[Gabriel39]

What problem does this PR solve?

Flight SQL preparation currently returns the same ResultMeta: string schema for every SQL statement without analyzing it. Python ADBC adbc_execute_schema() consumes that prepared dataset schema directly, so valid queries report incorrect columns and invalid queries appear to succeed. Direct GetSchema is also unimplemented.

Resolve output slots with the FE-only Nereids prepare planner and check table/column privileges before converting them to Arrow fields. Set query state before binding result columns, matching execution so unaliased expressions retain their SQL labels instead of synthetic command aliases. Both Prepare and GetSchema use this resolver, including nested types, nullability, decimal precision and timestamp units. Invalid queries fail before a prepared handle is registered; metadata requests do not construct a query executor or dispatch BE fragments.

Preserve JDBC preparation of commands using their FE result metadata (including the actual StatusResult protocol schema for commands without rows), and resolve USE/SWITCH prefixes within a temporary namespace scope. Session hints, statement resources and worker thread context are restored on success and failure. Schema analysis and execution share a per-session lock, so concurrent execution cannot update temporary analysis variables. Dialect conversion and the configured original-SQL parse fallback run before analysis, and USE/SWITCH prefixes honor remembered catalog databases without changing session history. Parameter binding, arbitrary multi-statement batches, and commands without discoverable result metadata remain explicitly unsupported by schema discovery.

Prepared handles are invalidated when used from a different catalog or database. After a successful prepared USE/SWITCH, that handle is rebound to its resulting namespace so it remains reusable; other handles retain their original namespace checks. Before GetSchema or prepared execution, reanalyze and compare against the schema advertised at Prepare; result-affecting session changes such as sql_mode require preparing again, while settings that preserve the schema keep the handle valid. Check the actual execution schema again before returning FlightInfo to cover concurrent DDL between analysis and replanning. For BE results, compare the final planner's semantic column labels and positional Arrow types: wire labels may contain expression text or a TVF prefix, and constant folding may narrow nullability. Keep nested field names, type metadata and dictionary encoding checks, and reject nullable widening. Retain the final executor alongside FlightInfo before processor cleanup clears the session context. On mismatch, expire the handle, cancel that retained executor before finalizing deferred executors, and clear buffered results/endpoints.

Resolve HELP headers from the local help catalog and SHOW SNAPSHOT headers from validated predicates without reading snapshot data. Resolve fixed headers for SHOW CONFIG, SHOW CREATE TABLE, procedure SHOW commands, ordinary EXPLAIN, PLAN REPLAYER and WARM UP SELECT. Concrete no-row DML and transaction commands advertise StatusResult; Lance CREATE/DROP INDEX advertise JobId without admission or remote table resolution. SHOW PROC exposes static headers for / and /current_queries without opening nodes or fetching rows. Other PROC nodes, EXPLAIN PLAN PROCESS (which lacks Flight result serialization), DESCRIBE, SHOW DATA, SHOW PARTITIONS, SHOW QUERY STATS, SHOW PYTHON PACKAGES, COPY INTO and WARM UP CLUSTER remain explicitly unsupported by schema discovery. The string-serialized group_concat_state output maps to Arrow UTF8; unmodeled aggregate-state serializations remain unsupported.

Prepare and prepared execution now incur an additional FE analysis pass to discover and validate the advertised schema. Physical planning and query execution remain in the existing execution path.

The External Regression investigation also exposed an existing remote Doris metadata timeout unit mismatch: the catalog supplies seconds, but pooled Thrift sockets require milliseconds. Convert the configured value with overflow checking so the default 10 seconds does not become 10 milliseconds.

Implement GetTableTypes data streaming for ADBC/JDBC metadata discovery. GetFlightInfo previously returned the standard schema and ticket, but DoGet always failed with UNIMPLEMENTED. Return sorted, distinct non-null public names from the existing TableType.toMysqlType mapping (BASE TABLE, SYSTEM VIEW, VIEW), using the standard non-null UTF8 field and releasing vectors on success or failure. This does not enumerate catalogs or execute queries; GetTables type-filter behavior is outside this change.

