Found two P2 regressions at head 7af1754: empty backing mappings abort otherwise readable alias scans, and mixed native esquery/source predicates become unevaluable local residuals. Both are described inline. No existing inline findings were present; existing_blocking_comment_ids is empty. No additional user focus was supplied.

Critical checkpoint conclusions:

1. Goal and correctness: source/index divergence is handled through local predicates and source reads, and alias targets are preserved, but the two inline failures leave valid queries broken. FE unit tests and the ES8 regression cover the principal intended cases; neither reported failure is covered.
2. Scope and reuse: changes are focused on four connector classes and tests, reuse the existing field contexts and scan-range contract, and keep generic planning connector-agnostic.
3. Concurrency and locks: no new threads or locks; field sets are built synchronously in per-scan metadata state before consumption. No new shared mutation or lock-order issue found.
4. Lifecycle/static initialization: existing per-statement raw-mapping reuse and per-scan provider state remain intact; reader/scroll cleanup is unchanged. No new static-initialization dependency or resource ownership issue found.
5. Configuration: no new settings; existing docvalue, keyword, parallel-scroll and scanner options are respected.
6. Compatibility: no new SPI, storage format or wire contract. The existing DSL overload remains and alias scans reuse the established shard_id=-1 sentinel. Empty backing mappings are a behavioral compatibility defect described inline.
7. Thrift/protobuf: no changes, so field-ID/enum migration checks are inapplicable.
8. Parallel paths: traced both legacy and V2 ES readers; both receive the original alias target and share the ES query/scroll implementation. Concrete indices retain shard-parallel scans.
9. Conditional checks: the alias mismatch condition correctly explains filters, routing and repeated shard IDs. Unsafe-field checks preserve ordinary Boolean residual indices; the native-function exception does not protect a containing OR/NOT, as reported inline.
10. Test coverage: FE tests cover unsafe attributes, predicate families, mixed ordinary Boolean filters, source filters, copy targets, runtime overrides, a second mapping, aliases and filter columns. The regression covers eight option combinations, pagination, residual LIMIT, NULLs, source filtering, alias filters/routing and counts. Add empty-backing and mixed-native Boolean cases.
11. Expected results: all 288 output blocks were checked against their fixtures and across eight configurations. Queries are ordered, except LIMIT cases with one matching row. Static inspection cannot establish output-generation provenance.
12. Observability: existing scan/debug and ES request logs expose target/DSL/mode; no additional metric requirement identified for this planning change.
13. Transactions/persistence: no EditLog, replay, commit, failover, visible-version or delete-bitmap changes.
14. Data writes/atomicity: production code changes only reads; transactionality and crash-safe writes are inapplicable.
15. FE-BE variables: no new transmitted variable; existing es_params/es_properties carry index, shard, DSL and docvalue mode through both reader paths.
16. Performance/memory: the additional schema pass is linear in backing-mapping size with no new remote fetch. Alias serialization and conservative pushdown/docvalue fallback are intentional correctness tradeoffs. No new per-row allocation or unbounded retry found; performance was not measured.
17. Other issues/final sweep: checked heterogeneous/missing fields, selected keyword subfields, native functions, empty/routed aliases, slot dependencies, residual indices and early LIMIT. Source-disabled/source-excluded value changes follow this PR's explicit stored-source policy and are not submitted as another defect.

Static review only: no builds, tests, benchmarks, services or source modifications were performed. End-to-end additions target ES8; typed ES5/6 execution and generation of the checked-in output were not independently run. These validation limits remain explicit; no unresolved review candidate or uncovered changed file remains. Review converged in two rounds after independent reviewers checked the unchanged findings and payload.

Review converged after 2 round(s).

Found four P2 issues: bounded keyword pushdown disagrees with source arrays exposed as STRING, and unrelated source filters, unused normalized subfields or searchable doc-value-only keywords reject otherwise exact native Boolean queries. All four are new instances, detailed inline.

