Review complete at head 07f16ef8ca737471c82cdc643ca69942325789b0 (three review rounds; all final reviewers found no new valuable issue). Four distinct findings are inline: two P2 test defects and two P3 cacheability cases. There were no existing inline comments or blocking findings to carry forward. This is a static review; builds and tests were prohibited and were not run.

Goal and proof: The new shared expression check excludes volatile scan predicates and virtual columns from condition cache and extends the query/SQL cache checks. The FE tests check the serialized scan flag and query-cache eligibility; the new regression compares result counts. Those tests do not prove the BE cache gate rejects volatile scans or still hits for deterministic scans (P2). The seeded-shuffle expected count is also dependent on a supported scan batch size (P2). Deterministic new-FE/new-BE scans are enabled by the translator and BE gate in the static trace, but the positive runtime path is not established by the regression.

Scope, reuse, and parallel paths: Moving the prior query-cache exclusion logic into ExpressionUtils.containsNonCacheableExpression is a focused reuse. The changed scan path checks attached filters and OLAP virtual columns, including lazy OLAP translation. File and TVF scanner cache predicates derive from the scan conjuncts; a filter retained above a scan does not enter its cache key. Query-cache aggregate/filter/project checks and the SQL-cache post-fold check were traced separately. The two typed-NULL P3 cases show the new exclusion is broader than needed. Seeded shuffle's volatile identity is appropriate for non-NULL arrays: the BE generator is reseeded per input block, and UniqueFunction preserves identity through rewrites. Non-NULL TIME-to-date casts read the query timestamp and must remain volatile.

Concurrency, lifecycle, and memory: The new per-scan flag is selected during FE planning, serialized into the plan, and read as BE operator state before opening the scan. No new shared mutable state, lock, thread entry, static initializer, ownership transfer, or significant allocation was added; the existing scanner and cache lifecycles are unchanged. No new error/status handling or silent-failure path was found.

Configuration and compatibility: No new configuration item or persistent format was added. The flag is an optional Thrift field with default false, so an old FE paired with a new BE safely disables condition cache. The documented Doris upgrade order upgrades all BEs before FEs. A new FE paired with an old BE would ignore this flag, but that is outside the documented rolling-upgrade sequence. Deterministic scans on a new FE/new BE can still enable the gate under the existing session switch. FE-to-BE field setting, copying, serialization, and BE consumption were checked.

Conditions, tests, and observability: The scan gate combines FE eligibility, the existing query digest, and the TopN restriction; the condition matches the cache's scan-local predicate inputs. FE unit cases cover volatile and deterministic flags, union branches, shuffle identity, TIME casts, and SQL-cache eligibility. The BE already exposes ConditionCacheHit, but the new regression never asserts it or lookup absence (P2). The hardcoded shuffle result is not valid for all supported batch sizes (P2). No runtime test result was independently verified. Existing counters are sufficient to investigate a hit, while a per-scan lookup assertion or isolated global search measurement is needed to prove volatile scans never search.

Data and operations: This PR changes no write, transaction, visible-version, delete-bitmap, EditLog, persistence, or crash-recovery path. No additional data-correctness, error-handling, resource, or performance issue was substantiated in the final sweep of all 14 changed files. The remaining performance issue is safe cache reuse being denied for the two always-NULL expressions (P3).

Review complete at 6228602 (static inspection only). I request changes for three distinct P1 correctness gaps: a new FE can still send an unsafe condition-cache digest to an older BE; a container cast with nested TIME-to-date conversion is classified as deterministic; and a query-clock cast in a hash-join runtime-filter probe can bypass the per-scan eligibility gate. All three have concrete cache-reuse paths in the inline comments. There are no unresolved candidates and no already-reported P0/P1 comments on this head.

Goal and scope: The new BE scan gate and shared FE expression check protect ordinary volatile scan filters on updated FE/BE pairs, and the change is focused on cache eligibility. The three cases above prevent the broader data-correctness goal from being fully met. I traced FE filter/virtual-column translation, Thrift serialization, BE scan digests and segment/file cache paths, SQL/query-cache folding, and the affected tests. No new data-write, transaction, persistence, visible-version, or MoW path is modified; no discarded Status, untracked large allocation, or cross-TU initialization change was found.

Concurrency and lifecycle: The new flag is per scan and read during BE scan open. Runtime filters may arrive before open and enter the digest/segment common-expression path; that lifecycle makes the third finding reachable. No new lock or lock-order issue was found. The existing condition-cache session switch is dynamically sent in query options; no new configuration item is added. Old FE to new BE defaults safely to disabled, while new FE to old BE has the rolling-upgrade gap in the first finding. The changed field is the only new FE-to-BE value; no storage-format or function-symbol change was found.