Release note

Fix Arrow Flight SQL schema discovery returning placeholder metadata and implement table-type metadata streaming for ADBC/JDBC clients.

Check List (For Author)

• Test
  • Unit Test
  • Regression Test
  • Manual test
• Behavior changed:
  • Yes. Prepare validates queries and reports their real output schema; GetSchema is implemented without executing the query.
• Does this need documentation?
  • No.

Validation:

• FE Checkstyle passed. Recompiled changed FE classes and tests against the existing FE test build; all 62 focused tests passed. Execution-schema tests now run real FlightSqlConnectProcessor.handleQuery(), covering the query-state transition that direct planner tests missed. The updated tests failed for SELECT 1, unhex and CAST before the fix and passed afterward. Negative schema-drift and deferred-executor lifecycle tests remain covered.
• Ran the actual regression framework against an isolated local FE + BE cluster using the BE binary from this PR's CI compile and current FE classes. With the previous FE classes, all five Flight suites from P0 build 1060727 reproduced schema-drift failures. With the fix, test_flight_timestamp_range, test_flight_utf8_validation, test_sql_cache_over_arrow_flight, test_select and test_auth_remote_ip all passed, along with the expanded test_schema_only and new test_table_types suites: 7 suites, 0 failures, 0 skipped.
• GetTableTypes tests cover the full GetFlightInfo-to-DoGet RPC, standard schema, sorted unique values, repeated requests, JDBC metadata, subsequent same-connection queries, and vector cleanup on stream failure. Python ADBC also passed 10 consecutive adbc_get_table_types calls followed by a query against the local cluster. Both new FE tests failed before implementation and passed afterward.
• The new persistent regression checks unaliased literal/function/CAST metadata and performs prepared execution, so metadata-only agreement cannot hide a mismatch with execution.
• The separately muted mtmv_with_sql_cache failure also occurs on unrelated PRs and is not changed here. Full P0/external CI reruns remain separate from the focused local validation.
• Earlier validation covered prepared namespace reuse, HELP and snapshot headers, session isolation, and the remote Doris timeout correction using a real local Thrift server with a delayed response.

Check List (For Reviewer who merge this PR)

• Confirm the release note
• Confirm test cases
• Confirm document
• Add branch pick label

[hello-stephen]

Thank you for your contribution to Apache Doris.
Don't know what should be done next? See How to process your PR.

Please clearly describe your PR:

1. What problem was fixed (it's best to include specific error reporting information). How it was fixed.
2. Which behaviors were modified. What was the previous behavior, what is it now, why was it modified, and what possible impacts might there be.
3. What features were added. Why was this function added?
4. Which code was refactored and why was this part of the code refactored?
5. Which functions were optimized and what is the difference before and after the optimization?

[Gabriel39]

run buildall

[Gabriel39]

/review

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: Review recovery stopped: No final review submission was declared
Workflow run: https://github.com/apache/doris/actions/runs/36544654197

Please inspect the workflow logs and rerun the review after the underlying issue is resolved.

[Gabriel39]

/review

[github-actions]

Complete static review of PR #68631 at d117502. Three review rounds covered all eight changed files; the final normal and risk-focused reviews returned NO_NEW_VALUABLE_FINDINGS. I recommend changes for the one P1 and eight P2 inline findings. There were no existing inline threads or previously reported P0/P1 findings to carry forward. No additional user focus was supplied.