The review reached its three-round cap without convergence: the final mapping sweep added the fourth verified finding after the other reviewers audited the previous three-comment set. The expanded final payload has not received another unchanged-set audit. All verified findings are included; no fourth round was started.

Checkpoint conclusions from static inspection:

• Goal and correctness: scalar source-value safety and alias preservation are implemented; the four inline cases remain. Added tests exercise the intended scalar behavior but omit these cases.
• Scope and clarity: changes stay within the ES connector and its tests, with no unrelated production changes.
• Concurrency and locks: the new field context is constructed and consumed per scan on the existing synchronous planning thread; no new threads, locks or lock-order changes.
• Lifecycle and initialization: per-scan memoization and reader cleanup remain intact; no new global initialization or ownership cycle.
• Configuration: no new configuration; keyword sniff, docvalues and parallel-scroll settings use existing paths.
• Compatibility: existing overloads and wire/storage formats are preserved. The planning regressions are described inline.
• Thrift/protobuf: no definition or field-ID changes.
• Parallel paths: traced both scanner generations through the shared ES query builder, reader and parser; aliases retain the original target in one scroll, including shared backing shard IDs.
• Special conditions: minimum bounds, source transformations, missing/empty backings and remote-only esquery guards were checked; the four overly broad exactness/safety decisions are the findings.
• Test coverage: FE boundary tests and eight SQL configurations cover both readers, docvalue modes and parallel modes, with independent internal-table and native-ES oracles. Add source-array, unaffected-source-field, unused-normalized-subfield and doc-value-only native-OR cases.
• Expected results: all 344 stored results have consistent case sets/results across the eight configurations and agree with the inspected scalar/alias fixtures. Generation provenance was not independently verified.
• Observability: existing query/range logging, read profiles and explicit planning diagnostics cover the changed paths; no additional concrete gap found.
• Transactions and persistence: no EditLog, transaction, failover or persistence changes.
• Writes and atomicity: read-only scan changes; write atomicity and crash recovery are unaffected.
• FE/BE propagation: source/docvalue selection and residual/LIMIT facts travel through existing properties for both readers; no new wire variable.
• Performance: aliases intentionally use one coordinating scroll; physical indices retain shard parallelism. Mapping safety is derived per scan and reused; no further concrete performance issue found.
• Other issues and limits: no further substantiated introduced defect after the full sweep. No additional user focus was supplied. This was a static-only review: no builds, tests, benchmarks or runtime validation were run.

Revalidated existing threads 4386047138 and 4386047135: their original native-residual and populated-plus-empty-backing instances are addressed at this head, so neither is reposted. There are no existing P0/P1 comments to retain. All seven changed files and assigned risk areas were covered; no unresolved review candidate or source-coverage gap remains.

Review did not converge after 3 rounds. This automated review cannot pass the review check.

Round 3 added verified M4: index-disabled keyword fields with docvalues are still searchable, but the new safety check rejects their supported native Boolean queries. Not all latest-round reviewers returned NO_NEW_VALUABLE_FINDINGS; the final four-comment payload has not received an unchanged-set audit after this addition. The three-round cap prevents a fourth round. All four verified findings are included, all seven changed files and assigned risks were covered, and no unresolved candidate remains. Validation was static only; runtime behavior and generated-output provenance were not independently executed or verified.

Reviewed exact head cff8803bf3a04e17eb7393d31f4808ebb5e2ccc2 using the supplied aggregate diff. One new P2 correctness finding is inline: forcing source reads for one projected column can leave an unchanged sibling's predicate incorrectly pruned. The PR's source-value preservation goal is therefore incomplete.

