test: parallel_flush_optin gains its pytest twin, and a flush no worker touched no longer passes as parallel (#1341) by jdatcmd · Pull Request #1643 · commandprompt/pgcolumnar

jdatcmd · 2026-10-21T22:06:34Z

Stacked on #1642; merges after it. This head sits on my restated stack (#1622, #1628–#1631), so the registry counts in the commit message are re-derived against that stack. Counts quoted in the body below are from the port's original base.

What I re-ran myself before opening (PG17): the noworkers mutant (no worker ever registered). Main's shell passes 5/5. This shell fails 1: with the GUC on, a wide-numeric flush dispatches parallel (not a silent serial fall-back): got [serial fall-back:[parallel_flush: 0 of 8 columns by 0 worker;.... The twin fails the same name.

@OffgridwithJD, this is the pytest twin of test/parallel_flush_optin.sh (#1341). It also fixes two shell checks that main's shell passed with the property broken. One commit, 90b2ba80, on 1d1e0850.

What was fixed

"with the GUC on, a wide-numeric flush dispatches parallel (not a silent serial fall-back)" read only the dispatch line. The flush prints that line, ending -> parallel, before it registers any worker. I mutated flush_columns_parallel so that no worker ever registers. Every column then fell back to serial, and main's shell still passed 5 of 5. The flush already logs how the work was split (parallel_flush: N of M columns by K worker(s)). The check now also reads that split line and needs worker columns in it.
"opting in stores byte-identical chunks to the serial path" flushed eight copies of g. I mutated the collection step so that each worker's column is stored under the next column index. Because every column held the same values, the stored bytes did not change, and main's shell passed 5 of 5. The eight columns now have eight different shapes. The fingerprint also adds the rows read back, which catches bytes that change without changing a length.
"premise: parallel_flush is off by default" becomes parallel_flush is off by default under pytest audit: 260 confirmed checks that cannot fail, prove something else, or carry the wrong label #1339. The boot default is pgcolumnar's own setting, and changing it to true reddens the check.

Main's shell, this shell and the twin under each mutant (PG17)

mutant	main's shell	this shell	twin
`columnar_write_state.c`: no worker ever registers (`if (false && RegisterDynamicBackgroundWorker(...))`)	5/5 pass	1 red: the ON check (`0 of 8 columns by 0 worker`)	1 red, same name
`columnar_write_state.c`: each worker's column is stored under the next index (`colResults[(col + 1) % natts]`)	5/5 pass	1 red: the chunks check	1 red, same name
`columnar_tableam.c`: the `pgcolumnar.parallel_flush` boot default is set to true	2 red: default off, default serial	2 red, same	2 red, same names
`columnar_write_state.c`: `pflush_metrics` adds each value stream's first byte to `valbytes` (a content term)	1 red: the metric check	1 red, same	1 red, same name
clean, PG15 to PG19	5/5	5/5 on each	5/5 on each

None of the mutants crashed the backend.

The twin reads the DEBUG1 lines as notices on the inserting session. On the shared pytest cluster, which keeps the default max_worker_processes, workers produced columns on every major. The clean twin runs passed the ON check, and that check fails unless the split line shows worker columns.

Gates (over `1d1e0850`)

Ledger: 5 rows, 5 dated by the four mutations and 0 never. The census moves 3718 -> 3718 never, 2026 -> 2031 dated, 5749 rows.
The ledger gate passes on each of the five clean logs: census stated 3718, ledger holds 3718: they agree, ceiling against origin/main: 117 -> 116, which does not rise.
Guards: 467 passed.
compare_to_bash.py: VERDICT: every bash property is covered (5 literal matches).
Collecting the cluster files finds 2338 tests, which matches cluster_tests 2338 (was 2333).
harness_selftest.sh: PASSED.
TESTS.md gets section 252, and unported_suites.txt loses the suite.

This branch conflicts with the parallel_degree twin branch. Both start from 1d1e0850, so both claim TESTS.md section 252, suites_not_covered 117 -> 116 and the next cluster_tests step, and both append to the ledger and the budget. Whichever lands second has to be rebased and its registries regenerated.

Logs

The logs are in /tmp/claude-1006/-home-jdatcmd/ed97498f-58ee-4453-98bc-cbf3cfaf41b1/scratchpad/portlogs/parallel_flush_optin/:

Main's shell: main-pg17.log and mainshell-<mutant>.log.
This shell: shell-*.log. The five clean logs merged into the ledger are shell-clean-pg15.log to shell-clean-pg19.log.
The twin: twin-*.log.
Gates: gate-pg*.log, guards.log and selftest.log.
Patches: mut-<mutant>.patch.