Critical checkpoints

• Goal and proof: The PR replaces placeholder Flight SQL prepare metadata and implements direct/prepared GetSchema through FE analysis. The new FE and Groovy tests cover basic SELECT, nested types, SHOW VARIABLES, invalid SQL, privilege denial, SET/SET_VAR isolation, and allocator cleanup. The nine inline cases show the goal is incomplete for supported dialects, commands, session transitions, and aggregate-state output.
• Scope and clarity: The eight-file change is focused on schema discovery, but the new resolver duplicates execution's dialect and namespace behavior (M01, M02) and assumes command metadata is ready before command validation (M03, M04, M07, M08).
• Concurrency: Flight RPCs reuse a mutable ConnectContext. Schema analysis takes its monitor while ordinary execution does not; a concurrent successful SET can be discarded when analysis restores its saved session variable (M05). I found no separate new lock-order or deadlock issue.
• Lifecycle and errors: The resolver restores statement, thread, and session references in finally, and the moved allocator test checks repeated successful prepares. Preparation now rejects executable SHOW commands with empty static metadata (M08), and the embedded prepared schema can diverge after USE (M09). SHOW PARTITIONS and SHOW QUERY STATS fail before metadata setup; the wrapper reports those failures as invalid schema requests (M07).
• Configuration and compatibility: No new configuration or persistent format is added. Existing SQL dialect conversion is skipped by schema discovery (M01). The prepared result schema and GetSchema/FlightInfo paths must agree with existing Flight clients; M03, M04, M06, M08, and M09 show concrete disagreements. No new FE-to-BE session variable or wire field is added.
• Parallel paths and conditions: I compared direct GetSchema, prepared GetSchema, Prepare, GetFlightInfo, FE-local result sets, and BE Arrow schema construction. The command stmtType/empty-metadata fallback and the ShowTable-only validation special case are insufficient for the cited commands (M03, M04, M07, M08). Ordinary scalar, date/time, binary, and nested Arrow mappings match the inspected BE path; AGG_STATE serialization is the confirmed exception (M06).
• Tests and results: The new tests assert baseline positive and negative cases but do not exercise the reported command shapes, aggregate state, configured dialect conversion, remembered catalog database, concurrent SET, or post-prepare USE. No .out result file changed. This runner's review instructions prohibit builds and tests, so this is static source evidence only; author validation claims were not used as independent proof.
• Observability, persistence, and writes: No new EditLog, transaction, data-write, or crash-recovery path is introduced by schema lookup. The resolver returns Flight errors; M07's internal null-state failure is currently surfaced as INVALID_ARGUMENT. I found no separate log or metric defect supported by this diff.
• Performance and remaining issues: Prepare and GetSchema add an FE analysis pass; the PR description acknowledges that cost. I found no separate measurable hot-path regression. The final changed-file sweep left no unresolved candidate or risk item; all substantiated defects are in the inline set.

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: Review recovery stopped: No final review submission was declared
Workflow run: https://github.com/apache/doris/actions/runs/36548250539

Please inspect the workflow logs and rerun the review after the underlying issue is resolved.

[Gabriel39]

/review

[github-actions]

Review status: INCOMPLETE under the required three-round convergence cap. Seven distinct [P2] findings are attached. The final round identified the EXPLAIN/REPLAY preparation regression, so the all-NO_NEW_VALUABLE_FINDINGS condition was not reached. All identified candidates were independently checked and deduplicated before submission. Static review only; no build or test was run.

Critical checkpoints:

