Bug triage results: 2026-08-31

[andygrove]

Triage pass over the open requires-triage queue, per the project Bug Triage Guide.

• Date: 2026-08-31
• Total issues processed: 77 (73 triaged, 4 skipped, 0 failed)
• Type counts: 24 bugs, 49 enhancements
• Priority counts applied: priority:critical 5, priority:high 2, priority:medium 13, priority:low 4
• Guide: docs/source/contributor-guide/bug_triage.md

Labels have already been applied. A reviewer should spot-check the calls below and close this issue when satisfied; corrections should be made directly on the affected issue.

Note: where an issue already carried a priority label applied by its author, that label was left in place. Per the guide, this process adds a priority label to bugs only.

Bugs

priority:critical

• Checked TIMESTAMP_MILLIS overflow for nested fields and nested-predicate scans is blocked on DataFusion nested-field pruning (#5553)
  • Area labels: area:scan
  • Rationale: nested-field and nested-predicate scans keep overflow-to-NULL where Spark's Math.multiplyExact throws, which is a silent wrong result under decision-tree step 1, even though the fix is blocked on upstream nested-field pruning.
• Match Spark ObjectHashAggregate decimal AVG buffer semantics (#5509)
  • Area labels: area:aggregation
  • Rationale: grouped high-precision decimal AVG returns a different value from Spark with no error when Spark uses ObjectHashAggregateExec, a silent wrong result.
• Match Spark ordering and rank semantics for floating values nested in arrays and structs (#5507)
  • Area labels: area:expressions
  • Rationale: nested -0.0/NaN comparison keys give different ORDER BY and RANK() output from Spark with no error, a silent wrong result.
• Support Spark-compatible Unicode case-insensitive Parquet field matching (#5495)
  • Area labels: area:scan
  • Rationale: filed as a feature request, but the content is a correctness divergence — the native adapter's ASCII-only eq_ignore_ascii_case can miss a present physical column and return SQL NULL instead of the stored value under spark.sql.caseSensitive=false.
• Date-to-timestamp casts can overflow or panic for wide dates (#5456)
  • Area labels: area:expressions
  • Rationale: unchecked multiplication wraps to a bogus timestamp in release builds where Spark throws, which is a silent wrong result; the second reproducer additionally panics.

priority:high

• Native Azure store lets ambient AZURE_* environment variables override or corrupt explicit Hadoop auth config (#5542)
  • Area labels: area:scan
  • Rationale: ambient env credentials silently win over explicitly configured Hadoop auth, so Comet and Spark can resolve different identities for the same table; major functional breakage with a security dimension.
• Iceberg native scan claims schemes it cannot execute; three scheme lists disagree (#5541)
  • Area labels: area:scan
  • Rationale: a gcs/abfs/abfss/wasb/wasbs Iceberg table passes validation, is claimed, and then every task dies with CometNativeException, while stock Spark reads it fine — major functional breakage on supported storage.

