Bug triage results: 2026-10-26 #7230

Summary

• Date: 2026-10-26
• requires-triage issues found: 48. Labeled: 40 (31 bugs, 9 enhancements). Skipped: 8. Failed to label: 0.
• Bugs by priority: priority:critical 11, priority:high 3, priority:medium 10, priority:low 7
• Guide: docs/source/contributor-guide/bug_triage.md

Labels have already been applied and requires-triage removed from every labeled issue. Please spot-check the calls and close this issue when satisfied. Several issues already carried a type label, a priority, or both from their author; those were kept when they matched the guide.

Bugs

priority:critical

• Native string to timestamp cast returns NULL for zone offsets with a non-zero seconds part (#7224)
  • Area labels: area:expressions
  • Rationale: A valid input silently becomes NULL where Spark returns a value, which is a silent wrong result.
• Native Iceberg scan returns NULL for an INT96 column read through a name mapping when other columns are selected (#7212)
  • Area labels: area:scan, area:Iceberg
  • Rationale: Nothing fails and the column comes back NULL, which is a silent wrong result.
• String-to-timestamp cast at the far-past range limit uses -08:00 instead of local mean time in America/Los_Angeles (#7210)
  • Area labels: area:expressions
  • Rationale: The cast returns a value 422 seconds off from Spark's, a silent wrong result, although only at the extreme end of the range.
• Native string-to-timestamp cast rejects zone suffixes Spark accepts (PST, CET, Zulu, Etc/GMT+8, +08:00:00, no space) (#7209)
  • Area labels: area:expressions
  • Rationale: Inputs Spark accepts return NULL with ANSI off, which is a silent wrong result.
• REBALANCE followed by GROUP BY or a window returns duplicate groups on skewed data (#7208)
  • Area labels: area:aggregation
  • Rationale: Skewed keys come back as several partial rows, a silent wrong aggregation result.
• collect_set over floats beside count(DISTINCT) diverges from Spark on signed zeros and NaN (#7203)
  • Area labels: area:aggregation
  • Rationale: The set contents differ from Spark with no error; this is a 1.2.0 regression for the distinct-aggregate shape.
• Native Iceberg scan drops rows from data files that store a narrower type than the column (#7138)
  • Area labels: area:scan, area:Iceberg
  • Rationale: Rows are silently dropped by the scan and the post-scan filter cannot restore them.
• Native Iceberg scan drops NULL rows for NOT (col <=> value) (#7137)
  • Area labels: area:scan, area:Iceberg
  • Rationale: NULL rows are silently dropped, which is a silent wrong result.
• Native Iceberg scan drops -0.0 rows for floating-point predicates (#7136)
  • Area labels: area:scan, area:Iceberg
  • Rationale: Rows are silently dropped because -0.0 and 0.0 compare as different values, which is a silent wrong result.
• Native Iceberg scan applies equality deletes to the wrong column when column names differ only in case (#7129)
  • Area labels: area:scan, area:Iceberg
  • Rationale: Deleted rows come back or the wrong rows are dropped with no error, a silent wrong result.
• Native object-store errors and debug logs print whole URLs outside the Azure store (#7128)
  • Area labels: area:scan
  • Rationale: A password in the URL userinfo or a token in the query can reach task errors and logs; the guide places security vulnerabilities at priority:critical. The reviewer may want to downgrade it, since it needs credentials embedded in the URL.

priority:high

• Native ABFS scans fail every task instead of falling back when Hadoop declines the auth type (#7202)
  • Area labels: area:scan
  • Rationale: Every task of affected ABFS jobs fails where 1.1.0 worked, which is major functional breakage on a production path.
• Bloom aggregate merge panics on batched states and ignores sliced offsets (#7172)
  • Area labels: area:aggregation
  • Rationale: The merge asserts and panics when spill replay supplies several states for one group.
• Executor is OOM-killed while Comet's pool and allocation counter stay under the limit: glibc keeps several GiB of freed native memory (#7108)
  • Area labels: none
  • Rationale: The cgroup OOM killer terminates the executor (exit 137) on a TPC-H SF10 run, which is a crash on a real workload.