• Goal and proof: The change derives Flight SQL schemas without executing statements. New FE and Flight RPC tests cover ordinary queries, selected SHOW commands, namespace changes, Arrow types, and error paths, but they do not cover the seven cases reported here; this review did not run them.
• Scope: The 12 changed paths are focused on schema discovery and its producer/command support. The command metadata classifier still misses valid result and protocol OK shapes.
• Concurrency and lifecycle: Schema calls, normal execution, and prepared execution use the per-session monitor; the earlier P1 lost-SET race is fixed. Scoped analysis restores session, statement, executor, state, namespace, and thread context. SHOW PROC now fetches rows under that monitor (M3). No separate lock-order or resource-release defect was substantiated.
• Configuration: No new setting is introduced. Existing sql_mode changes can stale a prepared handle within one namespace (M1), and the existing dialect parse-retry setting is omitted from discovery (M2).
• Compatibility and parallel paths: No storage or wire-format change was found. Direct Flight execution and GetSchema/Prepare differ for Lance index ALTER, procedure SHOW, ordinary no-row commands, and EXPLAIN/REPLAY (M4-M7). The classifier must use actual command result shapes; WARM UP SELECT shows why a blanket INSERT rule is unsafe.
• Conditions and tests: Existing negative tests cover some unsupported dynamic SHOW and OTHER commands, but not the reported variants or EXPLAIN/REPLAY. The changed test assertions were inspected as source; their runtime results were not independently verified.
• Observability and performance: Existing error propagation exposes rejected schema calls; no separate logging or metric issue was substantiated. SHOW PROC's full fetch is the concrete avoidable CPU, memory, and possible BE RPC work (M3).
• Persistence, transactions, and writes: The diff adds no EditLog, persistence, or write implementation. It does change preparation compatibility for ordinary DML and BEGIN (M6); no transaction atomicity issue was found in the unchanged execution path.
• FE/BE variables and other correctness: No new FE-to-BE variable is transmitted. Arrow field mapping was compared with BE and FE-local result construction without another independently substantiated type mismatch. All changed files and nine existing inline threads were checked; no existing P0/P1 finding remains applicable. The user supplied no additional review focus.

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: Review recovery stopped: No final review submission was declared
Workflow run: https://github.com/apache/doris/actions/runs/36553557856

Please inspect the workflow logs and rerun the review after the underlying issue is resolved.

[github-actions]

COMPLETE static review of PR #68631 at head 37f71c9. Three convergence rounds finished; all final-round subagents returned NO_NEW_VALUABLE_FINDINGS. Four distinct P2 findings are inline. No P0/P1 remains on this head: the earlier P1 session-clone race (comment 4131745680) is addressed by the shared ConnectContext monitor around schema analysis and ordinary/prepared execution. Existing lower-priority threads were treated as duplicate fences.

Critical checkpoints:

• Goal and scope: GetSchema and Prepare now derive FE/BE-compatible headers without constructing StmtExecutor or dispatching query fragments. The analyzer uses Nereids output slots and privilege checks for SELECT, and explicit command metadata for FE results. The change is focused on Flight schema discovery and the command interfaces it needs, although the resolver must cover many command shapes. HELP and filtered SHOW SNAPSHOT remain incorrect (inline findings).
• Concurrency and lifecycle: Same-session schema analysis, Prepare registration, and GetFlightInfo execution share the session monitor. Analysis closes parsed statement contexts and restores the prior session variable, query state, executor, namespace, and thread context. No new lock-order cycle, static-initialization dependency, or unclosed Arrow allocation was established. Cross-session DDL can still change a prepared query schema between reanalysis and execution, and a prepared USE/SWITCH can invalidate its own handle (inline findings). Deferred-executor cleanup was checked on success/failure paths.
• Configuration and compatibility: No new config item, FE-to-BE variable, function symbol, storage format, or persisted protocol is added. Result-affecting session settings are reanalyzed before prepared use; namespace/command compatibility gaps are described inline. Direct statement, prepared statement, FE-local result, BE query result, cloud/non-cloud, and explicit unsupported-command paths were considered.
• Conditional paths and errors: Known execution-dependent SHOW metadata is declined; SHOW SNAPSHOT's WHERE-dependent header is still read before validation. Command errors are converted at the Flight boundary. Unsupported parameter binding and arbitrary multi-statement batches are explicitly reported. No separate silent-status or observability defect was substantiated.
• Tests and results: The new FE tests and Flight regression suite cover normal SELECT/SHOW/DML schemas, nested and Arrow types, privileges, dialect fallback, invalid SQL, session isolation, and namespace/schema drift. They do not cover the four inline cases, and Prepare/GetSchema agreement alone does not prove execution parity. No .out file or expected-result change was added. Builds and tests were prohibited for this review, so these conclusions are static and author/CI claims are not independent validation.
• Persistence, writes, performance, and remaining invariants: Schema discovery adds FE analysis but does not itself commit transactions, mutate table data, write EditLogs, or change BE visibility/MoW paths. An extra prepared reanalysis is intentional for schema drift detection; no separate substantiated CPU, memory, metric, or logging issue was found. Every changed file and accepted/unresolved candidate was swept before submission.