Existing context was independently checked and deduplicated. The original bounded-literal exemption is removed; affected native OR/NOT expressions fail during planning, separate AND conjuncts remain supported, selected-path safety and per-column source filtering address the reported scalar examples, and the keyword docvalue-only example is fixed. Two broader cases remain within existing threads and are not reposted:

• Empty backing mappings, r4386047135: a populated-first alias with another backing mappings:{"date_detection":false} still reaches the BooleanNode to ObjectNode cast in getRootSchema and aborts planning. The new tests cover {}, typed empty and _meta-only roots, but not flag-only roots permitted by ES mapping parameters.
• Docvalue-only searchability, r4386297855: isSearchable admits only keyword when index=false. A scalar long with docvalues enabled is still classified unsafe, so a=1 OR esquery(b,...) is rejected despite supported remote branches. ES also supports docvalue-only numeric, date, boolean and IP queries. No existing P0/P1 finding applies.

Critical checkpoint conclusions:

• Goal/data correctness: mapping transformations and original-target alias scans address the intended problems; the new sibling-representation finding and the two known cases above remain.
• Scope/clarity/reuse: changes stay within the ES connector and its tests; sharing mapping selection with schema inference reduces duplicate parsing logic.
• Concurrency/locks: no new threads or locks; per-scan state remains confined to synchronous planning, and statement scope shares the immutable raw mapping string.
• Lifecycle/static initialization: private parsed trees and scan contexts have bounded ownership; candidate probing preserves array metadata with a deep copy. No new lifecycle or initialization-order concern found.
• Configuration: no configuration items added; existing docvalue, keyword-sniff, scanner and parallel-scroll settings are exercised.
• Compatibility: no shared SPI, function symbol or persisted format change. Original-target alias scanning is intentional; native combinations needing local evaluation now fail at FE. Typed/typeless and empty-root handling was inspected, with the known flag-only gap above.
• Thrift/protobuf: no gensrc definitions changed.
• Parallel paths: both scanners share the query builder and parser and are affected by the new finding. Physical indices retain shard parallelism; aliases use one original-target range, preserving filters/routing and avoiding repeated shard IDs.
• Conditional checks: source transformations, selected query/read paths, missing backing paths, globs and remote-only Boolean checks have concrete purposes and controls; global read-mode effects remain incomplete.
• Test coverage: FE controls and the live-regression fixture span both readers, docvalue/parallel modes, Boolean/native cases, aliases, small pages and LIMIT. Missing cases are the sibling singleton-array interaction and the known extensions above. No tests or builds were run in this static-only review.
• Changed test results: all 488 stored sections were inspected, including canonical values and equality across eight configurations. Values are consistent with the fixture; test execution and output-generation provenance were not independently verified.
• Observability: existing scan/debug context and FE diagnostics support the changed decisions; no additional required metric was identified.
• Transactions/persistence/failover: no transaction, EditLog, replay or persistence path changed.
• Writes/atomicity/crash handling: production changes are read planning only; no data-write or atomicity mechanism changed.
• FE-to-BE variables: existing flat properties and docvalue contexts carry the changes; no new wire variable or missing serialization path found.
• Performance: every-backing field analysis adds local parsing work without new remote metadata calls; single alias scrolls deliberately trade parallelism for correct alias semantics. No separate substantiated performance defect found.
• Other issues/user focus: the full changed-file list, related control flow, tests and special paths were covered. No additional user focus was supplied.

Review converged after 2 rounds: all three agents reported NO_NEW_VALUABLE_FINDINGS in the final round, required static coverage is complete, and all candidates/risk items have verified or duplicate dispositions. No unresolved review question or uncovered assigned scope remains. Runtime/build/test validation was excluded by contract.

Review converged after 2 round(s).

Reviewed all 8 changed files at 74dc1718b9f495033354060fb3635190ea254034 with three reviewers and independent main-agent verification. One new P2 correctness finding is inline: source fallback leaves scalar scaled_float predicates classified as exact, so returned values can violate a pruned SQL predicate. The search/source difference is documented by Elasticsearch's numeric mapping reference.