priority:medium

• Fix remaining reported issues for arrays of map (#5544)
  • Area labels: area:expressions
  • Rationale: post-merge review of feat: support Map for CreateArray literal #5452 (merged) finds newly admitted complex literals and a deepNullable cast that changes Slice's nested map type while Slice still declares the original element type; a functional bug with a fallback workaround.
• Iceberg native scan fails queries on tables whose column names are case-distinct to Java but not to Rust (#5540)
  • Area labels: area:scan
  • Rationale: Rust's newer Unicode data folds columns Java keeps distinct, so the query dies with _LEGACY_ERROR_TEMP_2093 — visible breakage with a fallback workaround.
• Native Celeborn shuffle: validate_remote_schema rejects dictionary shapes the shuffle writer can emit (#5536)
  • Area labels: area:shuffle
  • Rationale: local and remote readers disagree about valid frames, so a frame that reads locally fails on Celeborn; a broken feature, and the path is not yet enabled end to end.
• NativeUtil.getNextBatch leaks Arrow structs when importVector fails (#5534)
  • Area labels: area:ffi, area:shuffle
  • Rationale: the third exit from getNextBatch is unguarded, leaking C data and wrapper buffers on every native operator path; resource exhaustion rather than wrong results or an immediate crash.
• unbase64 can fail on rows skipped by LIMIT and semi/anti joins (#5532)
  • Area labels: area:expressions
  • Rationale: batch-at-a-time evaluation turns a query Spark completes into a decode failure; a visible functional bug with an expression-level opt-out.
• Native Celeborn shuffle: the installed Celeborn bootstrap hook can break client creation for the whole executor (#5529)
  • Area labels: area:shuffle
  • Rationale: a Comet-specific bootstrap failure is fatal to all Celeborn client creation on the executor; scoped medium because the native Celeborn path is not enabled end to end yet (see escalations).
• Native Celeborn shuffle: reflectively replacing Celeborn's final fields can release push admission while payloads are in flight (#5528)
  • Area labels: area:shuffle
  • Rationale: reflective assignment to four private final Celeborn fields can release admission early; a broken feature on a path that is not yet enabled.
• Native Celeborn shuffle: default maxFrameBytes of 64 MiB is unreachable, and a large row fails the whole job (#5527)
  • Area labels: area:shuffle
  • Rationale: the two shipped defaults cannot both hold, and the consequence for a large row is a failed job rather than a slow one; configurable, so a workaround exists.
• Comet native broadcast fails under spark.kryo.registrationRequired=true (#5510)
  • Area labels: none
  • Rationale: Array[ChunkedByteBuffer] is unregistered so broadcast throws outright; broken feature, worked around by not setting that Kryo option.
• Avoid object-store cache and registry collisions across backends and configurations (#5502)
  • Area labels: area:scan
  • Rationale: the cache key omits the backend, so a second URL can be served the first URL's store and the DataFusion registry can replace a mapping; a functional bug reachable only with mixed fs.comet.libhdfs.schemes routing.
• Large-offset Arrow vectors from PyArrow UDFs cannot be serialized for broadcast or collect (#5488)
  • Area labels: area:ffi
  • Rationale: Utils.getFieldVector throws Unsupported Arrow Vector for serialize for representations Comet deliberately produces elsewhere; visible failure on a narrow path.
• AQE + DPP + spark.comet.exec.transitionRevert.enabled fails with "SubqueryAdaptiveBroadcastExec does not support the execute() code path" (#5486)
  • Area labels: none
  • Rationale: reversion leaves the plan in a state where PlanAdaptiveDynamicPruningFilters no longer matches, failing the query; gated on a non-default config, so a workaround exists.
• CometExecRule overwrites direct AQE LogicalQueryStage links during replanning (#5482)
  • Area labels: none
  • Rationale: unconditional restoration of originalPlan.logicalLink breaks the correspondence between the current physical root and the active logical stage; a planner defect with no reported wrong-result or crash path.

priority:low

• Surface Parquet TIMESTAMP_MILLIS overflow as a Spark-faithful exception instead of a raw Arrow error (#5517)
  • Area labels: area:scan
  • Rationale: the value is correctly rejected, only the exception type and message diverge from Spark's ArithmeticException("long overflow"); error-fidelity, no wrong results.
• Cancel background batch producers before collecting final plan metrics (#5504)
  • Area labels: area:ffi
  • Rationale: a drop-time metric guard can update counters after the final snapshot, so the reported metrics are incomplete; observability only.
• Explain ObjectHashAggregate fallback when Comet shuffle is disabled (#5500)
  • Area labels: area:aggregation
  • Rationale: the shuffle guard bypasses withFallbackReason, so the strict check can report an unannotated decline; a diagnostics gap, not an execution defect.
• Track provenance of fallback reasons copied through shared expression nodes (#5499)
  • Area labels: area:expressions
  • Rationale: a stale FALLBACK_REASONS tag on a shared node can satisfy reportUnexplainedFallback and hide an unexplained fallback; diagnostics only.