User focus: review_focus.txt supplied no additional focus points.

[hello-stephen]

FE Regression Coverage Report

Increment line coverage 63.57% (178/280) 🎉
Increment coverage report
Complete coverage report

[Gabriel39]

run buildall

[Gabriel39]

/review

[Gabriel39]

Fixed the External Regression failure in 61a1c5d. The failing test_remote_doris_unique_table_select query uses the non-Arrow catalog and times out reading remote table metadata: the configured 10-second timeout was passed to a Thrift API expecting milliseconds, so it became 10 ms. The same failure also occurred on an unrelated PR. Convert seconds to milliseconds with overflow checking; do not mask the failure with test retries.

The latest four review findings are fixed in a781cf9. Validation passed: 48 focused FE tests, FE Checkstyle, a real local delayed-response Thrift test, and the expanded Flight RPC regression. The full external suite still needs validation on the updated CI build. The dedicated schema-only Flight suite is not included in the External Regression selection.

[github-actions]

Static review complete for PR #68631 at 18bfadc. Request changes: one new P1 inline finding shows that prepared BE query execution discards its executor before validating the actual schema. The previously reported P2 for WARM UP SELECT with an empty source (r4134752120) still applies and is not repeated. The earlier P1 schema-analysis/SET race (r4131745680) is fixed on this head, so there are no confirmed existing P0/P1 IDs.

Critical checkpoints:

• Goal, scope, and tests: GetSchema and Prepare now derive schemas through scoped parsing, planning, and command metadata without running query fragments. FE unit tests and a Flight RPC regression cover ordinary queries, commands, types, invalid input, namespace changes, and session restoration. The changed remote-catalog Thrift timeout also matches its seconds-based caller and millisecond socket API. The tests do not exercise prepared BE execution through the real FlightSqlConnectProcessor.close() path, which masks the new P1; the empty-source WARM UP case is also missing.
• Concurrency and lifecycle: Schema analysis and Flight execution share the per-session monitor; the prior concurrent SET issue is addressed. Prepared handles reanalyze result-affecting settings and track namespace changes. The new post-execution comparison loses the executor during processor close, so valid BE field labels can expire the handle and mismatch cancellation cannot signal the BE. No separate lock-order, static-initialization, or ownership issue was substantiated.
• Parallel paths, conditions, and compatibility: I traced direct and prepared GetSchema/GetFlightInfo, FE-local and BE results, USE/SWITCH prefixes, no-row commands, and dynamic SHOW metadata against execution. The empty-source WARM UP fast path remains the already-commented exception. No new protocol symbol, storage format, or FE-to-BE session variable was added; Arrow wire-type and nested-field mappings were checked against the BE serializer. The P1 breaks prepared-result schema compatibility at runtime.
• Errors, resources, observability, and performance: Unsupported metadata paths return explicit errors rather than guessing headers; schema discovery avoids query execution and expensive PROC row fetches. Existing logs expose failures, but the P1's cancellation call is ineffective after executor cleanup. The removed prepared placeholder avoids its prior Arrow allocation. No transaction, persistence, storage-version, delete-bitmap, or data-write path is modified by schema discovery; no additional metric or hot-path regression was substantiated.
• Verification limits and focus: This was a static review only, as required; no build or test was run, so test results are not independently verified. There was no additional user-provided focus.