Critical checkpoint conclusions:

1. Goal and correctness: original-target alias scans and source-safe string residuals address the intended failures; scalar numeric representation safety remains incomplete as described inline.
2. Scope and reuse: changes stay within the ES connector and its tests; sharing mapping selection removes duplicate inference logic, and existing SPI residual tracking/request/parser paths are reused.
3. Concurrency and locking: planning remains synchronous on a per-scan provider; no new threads, shared lock state, heavy locked operations or lock-order changes.
4. Lifecycle and initialization: mapping safety uses newly parsed JSON and per-scan context; shared statement mapping strings stay immutable. No new ownership/teardown or static initialization dependency.
5. Configuration: no configuration added. Existing docvalue, keyword-sniff, field-count, scanner and parallel-scroll settings are read through the existing paths.
6. Compatibility: no incompatible interface or storage-format changes, and no serialized-field definition changes. Existing DSL/mode/context properties feed both readers.
7. Thrift/Protobuf IDs: no gensrc changes; branch field-ID comparison is inapplicable.
8. Parallel paths: traced both reader versions and their shared query builder/parser. Alias ranges preserve the original target and omit shard preference; physical-index shard parallelism remains available. Cloud/shared-nothing use the same ES connector path.
9. Special conditions and failures: per-backing path/source checks and the source-mode gate have concrete semantic reasons. Native-only esquery is rejected during planning when an indivisible Boolean expression would require local evaluation; independent AND conjuncts remain supported. The string-only gate misses the numeric case inline.
10. Test coverage: added FE controls and end-to-end fixtures cover mapping transformations, selected query/read paths, empty backings, aliases, arrays, filter-only columns, native Boolean boundaries, small pages and LIMIT across eight scanner/docvalue/parallel combinations. Lossy numeric source fallback is missing.
11. Changed results: all 612 stored output blocks were checked statically against the fixture expectations, including independent SQL/native oracles; no snapshot mismatch found. Test execution and output-generation provenance were not verified in this review.
12. Observability: existing mode/DSL/range diagnostics and reader profiles remain available; planning failure explains unsupported native residuals. No additional critical metric or hot INFO logging requirement identified.
13. Transactions and persistence: production changes are external reads; EditLog, failover, visible-version and delete-bitmap rules are unaffected.
14. Writes and crash atomicity: no production data-write path changes; regression fixture writes are confined to the test setup. No new write atomicity or crash-lifetime issue identified.
15. FE/BE propagation: mode, DSL, contexts and original-target scan parameters use existing fields in both readers. Materialized filter slots are included before FE's mode choice; residual indices and source LIMIT gating remain aligned.
16. Performance: local source predicates and one scroll per alias intentionally trade pushdown/parallelism for correctness. Added analysis is per-planning mapping/expression work, with no new per-row operation. No benchmark was run.
17. Other issues: no further distinct verified finding after the final assigned-scope rechecks. The already-covered mapping variants below were deduplicated rather than posted again.

Existing-thread disposition: the seven original reproductions (empty backing, native residual, bounded keyword arrays, per-column source filtering, selected keyword path, docvalue-only keyword, and STRING sibling fallback) are addressed for their reported cases. Two extensions remain covered by prior P2 threads: docvalue-only numeric/date searchability is still rejected by the keyword-only isSearchable exception, although ES supports numeric docvalue-only queries; and configuration-only backing mappings, such as a root enabled=false without properties, can still reach the typed-mapping BooleanNode cast (supported ES mapping shape). These substantially overlap known comments and are not reposted. No existing P0/P1 finding needs to be carried.

User focus: no additional focus was supplied; the complete PR was reviewed. Coverage limits: static inspection only, including external API semantics checked against primary documentation. No builds, tests, Checkstyle, benchmarks or runtime queries were run, and repository source was not modified. No unresolved review candidate remains after the recorded acceptance/deduplication/dismissal decisions.

Review converged after 3 round(s).

FE runtime validation found a normalizer-triggered wrong-result regression. Reproduction steps and actual merge-base/head results are inline.

Static review completed against the exact PR head ba56fc4 in two bounded rounds. One new P2 is reported inline.

