Repository navigation
test: iceberg_rest_scan tests the namespace it was asked for, and a client that ignored it no longer passes (#1341) - #1656
Merged
Conversation
jdatcmd
approved these changes
Oct 22, 2026
jdatcmd
left a comment
jdatcmd
left a comment
Collaborator
There was a problem hiding this comment.
Approved at 383a397. Your table reproduces on PG17, at this head and at its parent (#1655's head), with the client sending "db" whatever it was given:
| mutant | parent: shell / twin | this head: shell / twin |
|---|---|---|
nsscan (columnar_iceberg_rest.c:547) |
9/9 / 9 passed | 1 failed: "iceberg_rest_scan of a table in a namespace the catalog does not have is the catalog's 404: got [no error (1)]" / the same 1 |
nslist (line 844) |
9/9 / 9 passed | 1 failed: "iceberg_rest_tables of a namespace …: got [no error (events)]" / the same 1 |
The fixture's other users still pass at this head on PG17: iceberg_rest 14/14, iceberg_rest_vended 10/10, iceberg_rest_server 29/29. Clean iceberg_rest_scan: 11/11 shell and 11 passed twin on PG15 and PG19.
It merges after #1655, which still needs its pyarrow fix (my review there), and after the stack's restate onto main.
87 tasks done
OffgridwithJD
force-pushed
the
test/1341-rest-namespace-honoured
branch
2 times, most recently
from
October 22, 2026 05:57
5c71b77 to
d527919
Compare
… guard's (#1341) iceberg_dv.sh gains test/pytest/test_iceberg_dv.py, one check per test under the shell's names, and two holes in the shell are closed, each measured red-first on PG17: - "a manifest/footer offset mismatch is refused (XX001)" read the SQLSTATE alone, and several guards raise XX001. With the offset match removed (the !found refusal skipped), a later XX001 refused the same file and main's shell passed 26/26. Every refusal now reads its own guard's message: refusal() answers the SQLSTATE only when the message carries the guard's phrase, and "<SQLSTATE> (another refusal: ...)" otherwise. - The three "backend still up" checks read SELECT 1, which answers again after crash recovery. With the referenced_data_file check removed a NULL path reaches strcmp and the backend dies on that arm; main passed "backend still up after the DV refusals". They now also read the server log's count of killed backends (crashes/survived, as avro_manifest), and each refusal waits for the server to answer first (await_up), so a crash reddens its own arm and its section's survival check, the same names in both harnesses, and not a timing-dependent tail of collateral. The eight survivor comparisons move to check_text, so an empty oracle no longer matches an empty scan. Measured on PG17, shell then twin as shipped: seventeen mutations, each with the same reds in both (strictgt 7, anyfile/nodup/nov2/noflags/noref/nobigoff 2, the rest 1); 26/26 in each on 15-19. 25 rows dated, 1 never (the survival check after the DV applications, which no mutation crashed). Census 3884 -> 3885 never, 2286 -> 2311 dated, 6196 rows; suites_not_covered 102 -> 101; cluster_tests 2753 -> 2779; TESTS.md §268. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d slot no longer passes (#1341) The pytest twin of test/logical_decoding_source.sh (#754), one check per test, every check by the shell's name, on a private cluster at wal_level=logical. The shell's second slot (pgc_754b) had no heap control. Created after the 5000-row insert, it decoded nothing, and both "and it too emits no decodable change of its own" and "100x the rows does not give 100x the internal records" (0 < anything) passed: main 11/11. A heap table now writes through that slot in the same window and must decode its 5000 inserts ("CONTROL: the second slot decodes the heap table's 5000 inserts"), and the ratio check refuses a slot that carried no internal records. Under that harness mutation both are red in both twins (SAME 2). The pgc_skip message for a missing test_decoding.so was built from pg_config --pkglibdir, so the grader could not read it; it is now a literal, and the twin records the same name (with the plugin moved away on PG17: shell 1 FAIL, twin the same 1 FAIL and 12 declined). No source mutation reddens a check: the claim is that pgcolumnar writes nothing logical decoding can read. A default chunk group of 100 rows still keeps the internal records following groups (114 for 50 groups, under the 10x bound), so it dates nothing. 12/12 shell and 13/13 twin on each of 15-19. Census 3885 -> 3897 never, 2311 dated, 6208 rows (the skip name has no ledger row: it is not a check); suites_not_covered 101 -> 100; cluster_tests 2779 -> 2792; TESTS.md §269. