[OffgridwithJD]

#1341 item 29: four fixes landed after the twin rule with shell arms only:

• A partitioned columnar table aborts the backend in four entry points: PgColumnarIsColumnarRelation tests relam and never relkind #1259: a partitioned parent with relam=pgcolumnar;
• import_parquet segfaults on a partitioned columnar table: the columnar check is absent, so #1261 does not reach it #1263: import_parquet into one;
• set_options refuses a columnar matview while reset_options accepts a plain view: the pair disagrees in opposite directions #1265: set/reset_options parity, and a matview's options row at DROP.

Each now has a partial pytest twin holding exactly its arms, by the shell's names, declared INCOMPLETE. Porting them found that two of the shell's refusal arms couldn't tell the bug from the fix.

Stacked on #1446, which is stacked on #1444. Merge them in that order; this branch is #1446's head plus two commits.

What changes

• Four partial twins, each declared INCOMPLETE in test_compare_to_bash.py with what it ports and what it doesn't:

  twin	arms ported
  test_inheritance.py	the A partitioned columnar table aborts the backend in four entry points: PgColumnarIsColumnarRelation tests relam and never relkind #1259 arms: 10 for the partitioned parent and 4 for the columnar matview
  test_parquet_import.py	the import_parquet segfaults on a partitioned columnar table: the columnar check is absent, so #1261 does not reach it #1263 arms: 7
  test_phase5.py	the set_options refuses a columnar matview while reset_options accepts a plain view: the pair disagrees in opposite directions #1265 parity arms: 11
  test_drop_cleanup.py	the set_options refuses a columnar matview while reset_options accepts a plain view: the pair disagrees in opposite directions #1265 drop-hook arms: 8

  • The two partitioned-parent tests decline on PG15 and 16 (UNSUPPORTED_MAJOR), where PARTITION BY … USING is refused and the fixture can't exist. Each has a row per major in expected_unrunnable.txt; measured, pytest exits 0 on 15 and 16 with 1 unrun each.

• inheritance's A partitioned columnar table aborts the backend in four entry points: PgColumnarIsColumnarRelation tests relam and never relkind #1259 refusals counted any error.

  • Measured: with the A partitioned columnar table aborts the backend in four entry points: PgColumnarIsColumnarRelation tests relam and never relkind #1259 predicate reverted, a release build doesn't abort. Each call opens relfilenode 0 and fails with 58P01: could not open file ".../0/0", which is the bug itself, and inheritance.sh stayed 26/26. CI runs release builds.
  • The fix: both refusals now require 42809, and their liveness comes from the server log's "terminated by signal" count, not from a new connection.

• parquet_import's import_parquet segfaults on a partitioned columnar table: the columnar check is absent, so #1261 does not reach it #1263 refusal passed on the SIGSEGV itself, since a dropped connection is an error too. Its liveness check caught the crash only by reaching the server before crash recovery finished, which I measured as timing-dependent on corruption.sh in test: corruption gets its pytest twin, and a backend crash can no longer pass for a refusal (#1341 item 25) #1446.

  • The fix: it now requires 42809 and counts crashes in the server log.

• phase5 moves onto lib.sh's check, pgc_fail and pgc_summary. A twin requires that, because the grader can't read a suite whose local check() calls pgc_record. It leaves suites_without_accounting.txt. Its 47 records are unchanged in names and verdicts, compared record by record against a run before the change.

• Ledger: inheritance's two refusal arms leave never, re-dated by the reverted predicate. The census is 3612 rows, 2673 − 2 = 2671 never and 939 + 2 = 941 dated, matching the tsv. The gate gives rc=0. parquet_import, phase5 and drop_cleanup aren't ledger-covered suites, so they owe no rows, as with test: iceberg_data_files refuses a delete entry inside a data manifest (#1341 item 49) #1439.