The changed scan-wide source fallback can make valid default Elasticsearch date values fail materialization. A date mapping without format accepts epoch-millis source strings such as "0" and fractional values such as "1.5". When a projected normalized keyword sibling loses its docvalue path, useDocValueScan switches the complete projection to _source; get_date_value_int only recognizes 13-character positive epoch strings, so these values fall through datetime casting (and a JSON fractional number reaches GetInt64() despite being a double). Both legacy and v2 readers share this failure path. The inline comment requests ES-compatible epoch parsing and regression coverage for both readers.

I revalidated all existing inline threads: none of the existing P0/P1 findings remains applicable, so existing_blocking_comment_ids is empty. Mapping/alias/REST routing, source-filter safety, Boolean residual planning, hidden slots, full-source propagation, LIMIT gating, parser coercion, arrays, NULL handling, and the user focus (no additional focus supplied) were reviewed. No builds, tests, source edits, or runtime validation were performed under the static-only contract.

Review converged after 2 round(s).

Elasticsearch filters indexed values while Doris may materialize different source values. This PR determines the scan-wide read mode before predicate pruning and preserves alias filtering/routing through the original target.

Static FE review completed at ba56fc4, including the necessary shared BE reader/parser boundary. Request changes: the existing epoch-millis date finding remains valid. A normalized sibling can force an otherwise readable date projection into source decoding and cause a format error. Fix the shared decoder and cover both readers; retaining a local predicate does not fix materialization.

Earlier mapping, predicate, numeric and array findings were checked against current code. The supplied current-head P0, compile and unit-test statuses are successful, but the AI review body still contains this unresolved P2. No builds or tests were run during this review. This conclusion covers FE and its affected execution contracts, not an independent full BE review or merge authorization.

[P2] Still requires changes: #69672 (comment)

Automated FE-focused static review (GPT-6.1 Sol, medium). No build or tests were run.

Elasticsearch filters indexed values, while Doris may materialize different source values. This PR selects the scan-wide read mode before predicate pruning, preserves alias filters/routing, and updates shared source decoding.

Independent static FE review completed at ad9d044 against merge base af9dad5. No remaining actionable P0/P1/P2 issue was found in FE or its affected BE execution contracts. Mapping safety, residual tracking, LIMIT gating, alias routing, both readers, and decoding boundaries were checked. The previous epoch-millis finding is addressed by the shared parser changes and added coverage. Approve.

No builds or tests were run. The supplied current-head snapshot reports successful P0, compile and external regression checks; unit tests and automated review remain pending. This is an FE-focused code opinion, not a full BE review or merge authorization.

Automated FE-focused static review (gpt-6.1-sol, high). No build or tests were run.

Reviewed #69672 at exact head ad9d044 against the supplied merge-base diff and all existing inline threads. The review converged after two bounded rounds.

No new actionable P0, P1, P2, or P3 finding was verified. The FE mapping/type, REST metadata, DSL/residual, scan-plan/property, alias-routing, and BE parser/wire scopes were covered. I traced mapping safety through scan-wide docvalue/source selection, Boolean and NOT/NULL pushdown, native esquery handling, read_full_source propagation, alias single-scroll routing, and both BE readers. Source/date/numeric/Boolean/ARRAY/JSONB boundaries and the changed unit/regression fixtures were inspected statically. A tentative epoch-millis plus-sign candidate was checked against the Elasticsearch formatter grammar and dismissed.

All existing inline threads and replies were revalidated at this head. They are prior P2 discussions with fixes or explicit scope decisions; no existing P0/P1 inline finding remains applicable, so existing_blocking_comment_ids is empty. The user focus file supplies no additional focus.

This was a static review only. I did not run builds, tests, benchmarks, live Elasticsearch reproductions, or runtime validation; author-reported results were treated as claims rather than execution evidence. No Thrift/Protobuf definitions, persistence, transaction, or write paths changed.

Review converged after 2 round(s).