priority:medium

• Native window over a sorted cache fails with "Expects PARTITION BY expression to be ordered" (#7223)
  • Area labels: none
  • Rationale: The query fails but only with spark.comet.convert.inMemoryCache.enabled=true and spark.comet.exec.inMemoryCache.enabled=false, so the user can turn the config off.
• CAST(timestamp_ntz AS TIMESTAMP) fails the task for values beyond year 262143 where Spark returns a value (#7211)
  • Area labels: area:expressions
  • Rationale: Loud failure (not a panic) on extreme values only, and the same bug is in 1.1.0.
• fair_unified: a sort fails on its first request when consumers that cannot spill hold the task's share (#7197)
  • Area labels: none
  • Rationale: A query fails with a memory error under a non-default pool, and switching the pool is a workaround.
• Broadcast exchange fails with OversizedAllocationException when a string or binary column holds more than 1 GiB (#7192)
  • Area labels: none
  • Rationale: A broken feature for very large build sides, with the workaround of avoiding the broadcast.
• Native string to negative-scale decimal cast doesn't match Spark 4.0.5, 4.1.4 and 4.2.1 (SPARK-60119) (#7187)
  • Area labels: area:expressions
  • Rationale: The divergence only applies to Spark releases that are still in RC and to spark.sql.legacy.allowNegativeScaleOfDecimal=true; see the escalation note.
• sequence and string to decimal cast don't match Spark 4.0.5, 4.1.4 and 4.2.1 (SPARK-58821, SPARK-60118) (#7170)
  • Area labels: area:expressions
  • Rationale: The divergence only applies to Spark releases that are still in RC; see the escalation note.
• Native Parquet scan: casts the planner inserts into pushed-down filters are rejected as Parquet type promotions (Spark 3.x, row-level filter pushdown) (#7158)
  • Area labels: area:scan
  • Rationale: Queries fail at runtime, but only with spark.comet.parquet.rowFilterPushdown.enabled=true on Spark 3.x.
• Native filters and non-collect limits can evaluate ANSI casts on rows Spark skips (#7147)
  • Area labels: area:expressions
  • Rationale: A query Spark completes raises CAST_INVALID_INPUT; it is a loud failure limited to ANSI mode and LIMIT or filter shapes.
• NativeUtil.exportBatch rejects large_string and large_binary vectors (#7109)
  • Area labels: area:ffi
  • Rationale: The export fails with Unsupported Arrow Vector for export for a native operator above CometMapInBatchExec.
• Arrow struct writer should respect projected schema width (#6576)
  • Area labels: area:Iceberg
  • Rationale: The writer can walk past the Arrow children, but only on the in-progress merge-on-read path.

priority:low

• Native NUMERIC_OUT_OF_SUPPORTED_RANGE error surfaces as SparkException instead of SparkArithmeticException (#7216)
  • Area labels: area:expressions
  • Rationale: The error class and message match Spark; only the exception type differs.
• Native Parquet scan reads timestamp columns as string, int, double or decimal, and INT96 as bigint, where Spark rejects them (#7214)
  • Area labels: area:scan
  • Rationale: Comet is more lenient than Spark on reads Spark rejects; valid queries are unaffected.
• Native Parquet scan read errors name no file, or every file in the task, instead of the file that failed (#7213)
  • Area labels: area:scan
  • Rationale: Cosmetic; the read fails as it should and only the error message is wrong.
• Native string to decimal cast raises the wrong error under ANSI for values with more than 38 integral digits (#7188)
  • Area labels: area:expressions
  • Rationale: Both engines raise an error and return NULL outside ANSI; only the error class differs.
• CometTaskMemoryManager.releaseMemory lowers used before Spark accepts the release (#7176)
  • Area labels: none
  • Rationale: Accounting is wrong only on the path where Spark throws during a release.
• A short grant stays with Spark when handing it back fails, but the pool records nothing (#7175)
  • Area labels: none
  • Rationale: Accounting is wrong only when the JNI hand-back call fails.
• Concurrent unified pool try_grow calls can both ask Spark for the same overcommit (#7174)
  • Area labels: none
  • Rationale: A race that makes a request larger than needed; no wrong results or crash.