Enhancements

• Derive a native UDF's return type from the library instead of requiring the caller to declare it (#5597)
  • Area labels: area:udf (pre-existing)
  • Rationale: an API ergonomics improvement to the native UDF registration surface; nothing is broken today.
• Align Comet's native UDF surface with SPARK-55278's language-agnostic UDF protocol (#5596)
  • Area labels: area:udf (pre-existing)
  • Rationale: forward-looking alignment with an accepted SPIP targeting Spark 4.3/4.4.
• Rename the .claude directory to a vendor-neutral .ai (#5592)
  • Area labels: none
  • Rationale: repository layout change matching apache/datafusion and apache/datafusion-python; no functional impact.
• next_day and levenshtein fall back to Spark on collated strings (#5591)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; the current fallback is correct, the request is to route it through the codegen dispatcher instead.
• map_sort falls back to Spark for non-scalar map key types (#5590)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; widening dispatcher coverage, not a defect.
• map_from_arrays falls back to Spark under mapKeyDedupPolicy=LAST_WIN, unlike map_from_entries (#5589)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; requests a CodegenDispatchFallback mixin for parity with a sibling serde.
• timestamp_seconds falls back to Spark for decimal, byte and short input (#5588)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; new input-type coverage.
• abs on interval types falls back to Spark (#5587)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; new input-type coverage, already flagged good first issue.
• named_struct with duplicate field names falls back to Spark (#5586)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; documented fallback, request is to dispatch instead.
• translate falls back to Spark by default instead of using the codegen dispatcher like the other string functions (#5585)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; the Incompatible marking is correct, the ask is dispatcher coverage.
• length / bit_length / octet_length fall back to Spark on binary input (#5584)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; new input-type coverage.
• arrays_zip falls back to Spark for map element types (#5583)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; new element-type coverage.
• Array functions fall back to Spark for binary and struct element types (ArraysBase type gate) (#5582)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; widening a type gate across seven serdes.
• Hash functions fall back to Spark for decimal precision > 18, and sha2 for a non-literal numBits (#5581)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; a performance/coverage gap, the fallback itself is correct.
• Map lookups with float, collated or complex keys fall back to Spark (map_col[key], element_at) (#5580)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; the gate exists for real semantic reasons, the ask is to dispatch rather than fall back.
• lpad / rpad with a non-literal pad argument falls back to Spark (#5579)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; coverage for an ordinary query shape.
• to_csv never runs inside Comet by default, unlike to_json / from_csv / schema_of_csv (#5578)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; coverage plus a documentation update.
• unix_timestamp on string input falls back to Spark, while to_unix_timestamp already uses the codegen dispatcher (#5577)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; parity with an equivalent serde.
• round on float/double falls back to Spark, while bround already uses the codegen dispatcher (#5576)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; parity with an equivalent serde.
• Route unrecognized StaticInvoke and Invoke through the codegen dispatcher instead of falling back (#5575)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; turns a hard fallback default into in-pipeline execution.
• Codegen dispatcher is unreachable from convert, so serdes that decline there never get a dispatch attempt (#5574)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; a structural gap in dispatcher reach rather than incorrect output.
• Codegen dispatcher: guard the closure-serialize step so a non-serializable tree falls back cleanly instead of throwing at plan time (#5573)
  • Area labels: area:expressions
  • Rationale: part of [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work #5572; hardening an unexercised failure mode with no reported occurrence.
• [EPIC] Codegen-dispatch coverage audit: expressions that fall back to Spark where the JVM dispatcher would work (#5572)
  • Area labels: area:expressions
  • Rationale: umbrella for the coverage work above; a performance and coverage effort.
• Narrow the invalid-UTF-8 Comet opt-out in the sketch and hll SQL test files (#5571)
  • Area labels: area:scan
  • Rationale: the opt-out is correct but coarser than the problem; test-coverage improvement.
• SQLQueryTestSuite.ignoreList entries skip the Spark baseline, not just Comet (#5570)
  • Area labels: area:ci
  • Rationale: an intentional exclusion that is coarser than intended; narrowing it restores coverage rather than fixing a failure.
• Remove stale dev/diffs test exclusions whose tracking issues are closed (#5569)
  • Area labels: area:ci
  • Rationale: cleanup of exclusions whose cited issues are fixed; restores coverage.