Red first

Each fix was reverted one at a time on PG18, and run through both twins, with the .so md5 printed every time. The SQL mutants edit pgcolumnar--1.0-alpha5.sql, which each suite's CREATE EXTENSION reads.

revert	shell	pytest	what went red
#1259: IsColumnarRelation from relam alone (columnar_metadata.c:4231)	inheritance 24/26 (was 26/26 before this PR)	test_inheritance.py fails	both refusals: got [58P01] want [42809]
#1263: the storage guard removed (columnar_parquet_reader.c:3674)	parquet_import 23/25 (the refusal arm passed on the crash before this PR)	fails	the refusal, no SQLSTATE (server closed the connection unexpectedly), and the liveness, 1 backend crashes
#1265: reset_options' guard removed	phase5 44/47	fails	heap and partitioned-parent parity, and the partitioned-parent verdict
#1265: set_options refuses a matview again	phase5 45/47	fails	matview parity and the settled verdict
#1265: reset_options refuses a matview again	phase5 45/47	fails	the same two
#1265: the drop hook skips a matview again (columnar_tableam.c:3982)	drop_cleanup 15/16	test_drop_cleanup.py fails	dropping a columnar matview takes its options row with it: got [1] want [0]

Runs

major	inheritance	parquet_import	phase5	drop_cleanup
15	16 + 10 skipped / pytest 4 + 1 unrun	18 + 7 skipped / pytest 1 unrun	47/47 / 11/11	16/16 / 8/8
16	the same as 15	the same as 15	47/47 / 11/11	16/16 / 8/8
17	26/26 / 14/14	25/25 / 7/7	47/47 / 11/11	16/16 / 8/8
18	26/26 / 14/14	25/25 / 7/7	47/47 / 11/11	16/16 / 8/8
19	26/26 / 14/14	25/25 / 7/7	47/47 / 11/11	16/16 / 8/8

Also measured: the guard half (438 tests), harness_selftest 1245/1245 and docs_style. cluster_tests goes 651 → 658 by collection, and TESTS.md gets §139 to §142.

🤖 Generated with Claude Code

[jdatcmd]

Approving 06851ab.

At the head, PG17: inheritance.sh 26/26 and test_inheritance.py 14 checks. The census at this head is 3612 / 2671 / 941; inheritance's two refusal rows moved from never to dated, matching the budget.

What I checked myself:

• phase5's move onto lib.sh's recorders: I ran phase5.sh on main and at this head and compared the RESULT records, name and verdict, sorted. Both have 47 records, all PASS, and they are identical.
• The #1259 revert (PgColumnarIsColumnarRelation returns true from relam alone). On my assert build it aborts the backend rather than giving your release build's 58P01. Either way the arms now catch it: shell 22/26, with both refusals got [no SQLSTATE (server closed the connection unexpectedly)] want [42809] and both liveness arms DOWN (1 backend crashes) / DOWN (2 backend crashes). pytest's refusal is red the same way. (Its "1 error" is fixture teardown losing the connection to the crashed server.) On a release build, before this PR, the old any-error refusal passed on 58P01, which is the bug itself. Requiring 42809 closes that.

I took the #1263 and #1265 reverts from your table without repeating them.

Order: #1444 → #1446 → #1448 all have my approval. The chain needs one restate onto main e04e1e0, which now has #1445. If #1449 lands first, it adds 7 ledger rows to restate over. I'll merge whichever is green and current first.

[jdatcmd]

Re-approving 7e7612c, a restate of 06851ab (which I approved) onto main e04e1e0. Every file's patch is identical to the approved head's, with TESTS.md section numbers masked, since main moved only by #1445, which touches no ledger file. The census at this head matches its budget paragraph.

[jdatcmd]

Re-approving 61b5250, a restate onto a4eef27 (#1449). Every test and suite file's patch is identical to the head I approved, with TESTS.md section numbers masked. Only the budget and expected_tests.txt moved, re-anchored on #1449's numbers. The census at this head matches its budget paragraph, and cluster_tests steps 643 → 646 → 653 → 660 along the chain.