Enhancements

• Reduce duplicate right-side keys before sorting a native left-semi join (#7189)
  • Area labels: none
  • Rationale: An opt-in performance optimization request.
• Run the MinIO-backed S3 suites in CI, starting with ParquetWriteWithMagicCommitterSuite (#7150)
  • Area labels: area:ci
  • Rationale: A request for new CI coverage.
• Run SortAggregateExec as a native hash aggregate without its input sort (#7146)
  • Area labels: area:aggregation
  • Rationale: A performance and robustness improvement request, not a defect in existing behavior.
• Tune nested take compaction threshold with a query-level benchmark (#7145)
  • Area labels: none
  • Rationale: A benchmarking and tuning request.
• test: move make_date SQL regressions into SQL fixtures (#7144)
  • Area labels: none
  • Rationale: Test reorganization; the issue says current coverage already exercises the behavior.
• perf: retain native make_date filters behind redundant NULL checks (#7143)
  • Area labels: area:expressions
  • Rationale: Results are correct; the request is to keep a native filter that currently falls back, and the issue says the impact is unmeasured.
• refactor: consolidate make_date and shared evaluation guards (#7142)
  • Area labels: area:expressions
  • Rationale: A refactor request.
• Keep wide-decimal scalar AVG native with a wider running sum (#7116)
  • Area labels: area:aggregation
  • Rationale: Asks to keep a currently conservative fallback native.
• Support join runtime reader filters through column-only projections (#7113)
  • Area labels: area:scan
  • Rationale: A request to extend runtime-filter attachment.

Escalations to consider

• Native string to negative-scale decimal cast doesn't match Spark 4.0.5, 4.1.4 and 4.2.1 (SPARK-60119) (#7187)
  • The native cast returns NULL where the new Spark releases return 0, a silent wrong result. It is priority:medium only because those releases are still in RC; consider priority:critical when they ship.
• sequence and string to decimal cast don't match Spark 4.0.5, 4.1.4 and 4.2.1 (SPARK-58821, SPARK-60118) (#7170)
  • '0e79' casts to NULL where the new releases return 0, a silent wrong result. It is priority:medium only because those releases are still in RC; consider priority:critical when they ship.
• Native window over a sorted cache fails with "Expects PARTITION BY expression to be ordered" (#7223)
  • Currently a loud failure behind two non-default configs. If the cache path becomes default, move to priority:high.

Skipped — needs more info

• [EPIC] Regressions in 1.2.0 since 1.1.0 (#7204)
  • This is a tracking epic, not a single defect, so no priority fits. It already carries bug and EPIC. requires-triage was left in place for a human to decide.
• Audit the PRs in 1.1.0 for regressions since 1.0.0 (#6399)
  • A multi-phase audit tracker; it is neither a bug nor an enhancement request.
• TPC-DS q54/q68: local SF100 results and distributed reproduction limits (#6946)
  • A reproduction report that explicitly claims no defect; it does not say whether there is a bug.
• Bug triage results: 2026-10-18 (#7105)
  • A previous triage summary that picked up the auto-applied requires-triage. Not a bug or enhancement; remove the label when closing.
• Bug triage results: 2026-10-12 (#6892)
  • Same as above.
• Bug triage results: 2026-10-05 (#6665)
  • Same as above.
• Bug triage results: 2026-09-28 (#6321)
  • Same as above.
• Bug triage results: 2026-08-24 (#5454)
  • Same as above.