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… longer passes as up (#1341) The twin of test/iceberg_deletes.sh, one test per check under the shell's name. Fixed in both harnesses, measured red-first against main's shell on PG17: - The six "backend still up" checks read SELECT 1, which answers again after crash recovery. With abort() in place of the null-pos refusal, or the partition spec's fields guard removed (a NULL deref, signal 11), main passed the survival check once recovery had finished. They now also read the server log's count of backends killed by a signal, and every reading waits (bounded) for the server, so a crash reddens its own arm and its section's survival check rather than the two arms after it. - Every refusal now also reads its own guard's message: XX001 and 22023 are each raised by several guards here. The oracle comparisons use check_text, so an empty side is not agreement. - "the data file has 5 rows before deletes" reads pgcolumnar's own read_parquet, so it is not a premise (#1339). Twenty-three mutations each redden the same names in shell and twin on PG17; head is 49/49 in both on each of 15-19. The equal-sequence equality check is dated only with both layered guards relaxed: the per-file strict < alone passed, because the pre-skip at the oldest data file's sequence number drops that delete first. Census 3897 -> 3903 never, 2311 -> 2354 dated, 6257 rows; suites_not_covered 100 -> 99; cluster_tests 2792 -> 2841; TESTS.md §270. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…at applied nothing no longer passes (#1341) The twin of test/logical_subscriber.sh, on two private clusters (a heap publisher at wal_level=logical, a columnar subscriber), one check per test, each by the shell's name. Four holes, each measured red-first on main's shell (PG17): - With the insert-only subscription disabled before its UPDATE, main passed 20/20: "skipped by design" and "no apply errors" read after a fixed sleep. An INSERT after the UPDATE must now arrive before either is read. - With no UPDATE sent to the wedge publication, "the error does NOT say only 'row locking'" and "the wedge is real" passed. The first now needs the replication error in the subscriber's log; the second reads "publisher -> subscriber". - With every lock statement failing on a missing table, the three "is not told about replication" checks passed. They now need the generic row-locking error first. - "premise: the second subscription synced before the UPDATE" is reddened by COPY failing in a logical worker, so it is not a premise (#1339) and is dated. The lock loop is unrolled into literal checks so the grader reads each name. Five mutations (hintnoname, nogate, overgate, applyins, syncins) redden the same names in both twins; clean 20/20 on 15-19 in both, and 8/8 head repeats on 17. 20 rows, 16 dated, 4 never. Census 3903 -> 3907 never, 2354 -> 2370 dated, 6277 rows; suites_not_covered 99 -> 98; cluster_tests 2841 -> 2861; TESTS.md §271. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ck that undid nothing no longer passes (#1341) The twin of test/logical_decoding_cdc_recipe.sh, every check by the shell's name, one check per test, on a private cluster at wal_level=logical. Like the shell, it extracts the CDC recipe from docs/user-guide.md at run time and executes it. Two holes, both measured on PG17 against main's shell, fixed in both harnesses: - The rollback arm read the capture count and the row only after ROLLBACK, so a rolled-back INSERT that never ran passed both arms. With it made a duplicate key, so the block stopped there, main passed 35/35. A reading inside the transaction, before the rollback, now requires the row both captured and present ("1:1"). Under the same mutation it is red in shell and twin. - "and it saw the same values the table holds" compared two readings with check. With the server stopped before it, "" = "" passed. It is check_text now. The missing-plugin pgc_skip name is a literal the twin records, and the seven recipe-contains premises are unrolled into literal names the grader can read. Five mutations date seven checks, the same names red in both twins on PG17: pgcolumnar_tuple_delete marking nothing (1), the update writing its new version with every column null (1), the update leaving the old version live (3), and two mutations of the guide's recipe: the trigger not firing on DELETE (2), and every non-DELETE captured as an INSERT (4). 36/36 shell and 37/37 twin on each of 15-19. Census 3907 -> 3936 never, 2370 -> 2377 dated, 6313 rows (the skip name has no ledger row: it is not a check); suites_not_covered 98 -> 97; cluster_tests 2861 -> 2898; TESTS.md §272. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…wn guard's (#1341) The twin of test/iceberg_rest.sh, every check by the shell's name, one check per test. It runs a private cluster restarted with the bearer token in the postmaster environment (a good one, a wrong one, none) against the shell's own REST fixture. Two holes in the shell, each measured red-first on PG17 and fixed in both harnesses: - The 42501 arms read the SQLSTATE alone, and two guards raise 42501. With the allow-list refusing everything and the link-local refusal gone (denyall_noll), "a link-local catalog host is refused even if allow-listed" passed on the allow-list's 42501; with every address read as link-local and the allow-list gone (llall_noallow), "a catalog endpoint off the allow-list is refused" passed on the link-local guard's. Every refusal now reads its own guard's message. - "backend still up after the REST refusals" read SELECT 1, which answers again after crash recovery. With the backend aborting on the catalog's 404 (crash404), main passed it. It now also reads the server log's count of killed backends. Mutants (PG17), SAME reds in shell and twin: tokenlog 1, tokenurl 2, no401 2, no404 1, nocfgobj 1, nocap 1, crash404 2, noauth 4, denyall_noll 8, llall_noallow 8, denyall 8. Clean 14/14 in both on each of 15-19. Seeded: 14 rows, 12 dated, 2 never (the premises). Census 3936 -> 3938 never, 2377 -> 2389 dated, 6327 rows; suites_not_covered 97 -> 96; cluster_tests 2898 -> 2912; TESTS.md §273. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…no longer passes as up (#1341) The twin of test/iceberg_rest_scan.sh, every check by the shell's name, one check per test. It runs a private cluster restarted with the bearer token in the postmaster environment against the shell's own REST fixture, which points loadTable at the committed warehouse relocated into a work directory. One hole in the shell, measured red-first on PG17 and fixed in both harnesses: "backend still up after the REST read" read SELECT 1, which answers again after crash recovery. With the backend aborting on the catalog's 404 (crash404) and the probe 3s later, main passed it. It now also reads the server log's count of killed backends. The not-found arm now reads the catalog's own 404 message, not any 42P01. Mutants (PG17), SAME reds in shell and twin: no404 1, nslast 1, crash404 2, noauth 7, denyall 7. Clean 9/9 in both on each of 15-19. Seeded: 9 rows, 8 dated, 1 never (the premise). Census 3938 -> 3939 never, 2389 -> 2397 dated, 6336 rows; suites_not_covered 96 -> 95; cluster_tests 2912 -> 2921; TESTS.md §274. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nores storage-credentials no longer passes (#1341) The pytest twin of test/iceberg_rest_vended.sh, one check per test under the shell's names, on a private cluster restarted with an explicit environment (every AWS_* removed; the catalog token, and for the ambient arm the AWS_* credentials, passed in), against the shell's own SigV4 S3 fixture and REST catalog. Three holes, each measured on main's shell on PG17 first: - "storage-credentials longest-prefix match selects the right creds" passed with the storage-credentials array never read (nosc, 10/10): the catalog's flat config carried the same right credentials. The fixture catalog now vends a wrong flat config beside the array, so only the array's longest entry reads. - "a wrong vended secret is refused by S3, not bypassed (28000)" read the SQLSTATE alone, and 28000 is also "no credentials": with vended credentials never built (novend) it passed. It now needs the 403 message; the 42501 allow-list arm likewise needs the allow-list's own message. - "backend still up after the vended-credential arms" read SELECT 1, which answers again after crash recovery: with abort() in place of the 403 refusal (abort403) it passed. It now also reads the server log's count of killed backends, and each arm waits for the server first. pg_restart_env also unsets every AWS_* before starting the postmaster, so the "no ambient credentials" premise does not depend on the caller's environment. Six mutations each redden the same names in shell and twin on PG17: nosc 1, shortest 1, novend 4, leak 1, noallow 1, abort403 2. Clean: 10/10 shell and twin on each of 15-19. Seeded: 10 rows, 6 dated, 4 never. Census 3939 -> 3943 never, 2397 -> 2403 dated, 6346 rows; suites_not_covered 95 -> 94; cluster_tests 2921 -> 2931; TESTS.md §275. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d no longer passes as up (#1341) The pytest twin of test/iceberg_rest_server.sh, one check per test under the shell's names, on a private cluster restarted without and then with the catalog token in its environment, against the shell's own REST fixture. Four holes, each measured on main's shell on PG17 first: - "backend still up" read SELECT 1, which answers again after crash recovery: with the half-credential guard removed (the mint runs with a NULL secret and the backend segfaults) or abort() in place of the catalog's 401 refusal, it passed. It now also reads the server log's count of killed backends, and every reading waits for the server first (the segfault had also reddened the next arm as collateral). - "a warehouse option on the SERVER is accepted (not HV00D)" and "oauth options are accepted on a USER MAPPING (no error)" wanted an empty SQLSTATE, which a stopped server also gives: with the server stopped before them, both passed. A statement that ran now reads "ok". - "a link-local oauth_token_uri is refused by the allow-list (42501)" read the SQLSTATE alone, and the link-local guard raises 42501 too: with the allow-list removed it passed. Every refusal now needs its own message. - "a half OAuth2 credential ... is refused before any request" now also counts the fixture's requests. Five mutations each redden the same names in shell and twin on PG17: nohalf 2, abort401 3, nowh 1, noallow 2, nomaptoken 3. Clean: 29/29 shell and twin on each of 15-19. The suite already had its 29 rows: 10 more are dated (12 dated, 17 never). Census 3943 -> 3933 never, 2403 -> 2413 dated, 6346 rows; suites_not_covered unchanged; cluster_tests 2931 -> 2960; TESTS.md §276. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…reads what the last key left (#1341) The twin compiles src/columnar_curve.c with a small driver into a shared library and checks the curve in Python: the permutation test (C1), the memcmp-order adjacency walk (C2), dyadic contiguity (C3), the production-width bridge (C4a), a Python transcription of Skilling's TransposetoAxes for the round trips (C4b), the goldens and the pack's extent (C5), ncols == 1 (C6) and the frozen Z-order bytes (C7), plus the serpentine, Z-order and truncating controls through the same checkers. The driver only encodes and counts. 184 tests, one per check, all in NO_CLUSTER: no connection and no cluster. Two shell changes: - The battery fills every key buffer with 0xa5 before the pack writes it. Under a memset one byte short, which goldens went red depended on what the previous call left in the buffer (121 red on main, a stable 124 now), and a buffer's first use read uninitialized memory. - arms_unrunnable/arms_failed are written out as loops at their call sites. They took a list of names in one argument, so compare_to_bash refused the whole suite; it now reads and grades it. The refusal arm in test_compare_to_bash.py keeps its coverage against a fixture holding the two helpers as they stood, and self-test 480's pinned refusal emitters lose arms_unrunnable (its check renamed from "the two" to "the ones the tree defines", ledger row with it). Measured on PG17: seventeen mutations of src/columnar_curve.c each give the SAME failing names in the shell and the twin, and between them redden every arm except the six premises and the ten INSTRUMENT controls. Clean 184/184 in both harnesses on each of 15-19. Seeded: 184 rows, 168 dated, 16 never. Census 3933 -> 3949 never, 2413 -> 2581 dated, 6530 rows; suites_not_covered 94 -> 93; guard_tests 474 -> 658; TESTS.md §277. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…was gone no longer passes (#1341) bench_guards.sh kept its own tally and emitted no RESULT records, so it could not be seeded or twinned. It now goes through lib.sh's check with no cluster (as wal_envelope.sh does), and test_bench_guards.py is its twin, in NO_CLUSTER. The per-runner checks grepped each bench runner for words. Every header mentions --enable-cassert, and the cannot-tell block's exit 1 sits within 24 lines of that mention, so main's suite passed all 84 checks with the assert refusal warning past it, with the empty-capture refusal removed, with every build refused, and with the override ignored, in every runner. Each runner's own guard block is now RUN against a stand-in pg_config (assert, plain, or no output), and a sixth check per runner wants a plain build let through. Only a complete guard is run: without one, nothing in the runner is executed. Also fixed, each measured red-first on main's shell: - "and the verdict is asked of cb_verdict" counted a comment that kept the call's shape; it now reads code only. - the ratio premise matched rr1= as r1=, so a rename left