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Complete static review of PR #68631 at head f311672 (base 76d173e). I reviewed all 21 changed paths, the upstream and downstream execution paths, the existing inline threads, and the FE/BE Arrow schema contract. Two normal full-review subagents and a separate risk-focused subagent converged in one round with NO_NEW_VALUABLE_FINDINGS. There are no new inline findings.

The existing [P2] WARM UP empty-source thread (comment 4134752120) remains applicable: the early empty-insert return yields StatusResult while schema discovery advertises six statistics columns. I independently confirmed this path and did not repeat the comment. Historical [P1] comments 4131745680 and 4140024678 no longer apply: session execution now shares the schema-analysis monitor, and the final executor is retained before processor close for prepared schema comparison and cancellation. Accordingly, there are no still-applicable existing P0/P1 IDs.

Critical checkpoints:

• Goal and scope: the new schema-only path derives Flight headers without running the query; Prepare/GetSchema, prepared validation, command metadata, Arrow type conversion, and the remote Thrift timeout fix are localized to the affected components. The known WARM UP edge prevents complete execution/schema parity for that case.
• Concurrency and lifecycle: Flight session analysis and execution use the same context monitor; temporary session, statement, namespace and thread-local state is restored after analysis. Prepared execution captures its final executor before processor cleanup, compares the actual schema, then cancels and releases deferred resources on drift. The deferred executor list is detached under its own lock and finalized outside that lock. No new deadlock, state-loss, or cross-translation-unit initialization issue was substantiated. Planning and BE schema fetch can hold the per-session monitor, so concurrent RPCs for that session wait.
• Configuration and compatibility: no new configuration item, FE-BE transmitted variable, storage format, function symbol, or persistent record was introduced. FE Arrow type, nested child, nullability, timezone, and top-level logical metadata construction was checked against the BE Arrow serializer on this checkout. Rolling-upgrade behavior was not exercised.
• Parallel paths and conditions: direct and prepared schema calls, FE-local command results, BE query results, USE/SWITCH prefixes, no-row command categories, unsupported dynamic headers, and the empty-insert fast path were traced. The command-header exceptions avoid executing commands solely to discover metadata.
• Tests and results: the patch adds FE unit tests and a Flight RPC regression for positive and negative schema cases, session restoration, handle expiry, and drift cleanup. The new regression does not create schema_source/schema_target, so its DML and WARM UP cases prove metadata discovery rather than execution parity; the known empty-source case is not covered there. No builds or tests were run in this review, and no test outcome is claimed as independently verified. No result .out file was changed.
• Observability, persistence and writes: existing error propagation and logging cover schema-analysis and prepared-drift failures; no concrete missing metric was identified. Schema discovery does not start transaction or data writes and changes no EditLog/replay path. Direct command execution retains its existing write and transaction flow.
• Performance and other risks: prepared calls reanalyze SQL and session locking serializes same-session requests; this is visible in the implementation, but no separate correctness or material performance regression was substantiated. The FeServiceClient timeout now converts the caller's seconds to pooled-Thrift milliseconds for both affected RPC paths.

User focus: no additional focus was provided. Review execution is complete and static-only; the already-reported WARM UP P2 remains for the author to address.

[hello-stephen]

FE UT Coverage Report

Increment line coverage 79.20% (278/351) 🎉
Increment coverage report
Complete coverage report

[hello-stephen]

FE Regression Coverage Report

Increment line coverage 64.67% (227/351) 🎉
Increment coverage report
Complete coverage report

[Gabriel39]

run buildall

[Gabriel39]

/review

[Gabriel39]

Fixed the Flight failures from P0 build 1060727 in f44c66c.

Schema-only analysis created QueryState with isQuery=false, so result-sink binding inferred command-style names such as __literal_0 and __unhex_0. Real StmtExecutor execution sets isQuery=true and keeps SQL labels such as 1 and unhex(encoded). Preparation now sets query state before analysis, preserving the execution labels without weakening schema-drift checks.