🤖 Generated with Claude Code

…k no longer tests one path twice (#1341) test/pytest/test_read_stream.py ports the checks of test/read_stream.sh by name. The suite compares five query shapes with a heap twin, with pgcolumnar.enable_read_stream on and then off; one shell fix, in both twins: - The setting arrives by ALTER DATABASE SET, whose error the shell's q drops. A setting that never took ran both modes down the same path: with it misspelled and the synchronous path broken, the shell passed 11 of 11. Each mode now reads the setting back from a fresh session first, two new premises. - The ten comparisons were named through $mode in a loop; they are literal. The stripe premise carries its count and leaves lossy_arms.tsv. Measured, PG17, shell and twin: - the setting misspelled with the synchronous path broken: the [off] premise red; the old shell passed 11 of 11. - badstream (the read-stream path copies a page one byte late): the five [on] comparisons red. badsync (the synchronous path does): the five [off] comparisons red. Each red set is identical in both twins. - clean: 13 of 13 on PG15 to PG19 in both twins. Ledger: 13 rows, 10 dated by the two mutations and 3 never. suites_not_covered 113 -> 112, cluster_tests 2454 -> 2467, TESTS.md section 256. Census 3793 -> 3794 never, 2073 -> 2085 dated, 5879 rows. The ledger gate and the guards pass over 8d2bf1b. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… file no longer passes it (#1341) test/pytest/test_parquet_level_width.py ports all 7 checks of test/parquet_level_width.sh by name. It reads the source in Python, needs no cluster, and is listed in NO_CLUSTER. Three shell fixes, in both twins: - "no local copy of the helper survives" summed grep -c over the writer and the reader; grep prints nothing for a missing file, so with the writer's path wrong the sum read 0 and the shell passed 7 of 7. It now reads only when both files exist, and says so otherwise. - "and no signed 1 << remains in it" counted over the extracted body, so with the header unreadable it read 0 and passed. It now needs a body. - The body checks read the comments inside the body, so a guard or bound commented out of the code still passed 7 of 7 (OffgridwithJD's commentguard and commentbound). The body now has its comments blanked, multi-line ones included. Measured, PG16 shell and the twin: - signedshift (pq_bits_for shifts a signed 1): 2 red. unbounded (the b < 31 bound removed): 1 red. localcopy (a local bits_for back in the writer): 1 red. commentguard (the non-positive guard commented out): 1 red. commentbound (the b < 31 bound commented out): 1 red. commentml (the guard in a multi-line comment): 1 red. Each red set is identical in both twins (the comment mutants on PG17). - clean: 7 of 7 on PG15 to PG19 in the shell, and in the twin. Ledger: 7 rows, 5 dated by the four mutations and 2 never. suites_not_covered 144 -> 143, guard_tests 464 -> 471, TESTS.md section 224. Census 3543 -> 3545 never, 1798 -> 1803 dated, 5348 rows. The ledger gate and the guards (471) pass over 555e5ef. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… error no longer pass it (#1341) test/pytest/test_parquet_nested.py ports all 6 checks of test/parquet_nested.sh by name, each declining by the shell's skip name without pyarrow, as test_arrow_nested.py does for the Arrow sibling. Two shell fixes, in both twins, mirroring arrow_nested's: - "nested Parquet values match expectation" compared only the rows the file held, so a zero-row file of the same schema printed MATCH; it now requires every id 1..5000 once. - "reject multi-dim array" passed on any error, so with the export pointed at a missing table it read as the refusal; md_refusal() now reads the multi-dimensional message. Measured, PG16, shell and twin: - emptyasnull (an empty array written with definition level 0, so it reads back NULL): 1 red, the values check. nomultidim (the refusal removed): 1 red, the multi-dim check. Each red set is identical in both twins. - clean: 6 of 6 on PG15 to PG19 in both twins. Ledger: 6 rows, 2 dated by the two mutations and 4 never. suites_not_covered 111 -> 110, cluster_tests 2467 -> 2473, TESTS.md section 258. Census 3796 -> 3800 never, 2090 -> 2092 dated, 5892 rows. The ledger gate and the guards pass over 1b1922e. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…nts no longer agree (#1341) test/pytest/test_parquet_nested_import.py ports all 7 checks of test/parquet_nested_import.sh by name. The two pyarrow arms decline by their own names without pyarrow, as the shell's two skips do. One shell fix, in both twins: - "import count matches" compared two counts with check, under which two empty readings agree: with both tables missing, the shell passed it. It is check_num. The set hashes already gave each failed query its own value, so the round-trips could not agree on two failures. Measured, PG16, shell and twin: - readempty (the reader takes an empty Parquet list for a NULL array): 3 red, the two round-trips and the pyarrow oracle; the counts stay green, as they should. The red set is identical in both twins. - clean: 7 of 7 on PG15 to PG19 in both twins. Ledger: 7 rows, 3 dated by the mutation and 4 never. suites_not_covered 110 -> 109, cluster_tests 2473 -> 2480, TESTS.md section 259. Census 3800 -> 3804 never, 2092 -> 2095 dated, 5899 rows. The ledger gate and the guards pass over 9109cc2. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