• Native shuffle rejects nested types as hash partitioning keys although the native hasher supports them (#5566)
  • Area labels: area:shuffle
  • Rationale: the current gate is conservative but safe; the ask is to widen it now that nested hashing exists.
• Support useLargeVarTypes in accelerated mapInArrow/mapInPandas (#5555)
  • Area labels: none
  • Rationale: new configuration coverage for the accelerated Python UDF path.
• [FEATURE] Enable direct top-level Variant projection in ordinary native Parquet scans (#5551)
  • Area labels: area:scan
  • Rationale: new functionality, one of five atomic PRs under [FEATURE] Native top-level Variant projection from ordinary Parquet #5546.
• [FEATURE] Match Spark physical Parquet semantics for projected Variant columns (#5550)
  • Area labels: area:scan
  • Rationale: new functionality under [FEATURE] Native top-level Variant projection from ordinary Parquet #5546; the Variant path it constrains is not yet shipped.
• [FEATURE] Normalize marked Variant arrays at the native Parquet boundary (#5549)
  • Area labels: area:scan
  • Rationale: new functionality under [FEATURE] Native top-level Variant projection from ordinary Parquet #5546.
• [FEATURE] Carry Spark VariantType identity through Comet schema serialization (#5548)
  • Area labels: area:scan, area:ffi
  • Rationale: new protobuf and Arrow Field plumbing under [FEATURE] Native top-level Variant projection from ordinary Parquet #5546.
• [FEATURE] Export complete Arrow Fields through the native FFI boundary (#5547)
  • Area labels: area:ffi
  • Rationale: new FFI capability under [FEATURE] Native top-level Variant projection from ordinary Parquet #5546; today's exporter is limited, not broken.
• [FEATURE] Native top-level Variant projection from ordinary Parquet (#5546)
  • Area labels: area:scan
  • Rationale: parent tracking issue for re-landing feat: project Spark 4 VARIANT columns in native Parquet scans [Splitting] #5407 as five reviewable PRs.
• Native Celeborn shuffle: the remote read path costs up to 4.5x local decode and is unbenchmarked (#5535)
  • Area labels: area:shuffle
  • Rationale: validation is the right call for untrusted bytes; the ask is measurement and optimization.
• Native Celeborn shuffle: add a Celeborn test dependency so the reflection against its internals is verified (#5530)
  • Area labels: area:shuffle
  • Rationale: test-infrastructure improvement.
• Support NullType output types in codegen dispatch (#5525)
  • Area labels: area:expressions
  • Rationale: widening the dispatch type gate for common untyped literals; the current rejection is safe.
• Support Spark's pushed one-field VariantStruct (whole-value pushVariantIntoScan rewrite) in native Parquet scans (#5519)
  • Area labels: area:scan
  • Rationale: new scan capability; the current behavior is an explicit Spark fallback.
• Expose native Parquet scan I/O and read-amplification metrics (#5508)
  • Area labels: area:scan
  • Rationale: new observability counters.
• Narrow strict floating-point admission for corrected scalar sort keys (#5506)
  • Area labels: area:expressions
  • Rationale: admission is conservative but correct after fix: normalize scalar float sort and window rank keys #5469; the ask is to relax it for scalar keys.
• Include days in CalendarInterval hash when Spark 4.3+ is supported (SPARK-58236) (#5498)
  • Area labels: area:expressions
  • Rationale: Comet matches every currently supported Spark version; this is prospective work for a Spark 4.3 profile.
• ci: Prebuild Linux CI images with toolchains and dependencies (#5490)
  • Area labels: area:ci
  • Rationale: build-infrastructure improvement for reliability and setup time.
• Adopt techniques from Spark's ArrowCachedBatchSerializer (SPARK-57268) in Comet's cache format (#5487)
  • Area labels: none
  • Rationale: records design ideas worth adopting from an upstream implementation of the same idea.
• Cached reads that feed Spark operators are slower than Spark's own cache format (#5485)
  • Area labels: none
  • Rationale: a performance gap in a new opt-in feature, not a regression against previously shipped behavior.
• Comet cache decode ignores column projection, making narrow reads of wide cached relations slower than Spark (#5484)
  • Area labels: none
  • Rationale: a design limitation of the new cache serializer from feat: add experimental native support for in-memory cache, disabled by default #5051; optimization request.
• Track and remove native Variant compatibility workarounds after upstream fixes (#5477)
  • Area labels: area:scan
  • Rationale: maintenance bookkeeping so intentionally narrow workarounds do not become permanent.
• Remove Variant UTF-16 output rewriting (#5474)
  • Area labels: area:scan
  • Rationale: removal of a correctness-preserving workaround once upstream ordering lands; current behavior is correct.
• Make fair_unified account for spillable consumers (#5465)
  • Area labels: none
  • Rationale: memory-pool accounting improvement matching DataFusion's FairSpillPool; deliberately left out of the focused [EPIC] Memory pool and accounting audit sweep #5212 fix.
• Track exact SpillWriter page bytes for O(1) memory usage sampling (#5462)
  • Area labels: area:shuffle
  • Rationale: performance follow-up to [EPIC] Memory pool and accounting audit sweep #5212 finding [EPIC] Add support for all join operators #12; the correctness fix stays separately scoped.