Focus and tests: The four earlier P2/P3 threads appear addressed on this head: known-NULL TIME casts and shuffle arguments reach NULL propagation after cast legality checks; non-NULL shuffle remains volatile; the fixed seeded-shuffle count is gone; and the OLAP regression now checks zero BE lookup deltas at batch sizes 1 and 1024 plus a positive second-run ConditionCacheHit. Its temporary global default is restored in finally, and the generated output counts are consistent with the fixture. The FE tests cover these scalar cases and ordinary scan eligibility, but do not cover the three inline cases. The author reports FE and regression validation; I did not run builds or tests because the review instructions prohibit them.

Parallel paths, performance, and observability: The zero digest propagates to OLAP and external file readers for directly flagged scans; the OLAP lookup metric and per-query profile give useful evidence for the added fixture. The metric does not cover external file lookups, so I treat that fixture as OLAP-only evidence. I found no separate hot-path performance regression or observability blocker. No documentation or release-note issue beyond the correctness gaps was substantiated.

Review complete for head d78bb71e44213a9341a1d4d50afa15d2e32db99b after two convergence rounds and a final sweep of all 16 changed files. I found one distinct P1 correctness issue, posted inline: the new eligibility policy permits deterministic virtual-column scan predicates even though their definitions are absent from the BE condition-cache key. The digest weakness predates this PR; this head explicitly retains eligibility for the affected scans. The earlier nested-cast and runtime-filter P1s are fixed on this head. No prior P0/P1 inline finding remains applicable under the supported upgrade sequence.

User focus: recursive ARRAY/MAP/STRUCT TIME-to-date casts are now classified volatile, while known-NULL casts still pass legality checking and fold. Singleton and grouped runtime-filter targets, including implicit probe casts, downgrade scan eligibility before serialization. The added regression checks zero volatile cache lookups and a deterministic second-scan hit; its expected outputs and test logic were inspected. The Doris upgrade guide requires BE upgrades before FE upgrades, and the optional flag defaults false for old-FE/new-BE operation. No additional issue was substantiated in those focused paths.

Critical checkpoints:

• Goal, design, tests: Shared expression screening and the per-scan FE-to-BE flag address volatile predicates, with FE unit and regression coverage. M1 leaves a wrong-row cache collision for virtual predicates. The changes are otherwise focused; the missing digest identity needs a two-query regression.
• Concurrency and lifecycle: FE sets the plan flag before serializing the scan; BE copies it into the operator and checks it at open. The shared condition cache is used across queries. This PR adds no locks, async owners, static initialization dependencies, or new lifecycle cleanup. M1 is cross-query key reuse, not a newly introduced lock-order problem.
• Configuration and compatibility: No new configuration item was added. Existing session switches control condition cache and virtual-slot CSE. The optional Thrift flag is sent on scan nodes and defaults false when absent; the documented BE-before-FE upgrade order makes the supported mixed-version phase safe.
• Parallel paths and conditions: OLAP, file, TVF, and lazy-scan setup, physical filters, singleton/grouped runtime filters, and virtual columns were traced. RF downgrades are monotone and cover implicit casts. The deterministic virtual-column check at the inline location lacks a cache-key identity check.
• Coverage and expected results: The added FE tests cover nested casts, NULL folding, shuffle volatility, scan flags, and RF targets. The regression asserts lookup counts and a positive cache hit, and its generated .out values match the stated fixtures by static inspection. No test covers M1. Author-reported builds and test runs were not independently repeated: this review environment prohibits builds and tests.
• Observability, persistence, and writes: Existing condition-cache lookup, hit, and filtered-row counters support diagnosis. No new transaction, EditLog, data-write, version-visibility, delete-bitmap, or persistent-format path is changed. No new error-handling or memory-ownership defect was substantiated.
• Performance and remaining issues: The per-expression eligibility checks are bounded by plan size; deterministic scans remain eligible. The final changed-file and candidate sweep found no other unresolved point beyond the inline P1.

Complete static review of head 9180e96509b5b7b4816cde183bcc667b5c18326d. Two full coverage passes and a separate risk-focused pass returned NO_NEW_VALUABLE_FINDINGS; the final sweep covered all 18 changed files and left no unresolved candidate. No new inline comments are proposed.

User focus and data correctness. VirtualSlotRef::get_digest() returns zero for both construction paths. VExpr, CAST, and runtime-filter digest parents preserve zero; scan-open computation stops on zero before predicate normalization. Scanner refresh returns zero when the local scan digest is zero and rebuilds from its current conjuncts otherwise. OLAP and file condition-cache lookup and insertion require a nonzero digest. This statically addresses the virtual-definition collision in existing P1 4164802395. The regression asserts virtual-slot plans, compares both definitions with uncached reference queries, requires zero lookups, and retains a deterministic second-query cache-hit check. The original pre-fix missing k=9999 result is author-reported; the new BE was not run here.