the printf check reading nothing; it now matches a whole name. - "more than enough is accepted" named two checks; the worker-slot one is now "more than enough worker slots are accepted". - two premises that a one-line change to bench/cb_guards.sh reddens lost the prefix: the cold tag's presence and the verdict helpers'. selftest 320's premise required at least one suite keeping its own PGC_CHECKS tally, and bench_guards.sh was the last; converting it reddened the premise for the reason the rule exists. It now counts the files the sweep read ("premise: the sweep read the corpus"; its never row renamed). Twenty-two mutations of bench/ redden the same names in shell and twin on PG17; 91/91 shell and twin on each of 15-19. Seeded: 91 rows, 49 dated, 42 never. Census 3949 -> 3991 never, 2581 -> 2630 dated, 6621 rows; suites_not_covered 93 -> 92; guard_tests 658 -> 749; TESTS.md §278. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… no longer passes (#1341) docs_style.sh kept a private tally and emitted no RESULT records, so no ledger row could be seeded for it and nothing could grade a twin. It now records through lib.sh's check() and pgc_summary, with no cluster, as wal_envelope.sh does. Four arms read text a reader never sees. Measured on main, each left all 58 checks green: - the 1024-floor sentence kept only inside an HTML comment; - "join key" kept only inside an HTML comment; - a stale VERSION citation with the right one hidden in a comment beside it; - the roadmap nav entry commented out with #. Each arm now reads the page without its HTML comments (_md), and the nav check ignores YAML comments. Each of the four now reddens its own check in both twins. Two loops built check names from a variable, which the twin grader cannot read. The per-version changelog checks are three fixed names that list the failing versions in their reading, and the META.json pair is two literal checks. In a checkout without release tags the changelog arms stay notes in the shell (the nightly coverage job runs it tagless and counts an unrunnable suite as failed), and decline UNMET_PRECONDITION by name in the twin. The guards job, where the twin runs (it is in NO_CLUSTER), now fetches the release tags so they run there. "premise: CHANGELOG.md still has the union merge driver" asserted the same fact as the driver check above it and went red with it, so it is an ordinary check. The twin, test_docs_style.py: 53 tests, one check each, by the shell's names, read with Python's re and pathlib. 36 documentation mutations (pages, META.json, the badge, .gitattributes, the post-tag allowance file, untracked install scripts) each redden the same names in both twins, with no premise among them. 53/53 shell on each of 15-19, 53/53 twin. 36 rows dated, 17 never (the 16 premises and the checker's examined-count control). Census 3991 -> 4008 never, 2630 -> 2666 dated, 6674 rows; suites_not_covered 92 -> 91; guard_tests 749 -> 802 (cluster_tests unchanged); TESTS.md §279. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d's (#1341) The twin of test/audit.sh, one check per test, each by the shell's name, on a private database. Three holes in the shell, each measured red-first against main's shell on PG17, and fixed in both harnesses: - The refusal arms accepted any error. With set_options failing at entry for a reason that is none of its guards (a division by zero), all five passed. expect_error now requires the guard's own message; under that mutation both harnesses red the same 9 checks. - The parallel CREATE INDEX's outcome was discarded (>/dev/null 2>&1 || true). With a build participant raising an error, no index existed, and "create index no relation leak" and "create index rows indexed once" both passed, the count answered by a Seq Scan. The build is now its own check, the count is read only through the index, and the plan check is no longer a premise (a product mutant reddens it, #1339). - The "AUDIT TEST ABORTED" notice never printed: a second `trap cleanup EXIT` replaced it. It now lives in cleanup(). The two plain set_options calls no longer end the suite under set -e, so the checks that read them fail by name. The suite now records through lib.sh's check and ends in pgc_summary, so it prints an accounting line for the first time: its local check() called pgc_record directly, which test_compare_to_bash's seed arm refuses for a twinned suite. A PASS line now carries the name only; a FAIL line still carries both sides. Eleven mutations (set_options guards, the parallel build's scan and reader, the collation guard, missing values, the write buffer after ADD COLUMN, a failing entry) each red the same names in shell and twin on PG17. Head: 34/34 on 17-19 and 31 records on 15-16 (the partitioned-parent arm is one skip there, and the twin declines its four tests UNSUPPORTED_MAJOR under that name), shell and twin alike. 