Validated before committing against a real isolated local FE + BE cluster, using this PR's CI BE binary and the actual regression framework:

• Previous FE: all five affected Flight suites reproduced schema-drift failures; the new regression also failed its column-name assertion.
• Fixed FE, same BE: test_flight_timestamp_range, test_flight_utf8_validation, test_sql_cache_over_arrow_flight, test_select, test_auth_remote_ip, and test_schema_only all passed (6 suites, 0 failures, 0 skipped).
• All 60 focused FE tests and FE Checkstyle passed. The execution-schema unit tests now invoke the real FlightSqlConnectProcessor instead of calling the planner directly; three tests failed before the fix and passed afterward.

The separately muted mtmv_with_sql_cache failure also occurs on unrelated PRs and is outside this Flight fix. The local test cluster has been stopped. Full CI has been requested separately.

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: Review recovery stopped: No final review submission was declared
Workflow run: https://github.com/apache/doris/actions/runs/36683307092

Please inspect the workflow logs and rerun the review after the underlying issue is resolved.

[github-actions]

Complete static review of PR #68631 at d2c4a08 (all 23 changed files). One new P2 issue is inline: Prepare/GetSchema reject executable no-result SET and cloud USE prefixes before registering a prepared handle. The empty-source WARM UP result-schema mismatch still applies, but is already reported in r4134752120 and is not duplicated here. The earlier P1 session race (r4131745680) and lost-executor issue (r4140024678) are fixed on this head; no existing P0/P1 inline finding remains applicable. The review focus file specifies no additional points.

Critical checkpoints: The main goal—advertising analyzed Flight query/command schemas and table-type metadata—is implemented for the covered paths, subject to the new prefix regression and known WARM UP mismatch. FE schema analysis is isolated under the per-session monitor; normal and prepared execution use that monitor, the cloned session/query state is restored, and the retained final executor supports schema-drift cancellation and deferred cleanup. No independent lock-order, static-initialization, or lifecycle defect was substantiated. Command metadata was checked against FE-local result serialization, empty/conditional paths, and direct versus prepared execution; supported nested Arrow fields, scalar types, and table-type names were compared with the BE conversion and GetTables paths. The changed timeout uses existing seconds-based catalog configuration and the pooled Thrift socket's millisecond contract. No new configuration, FE-BE variable, persisted format, edit-log path, or transaction/write implementation was added, so those checkpoints are not directly exercised by this PR. Unsupported dynamic metadata fails explicitly; no separate observability or hot-path performance defect was substantiated. The changes are concentrated in FE schema/producer logic, command header helpers, the timeout fix, and supporting tests.

The added FE tests and Flight RPC regressions exercise representative query/command headers, types, session restoration, prepared lifecycle, table types, and negative cases. They do not cover the two executable multi-statement prefix examples or the already reported empty-source WARM UP path; metadata-only DML fixtures do not establish DML runtime behavior. This invocation ran no builds or tests, as required, so these are static code conclusions rather than independent runtime results. Three review rounds converged with no new valuable findings in the final full-review and risk passes, and the final changed-file/candidate sweep found no unresolved point.

[github-actions]

[P2] Preserve preparation for executable no-result prefixes. SET query_timeout=17; SELECT 1 and, in cloud mode, USE @valid_group; SELECT 1 both execute with a single final Flight result: their first command sends no result set. Prepare now invokes this analyzer before registering a handle, so this guard rejects both requests with UNIMPLEMENTED although they were preparable before. Cloud USE is a UseCloudClusterCommand, not a UseCommand; its database-qualified form can also change the final schema's namespace. Resolve these prefixes against isolated session state, including settings and cloud namespace/group changes, or retain a compatible prepared path.

[hello-stephen]

FE Regression Coverage Report

Increment line coverage 63.44% (236/372) 🎉
Increment coverage report
Complete coverage report

np