Reviewed pinned head fa30df3 across all nine changed files. One verified P1 requires changes: the Iceberg writer's new codec-error branch references undeclared open_status and has no final cleanup/status return. Static inspection establishes the build blocker; the inline comment covers this single incomplete refactor.

The intended codec flow is consistent: Hive/Iceberg table property -> existing FE GZ enum -> data/delete/MERGE writer -> shared Parquet GZIP mapping -> Arrow GZIP pages/footer -> existing footer-selected decoder. No additional user focus was supplied. No existing inline findings required carry-forward.

Critical checkpoint conclusions:

1. Goal and tests: GZIP support reaches Hive/Iceberg Parquet data and position-delete writers, but the P1 prevents the BE implementation from compiling. Positive FE property and Hive/Iceberg read/footer tests were added; none were run.
2. Scope: the shared mapping replaces three duplicate switches, with a focused FE whitelist change and regression additions. The error-handling branch is unfinished.
3. Concurrency: no new threads, shared mutable state, locks, heavy operations under locks, or lock-order changes. The helper is stateless.
4. Lifecycle and statics: open/write/close, sorted rollover, ownership and rollback paths were traced. No new global initializer or ownership cycle. The introduced invalid error path is covered inline; older dynamic-partition open-failure cleanup weaknesses are pre-existing.
5. Configuration: no new process/session configuration. Existing table codec properties are selected when binding each sink.
6. Compatibility: existing compression enums and standard Parquet GZIP pages/footer are reused; prior codec mappings and LZ4 framing remain consistent. Older binaries retain their prior unsupported-codec behavior. No storage-schema migration is introduced.
7. Thrift/protobuf: no definitions, IDs, enum values, required fields or wire variables changed; release-branch ID comparison is not applicable.
8. Parallel paths: Hive data, Iceberg regular/static/dynamic/sorted data, DELETE and MERGE data/deletes were checked. Existing legacy/native-v2 readers support footer GZIP, including non-x86 decoding. ORC/text and v3 Puffin deletion vectors retain separate paths.
9. Special conditions: the new Result failure branch is defective as reported. Other mappings preserve unsupported-codec rejection; the Hive test matrix correctly excludes ORC gzip and retains zlib there.
10. Test coverage: new positive coverage includes property storage, writes, ordered reads and actual footer codecs, including position deletes. No direct helper negative tests, multi-file stress or comprehensive merge/update cases were added.
11. Expected results: no output snapshots changed; added checks use explicit assertions and existing qt labels/results remain. Fixture warehouse and committed Iceberg metadata-file paths match the assertions by static inspection.
12. Observability: existing codec errors and Arrow writer status/logging paths are reused; no additional metric or distributed lifecycle was introduced. The invalid error return is the inline finding.
13. Transactions, persistence and failover: no EditLog, commit descriptor, snapshot/version or failover algorithm changes. Hive uses its normal table parameter; Iceberg preserves codec choice in the existing pinned write context.
14. Writes, atomicity and crashes: successful codec selection uses existing close/commit/abort paths, with no new commit point or visibility mechanism. Static inspection does not establish crash/runtime guarantees; the introduced error branch is reported.
15. FE/BE passing: the existing compression field reaches all relevant data/delete/MERGE descriptors and all three changed BE callers.
16. Performance and memory: a constant-size mapping replaces equivalent switches, without new significant owning allocations, copies, scans or repeated IO. Existing Arrow memory accounting is reused; GZIP CPU cost follows the requested codec.
17. Additional sweep: checked filename suffixes versus internal compression, pinned/default properties, Arrow zlib build capability, standalone unit-test setup, reader variants and available parallel rejection expectations. No other substantiated introduced defect was found.

Core data visibility, internal visible-version and MoW bitmap machinery are unchanged. The affected error-propagation invariant is covered by the P1.