Escalations to consider

• Native Azure store lets ambient AZURE_* environment variables override or corrupt explicit Hadoop auth config (#5542)
  • The guide lists security vulnerabilities under priority:critical. Ambient environment variables silently overriding an explicitly configured service principal (including the AKS workload-identity webhook case, where the configured client secret is dropped) is an identity-selection issue with a security dimension. Filed priority:high; a reviewer may want to escalate.
• Date-to-timestamp casts can overflow or panic for wide dates (#5456)
  • Matches the guide's escalation trigger in reverse: it both panics (NaiveDate + TimeDelta overflowed) and, in release builds, silently wraps to a wrong timestamp. Filed priority:critical on the wrong-result path.
• Native Celeborn shuffle: the installed Celeborn bootstrap hook can break client creation for the whole executor (#5529)
  • A Comet-specific bootstrap failure is fatal to all Celeborn shuffle on the executor, not just Comet's. Held at priority:medium because the native Celeborn path is not enabled end to end yet (see Native Celeborn shuffle: the remote read path costs up to 4.5x local decode and is unbenchmarked #5535); escalate to priority:high once it ships enabled.
• Checked TIMESTAMP_MILLIS overflow for nested fields and nested-predicate scans is blocked on DataFusion nested-field pruning (#5553)
  • Filed priority:critical on the silent-wrong-result rule, but the fix is blocked on upstream DataFusion nested-field pruning, so the priority reflects impact rather than actionability.

Skipped — needs more info

• Bug triage results: 2026-08-24 (#5454)
  • A prior triage summary, not a bug or an enhancement, so no type label applies. requires-triage left in place. Note that pass applied no labels (its token lacked label write access), so its 28 issues were re-triaged in this pass. A reviewer should close it.
• Bug triage results: 2026-08-17 (#5385)
  • A prior triage summary awaiting reviewer sign-off; no type label applies and requires-triage was left in place.
• Bug triage results: 2026-08-11 (#5325)
  • A prior triage summary awaiting reviewer sign-off; no type label applies and requires-triage was left in place.
• Bug triage results: 2026-08-03 (#5231)
  • A prior triage summary awaiting reviewer sign-off; no type label applies and requires-triage was left in place.