Critical checkpoints. Goal and tests: the code and added unit/regression cases cover the reported unsafe reuse, subject to the execution limit below. Scope and clarity: the zero sentinel reuses the existing digest failure convention, and the FE eligibility checks share one expression helper; no broader change was needed for this path. Concurrency: late runtime filters are added under the existing conjunct lock, scanners clone a snapshot, and the changed code adds no shared mutable cache state or lock inversion. Lifecycle/static initialization: per-reader cache context is cleared before refresh, and no new global initializer or retained owner appears. Configuration: no new setting is introduced; existing condition-cache and virtual-slot switches are used. Protocol/compatibility: FE serializes the scan flag through optional Thrift field 108, BE reads it before enabling the digest, and an absent old-FE field defaults false. The new-FE/old-BE scenario in P1 4144328662 is outside the documented BE-before-FE upgrade order.

Parallel paths and conditions: OLAP, file, TVF, scan filters, and singleton/grouped runtime-filter targets were checked; eligibility is only downgraded after scan creation, including the probe's implicit cast. Nested TIME-to-date casts, seeded/unseeded shuffle, and typed-NULL folding were traced through SQL, query, and condition-cache classification; the earlier P1s 4144328668 and 4144328681 no longer apply to this head. Test results and observability: the generated .out counts and test assertions are consistent with the fixtures, and existing lookup metrics/profile counters expose disabled cache use and positive deterministic hits. Transaction, persistence, and data writes: none are modified; the material risk was read-side omission of committed rows, addressed by disabling unsafe cache reuse. FE-to-BE variables and performance: the only new transmitted flag reaches scan-open, while ordinary deterministic scans retain the cache path. No additional substantiated correctness or performance issue was found.

Validation limit. This invocation performed static inspection only, as required by the review prompt. The author reports formatting, header hygiene, diff checks, clang-tidy, FE tests, and prior regressions, but those are not independently executed results here. The BE unit test and post-fix runtime behavior remain pending CI because the local dependency set lacks thirdparty/installed/lib64/libpaimon_c.a.

Review complete for PR #68672 at 5108dd2 against a53b182. The five retained commits and all 24 changed paths were reviewed. No new substantiated inline finding remains. This was static inspection only: no build, unit test, or regression suite was run in this review.

Critical checkpoints:

• Goal, size, and reuse: the per-scan FE eligibility flag gates BE condition-cache use; a shared expression classifier covers scan filters, virtual columns, runtime-filter probes, and query-cache checks. The change is focused on cache correctness, with NULL-only folding retained for shuffle and TIME casts. Inspected FE/BE unit tests and regression cases support the intended behavior, but their execution is not independently verified here.
• Concurrency and lifecycle: no new locks, threads, or cross-translation-unit static dependencies. The lookup counter is per reader and reaches the OLAP scan profile when its scanner closes. Cache digest selection occurs before scanner preparation; virtual-slot zero digests propagate through parent expressions and scan conjuncts.
• Conditions, errors, and parallel paths: volatile scan filters and RF probes, including implicit casts and singleton/grouped RF targets, downgrade the scan flag. Grouped RF translation still uses any non-blocking member to keep the merged filter non-blocking. Invalid or ignored RF targets do not assign a filter. The existing cast-legality and nonconstant shuffle-seed rejection paths remain in place.
• Compatibility and configuration: no new operator configuration or persistent format. The new optional FE-to-BE Thrift field defaults false, so an old FE disables the cache on a new BE during the documented BE-before-FE rolling upgrade. The reverse version order is outside that procedure. No transaction, EditLog, or data-write path changes.
• Performance and observability: stable scans remain eligible, while unsafe predicates avoid cache lookups. The ConditionCacheLookup counter counts attempted OLAP segment lookups and the regression selects each query by a unique profile marker; the deterministic control checks a warmed hit. The new SQL-cache plan traversal is limited to the SQL-cache rewrite path. No additional performance defect was substantiated.
• Tests and results: inspected both BE unit tests, all three FE test paths, the two regression suites, and the generated expected output. The regression uses query-scoped lookup assertions for unsafe OLAP scans and a positive deterministic hit. No test result from this review is claimed.

Focus response: the rebased ArrayShuffle retains UniqueFunction identity and rejects nonconstant seeds while NULL-only results fold; nested TIME-to-date casts are classified through ARRAY, MAP, and STRUCT types; virtual-slot digests disable condition-cache use; per-query ConditionCacheLookup assertions cover the OLAP fixture; singleton/grouped RF probes are checked without losing grouped non-blocking behavior; and the optional Thrift default protects the supported BE-before-FE upgrade. Existing P1 threads on virtual slots, RF probes, and nested casts are addressed at this head. The older new-FE/old-BE thread describes an unsupported upgrade order, so no existing P0/P1 inline comment remains applicable.