Review converged after two rounds: both complete reviewers and the separate risk reviewer returned NO_NEW_VALUABLE_FINDINGS on the exact unchanged inline set, with all changed paths and risk/candidate dispositions complete. Remaining validation limits: static analysis only; no builds, tests, submodule initialization or Spark/Trino execution. The separate external compression product suite referenced by the old comment is unavailable in this checkout; no stale executable GZIP rejection was found in available sources. No unresolved static candidate remains.

Review converged after 2 round(s).

Reviewed all nine changed paths at 413ec38, including upstream property handling, downstream writer/commit behavior, and both regression suites. No new actionable defect found. Review converged in round 1: all three independent reviewers returned NO_NEW_VALUABLE_FINDINGS against the same empty finding set, and all risk candidates were dismissed with source evidence.

Existing P1 comment 4323545185 is fixed at this head: VIcebergPartitionWriter::open now returns the codec error explicitly; the undeclared open_status and fallthrough break are gone. No existing P0/P1 finding remains applicable, and no duplicate comment is posted.

Critical checkpoint conclusions:

1. Goal and proof: Hive DDL now accepts gzip, and all three Parquet sink writers map GZ to the existing Arrow GZIP codec. Added tests inspect metadata, written rows and actual data/delete footer codecs.
2. Scope and clarity: the focused helper replaces three duplicated switches while preserving the four existing mappings.
3. Concurrency and locks: the helper is stateless; no threads, shared mutable state, atomics, lock scope or lock ordering are introduced.
4. Lifecycle: rejected codec selection occurs before transformer creation or appends. HDFS writer ownership still closes its handle; S3 has no upload in flight at that point. No new ownership cycle or static initialization dependency was found.
5. Configuration: existing table properties are reused; no new dynamic configuration or restart behavior is introduced.
6. Compatibility: existing codec enums and Parquet encodings are reused, including Hadoop-framed LZ4. No wire or storage-schema change; GZIP writes still require a supporting BE during rolling upgrades.
7. Generated definitions: no Thrift/protobuf definitions or IDs change, so release-branch ID comparison is inapplicable.
8. Parallel paths: Hive data, Iceberg data and position deletes use the helper; sorted/partitioned writes and MERGE data/delete fan-out transport the same enum. ORC/text paths retain their existing handling.
9. Conditional/error paths: unsupported codecs return the original error; cleanup failure is logged. The Parquet-only gzip regression and ORC skip match their respective codec support.
10. Test coverage: new FE metadata and Hive/Iceberg write/read/footer assertions cover the feature, including position deletes. Existing codec and ORC/LZ4 checks remain. Direct helper negative tests and cleanup failure injection are absent; error propagation was checked statically.
11. Expected results: no golden output changes are required because new cases use direct assertions; existing qt labels remain unchanged. Ordered reads and expected GZIP footer values match the code path.
12. Observability: codec errors retain their codec context, and the new cleanup warning includes file path and status. No new distributed stage requires added metrics.
13. Transactions and persistence: no EditLog, replay, transaction protocol, visibility-version or delete-bitmap changes; existing external metadata/commit paths are reused.
14. Writes and failures: successful close/commit ordering is unchanged. Rejected-open errors propagate through async writer status and prevent commit-data creation; no new crash-cleanup defect was substantiated.
15. FE/BE transport: existing GZ reaches Hive and Iceberg INSERT/DELETE/MERGE sinks without a missing send path; no new transmitted field is needed.
16. Performance and memory: mapping remains constant time with no new large allocation, repeated scan or hot-path copy. GZIP compression cost follows the selected table codec.
17. Other concerns and user focus: no additional user focus was supplied. Rechecks of Result/header availability, defaults, property precedence, filename handling, test setup and unchanged codec behavior found no further issue.

Limits: static review only. No builds, tests, runtime failure injection or Spark/Trino interoperability checks were executed, and repository sources/submodules were untouched. No unresolved candidate or uncovered changed path remains within that scope.

Review converged after 1 round(s).

Reviewed all 10 changed files at d25ac89. No new actionable findings were identified. Review converged in 1 round: both complete-scope reviewers and the separate lifecycle-risk reviewer returned NO_NEW_VALUABLE_FINDINGS on the unchanged empty inline set; all candidates and risk focuses were independently verified and dismissed with evidence.