35 rows: 25 dated, 10 never. Census 4008 -> 4018 never, 2666 -> 2691 dated, 6709 rows; suites_not_covered 91 -> 90; cluster_tests 2960 -> 2994; TESTS.md §280. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…othing no longer passes (#1341) fuzz, fuzz_avro, fuzz_listing, fuzz_arrow and fuzz_parquet each gain a twin that makes every check by the shell's name. Each hole below passed main's shell on PG17: - fuzz_avro / fuzz_listing: a generator that wrote nothing left the previous mutant (or, for the listing, no override, so the fixture served the real listing) and every check passed, 7/7. The mutant is now removed before it is generated, and "every iteration wrote a fresh mutant" counts it. fuzz_avro's reach check no longer counts a file the backend could not open. - fuzz_arrow / fuzz_parquet: a mutate that failed was skipped with `continue`, so a run that generated half its mutants passed, 4/4. "every mutant was generated and ran ..." now counts both. An empty or wholly rejected seed corpus was an unnamed FAIL; it is a named check, so the twin can mirror it. - fuzz: draws came from bash's $RANDOM (changed in 5.1) and check names carried each iteration's parameters, so names moved with the seed and no twin could draw the same run. The suite now draws from its own 31-bit LCG, which the twin transcribes (same tags, iteration for iteration), and asserts each property once over the run. Its set_options result was thrown away: with chunk_group_row_limit dropped (optdrop), main passed 250/250. Options now read back as set. Mutants on PG17, each reddening the same names in shell and twin: optdrop 1, gtmin 1 (fuzz); crashhdr 2, hangmagic 2 (fuzz_avro); crashtok 2 (fuzz_listing); crashmeta 1 (fuzz_arrow); crashpq 1 (fuzz_parquet). Clean on 15-19 in both. The sanitizer checks stay never on these non-sanitizer builds. Census 4018 -> 4051 never, 2691 -> 2700 dated, 6751 rows; suites_not_covered 90 -> 85; cluster_tests 2994 -> 3038; TESTS.md §281-§285. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lient that ignored it no longer passes (#1341) The REST fixture listed "events" for any namespace it was asked about, and every check in iceberg_rest_scan asked for "db". A client that ignored its namespace argument therefore passed everything: with loadTable always asking for "db" (nsscan) or the table list always asking for "db" (nslist), the shell and the twin were both 9/9 on main. The fixture now answers /v1/namespaces from its own --namespace, and 404s a table list for a namespace it does not serve. Two checks ask for one, in both harnesses: - "iceberg_rest_tables of a namespace the catalog does not have is the catalog's 404" - "iceberg_rest_scan of a table in a namespace the catalog does not have is the catalog's 404" Each requires 42P01 with the catalog's own 404 message. nslist reddens the first and nsscan the second, the same name in shell and twin on PG17. Head is 11/11 in both on each of 15-19, and iceberg_rest, iceberg_rest_vended and iceberg_rest_server (the fixture's other users) stay 14/14, 10/10 and 29/29 on PG17. Two rows added, both dated. Census 4051 -> 4051 never, 2700 -> 2702 dated, 6753 rows; cluster_tests 3038 -> 3040. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
OffgridwithJD
force-pushed
the
test/1341-rest-namespace-honoured
branch
from
October 22, 2026 15:29
d527919 to
af0f641
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
@jdatcmd A follow-up the iceberg_rest port turned up. It's stacked on #1655 (its base is #1655's branch), so it goes last.
The hole. The REST fixture listed
eventsfor any namespace it was asked about, and every check iniceberg_rest_scanasked fordb. A client that ignored its namespace argument therefore passed everything.nsscan: loadTable always asks fordb(columnar_iceberg_rest.c), beforenslist: the table list always asks fordb, beforensscan, afternslist, afterThe fix.
test/iceberg_rest_server.pyanswers/v1/namespacesfrom its own--namespaceand 404s a table list for a namespace it doesn't serve. Its other users (iceberg_rest,iceberg_rest_vended,iceberg_rest_server) stay 14/14, 10/10 and 29/29 on PG17.Counts: two rows added, both dated. Census 3992 → 3992 never, 2526 → 2528 dated, 6520 rows;
cluster_tests2810 → 2812.Gates: guards 795 passed, collection 2812, ledger gate rc=0, self-test PASSED.
🤖 Generated with Claude Code