (handled by the earlier early-return) or for non-prefix sets whose collision-slot types already match.

How it surfaced / verification

Surfaced by PPL chart ... by <field> span=... over an INTEGER measure on the analytics-engine route (e.g. chart max(balance) by age span=10, chart usenull=... avg(balance) over gender by age span=10). The equivalent stats ... by <field> span=... shape hits the same path.

Verified against CalciteChartCommandIT (opensearch-project/sql) force-routed through the analytics-engine path. The two span tests that crashed on planning go red → green:

Test (CalciteChartCommandIT)	Before	After
testChartMaxBalanceByAgeSpan	❌ aggCall type: BIGINT inferred type: INTEGER	✅
testChartUseNullTrueWithNullStr	❌ Type mismatch … $f2: BIGINT -> INTEGER	✅

Combined with the partition-boundary coercion fix (#21878), chart on the analytics-engine route goes from 8/15 → 10/15; the remaining 5 are out of scope here (datetime wire-format sql#5420 ×2, and the attributes.client.ip Binary/BinaryView scan mismatch ×3).

Follow-up

A self-contained QA IT under sandbox/qa/analytics-engine-rest exercising stats max(<int>) by <field> span=N would regression-cover this in OpenSearch CI (no existing Span/Stats QA IT covers the stats max(int-measure) by … span non-prefix shape). Not included here.

Check List

• Commits are signed per the DCO using --signoff.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

🔒 No security concerns identified

✅ No TODO sections

🔀 No multiple PR themes

⚡ No major issues detected

[github-actions]

PR Code Suggestions ✨

Explore these optional code suggestions:

Category	Suggestion	Impact
Possible issue	Add negative index validation The bounds check for aggIdx should occur before using it to access aggCalls. Consider reordering the condition to check aggIdx >= aggCalls.size() first, or validate aggIdx >= 0 to prevent potential negative index access if k < groupCount logic fails. sandbox/plugins/analytics-engine/src/main/java/org/opensearch/analytics/planner/rules/OpenSearchAggregateSplitRule.java [74-77] int aggIdx = k - groupCount; -if (aggIdx >= aggCalls.size() || k >= inputFields.size()) { +if (aggIdx < 0 || aggIdx >= aggCalls.size() || k >= inputFields.size()) { return true; } Suggestion importance[1-10]: 2 __ Why: The suggestion to add aggIdx < 0 check is unnecessary because the loop iterates over aggregate.getGroupSet().toArray() and explicitly continues when k < groupCount (line 71-73), ensuring aggIdx = k - groupCount is always non-negative when the check is reached. The existing bounds checks are sufficient.	Low

[ahkcs]

Folded into #21878 — combining the partition-boundary coercion and the non-prefix agg-split type fix into a single PR, since together they're what makes PPL chart work end-to-end on the analytics-engine route.

[github-actions]

❌ Gradle check result for 05562f7: FAILURE

Please examine the workflow log, locate, and copy-paste the failure(s) below, then iterate to green. Is the failure a flaky test unrelated to your change?