Existing review context: P1 comment 4323545185 is fixed: open() now propagates codec errors with DORIS_TRY and every format branch returns a Status; the undeclared open_status and erroneous break are gone. The creation-order request in thread 4332582986 is also implemented: all three Parquet writers validate the codec before creating the output file. No existing P0/P1 still applies, so existing_blocking_comment_ids is empty. No additional user-provided focus was supplied.

Critical checkpoint conclusions:

• Goal and evidence: Hive/Iceberg Parquet GZIP now follows the existing FE property -> GZ -> shared GZIP mapping -> Arrow GZIP route, including Iceberg position-delete files. New unit/regression tests cover property translation, invalid-codec file absence, row reads, and GZIP footers; these tests were inspected statically.
• Scope and reuse: one shared conversion replaces three duplicate switches; the FE allowlist and test changes are focused on GZIP and early rejection.
• Concurrency and locks: no new thread, shared mutable state, lock or lock order. Conversion is stateless; local create_file lambdas do not escape. Existing filesystem dispatch and sorted-writer synchronization remain in use.
• Lifecycle and static initialization: no new global initialization dependency or ownership cycle. Checked static/unpartitioned retained writers, dynamic writers, sorted/spill paths, cancellation, rollover, delete cleanup and MERGE sibling cleanup. Error paths guard absent transformers/file writers and avoid publishing commits; asynchronous filesystem dispatch completes within the caller lifetime.
• Configuration: no configuration/session variable is added; existing per-table codec properties and sink binding are reused.
• Compatibility: existing GZ/GZIP wire enums and standard Parquet codec are used; previous mappings and timestamp/version choices are preserved. Older BEs retain their prior explicit unsupported-codec error until upgraded.
• Thrift/protobuf: no definitions, IDs, enum values or required fields changed; release-branch schema comparison is not triggered.
• Parallel paths: Hive and Iceberg data/delete Parquet writers all use the helper; ORC and Hive CSV still create files before transformer setup with the same options. Generic Parquet already supports GZIP; Iceberg v3 deletion vectors retain their separate path.
• Special conditions and errors: unsupported codecs return non-OK Status before file creation. LZ4 retains LZ4_HADOOP and its existing page limits. The regression's ORC/GZIP exclusion documents the zlib spelling; no new unchecked Status or speculative null fallback was identified.
• Test coverage: FE metadata unit test, BE unsupported-codec test, Hive GZIP write/read/footer and Iceberg data/format-v2 position-delete write/read/footer checks were reviewed. Existing codec loops remain. The BE target's -fno-access-control permits the test's private-field access.
• Expected results: no golden files changed; new explicit row/codec assertions match the inputs and metadata schema. Ordered row assertions and existing result labels are preserved. No surviving GZIP expected-failure product test was found in this checkout.
• Observability: existing unsupported-codec Status, file-close/delete warnings and sink counters cover these unchanged operations; no new distributed phase calls for separate instrumentation.
• Transactions and persistence: no EditLog, visible-version, delete-bitmap or commit-protocol change; existing success-only commit metadata and rollback paths remain in use. No new failover-persisted state is introduced.
• Writes, atomicity and crash boundaries: successful creation/close/publication follows the existing transaction path; invalid codecs avoid creating output resources. Failure/cancellation propagation and cleanup were traced; no new write-boundary defect was identified.
• FE-BE variables: no variable is added. Existing compression fields were traced through Hive, Iceberg data/delete and MERGE routes.
• Performance and memory: constant-time shared conversion adds no row-path scan, large allocation or ownership container; invalid-codec file creation is removed. Existing memory pools and tracking context are unchanged.
• Further issues: rechecked Result/include visibility, test fixture conventions, filename versus internal Parquet compression, metadata URLs/columns, property defaults/overrides, and cleanup after failed open. No unresolved correctness candidate remains.

Validation limits: this was a static review only, as required; no builds, tests or submodule initialization were performed. Runtime cross-engine, mixed-version and crash/remote-store behavior were not exercised. The new BE test covers no-file rejection/destruction; retained-writer failure-close behavior was verified by source tracing rather than a direct integration test. No external product-test suite is present in this checkout. All assigned in-repository scope is covered; there are no remaining unresolved review questions.

Review converged after 1 round(s).

Reviewed all ten changed paths at f7456aa9b98e87244750713f9f3c8c40deb594f3. No new actionable findings. Hive and Iceberg data/delete/merge paths carry the existing GZ value into the shared helper, which selects Parquet GZIP and the existing Arrow GZIP builder. Existing codecs, including Hadoop-framed LZ4, remain intact.

Existing P1 comment 4323545185 is resolved: open_status is declared, assigned in both supported cases, and returned. The validation-before-file-creation request in thread 4332582986 is also implemented in all three Parquet writers. Neither is duplicated; no existing P0/P1 remains applicable. No additional user focus was supplied. The requested master-branch follow-up is outside this backport's changed scope.

Critical checkpoint conclusions:

1. Goal and tests: GZIP admission and routing are complete statically. Added tests cover Hive metadata translation, Hive INSERT/readback/footer, Iceberg data and position-delete readback/footer, and unsupported-codec rejection without file creation.
2. Scope: the shared mapping replaces three switches; changes and tests are focused on compression and open-time file handling.
3. Concurrency and locks: no new threads, shared mutable codec state, locks, or lock-order changes. The file-creation lambdas run synchronously within open().
4. Lifecycle and initialization: transformer/FileWriter ownership, static/nonpartitioned versus dynamic registration, and S3/HDFS/local close/destruction were traced. No introduced regression or static-initialization issue was established. The new best-effort delete is not a universal S3 orphan-prevention guarantee: later stream destruction can publish after it, but that same failed-open publication behavior predates this PR.
5. Configuration: no process configuration added; existing table properties select the codec.
6. Compatibility: existing GZ/GZIP values and standard Parquet page compression are reused; no changed wire layout, stored schema, or removed symbol.
7. Generated definitions: no thrift/protobuf changes; cross-release field-ID comparison is inapplicable.
8. Parallel paths: Hive new/existing/overwrite partitions, Iceberg data/delete/merge and sorted delegation were checked; ORC/CSV retain file creation and their existing transformers.
9. Conditions and errors: unsupported Parquet codecs fail before file creation; open failures propagate their Status. ORC/GZIP is correctly excluded from the regression loop.
10. Coverage: positive codec/footer assertions and a negative BE no-file test are present. Multi-file, external-engine and backend-fault-injection execution are not established.
11. Expected results: no output snapshots changed; new explicit assertions agree with source behavior, with deterministic row checks. The BE test target allows its direct private-member access through the existing -fno-access-control flag.
12. Observability: codec/open errors remain propagated and cleanup failures are logged; no additional metric-critical path identified.
13. Persistence and failover: no EditLog, replay or transaction-protocol changes; existing sink commit/rollback paths remain in use.
14. Writes and crashes: row conversion and visibility/commit semantics are unchanged; no introduced atomicity or crash-safety regression found within static review.
15. FE-BE transport: no new variable; existing compression fields were traced through data, delete and merge sinks.
16. Performance and memory: constant-time mapping adds no success-path allocation or per-row work; no new accounting or reservation requirement identified.
17. Further recheck: reviewed codec aliases/property precedence, pinned write context, file suffixes, position-delete triggering, metadata TVF prerequisites and parallel/error paths. No further actionable issue found.

Review converged in one round: both complete-scope reviewers and the separate lifecycle-risk reviewer returned NO_NEW_VALUABLE_FINDINGS; all changed paths, candidates and initial risks have evidence-backed dispositions. Static inspection only: no builds or tests were run, no source files changed, and no submodules initialized. Compilation, runtime behavior and cross-engine interoperability are not claimed as validated.

Review converged after 1 round(s).