OffgridwithJD

@jdatcmd Approving 91488075. Measured at this head on PG17, shell then twin as shipped (make clean per mutant). I reviewed this PR's own commit; the suites ran at the head.

run	main's shell	this shell	twin
head, PG17 / PG18	—	5/5, 5/5	5/5, 5/5
`noworkers` (`if (false && RegisterDynamicBackgroundWorker(...))`)	5/5 pass	1 red: the GUC-on check	SAME 1
`shift` (`colResults[(col + 1) % natts]`)	5/5 pass	1 red: byte-identical chunks	SAME 1
`defaulton` (boot value true)	—	2 red: off by default, default dispatches serial	SAME 2
`gateall` (mine, over-apply: `goParallel = (true \|\| pgcolumnar_parallel_flush) && …`)	—	1 red: a default flush dispatches serial	SAME 1
`swap01` (mine: columns 0 and 1 swapped on collection)	—	1 red: byte-identical chunks	SAME 1

Also read through:

a failed or missing dispatch reading can't pass: check_text refuses an empty expected side;
the twin's .split(" -> ") keeps the QUERY_ERROR prefix, so a failed session still reds;
the fingerprint's row read-back is what swap01 trips.

🤖 Generated with Claude Code

… not its fixture CI failed the harness guards and both cluster-test jobs with: REFUSED test_parquet_nested_import.py imports pyarrow outside require_module (line 63) REFUSED test_parquet_nested_import.py imports pyarrow outside require_module (line 64) The module fixture imported pyarrow itself to write the reference file. A fixture cannot decline, so main's guard refuses the import. The fixture now keeps the database and the directory, and the pyarrow arm runs once, from the first pyarrow test, with the modules that test's require_module returned. The check names and the count are unchanged. Measured: - Red, the parent's twin: test_harness_deps "the guard half of the corpus runs without a database driver" fails with the REFUSED lines above. - Green: the twin is 7/7 on PG17 and PG19. With pyarrow hidden it is 5 pass + 2 declined, each pyarrow arm by its own name. - Mutant, the reference file one row short: both pyarrow arms fail. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…nd no longer passes as survived (#1341) test/pytest/test_parquet_count_bounds.py ports all 17 checks of test/parquet_count_bounds.sh by name, using the shell's crafter and declining by the shell's skip name without pyarrow. Three shell fixes, in both twins: - The survival checks read SELECT 1, which answers again after crash recovery. With the num_values bound removed (#570), the 2^62 file crashed the backend and "and the backend survived it" and "survived the list over-read" passed on main. They now also need the server log's count of killed backends unmoved since setup, as avro_manifest does since #1576; the same crash now reads "no ([1], 1 backend crashes)". - The two phantom-row checks caught only the exact count the crafter declares (40, 8). They now catch any count above the file's true one. No mutation here produced a different count, so this hardening is not shown red first: under both #571 guards removed, the leak was the declared 40, which main caught too. - The crafter premise covers all seven files; the three list files were never checked. Measured, PG16, shell and twin: - no570 (the num_values range check removed): the backend crashes. Both harnesses red the rejection, the survival check and the detail; what else reds depends on how long recovery takes, so the counts differ (7 in the shell, 11 in the twin), all from reads during recovery. - no571 alone is masked: pq_require_entry rejects the same files at decode time with the same message, so the suite stays 17/17. - noentry (pq_require_entry refuses nothing): 2 red, the two list checks, the same in both twins. no571 with noentry: 9 red, the same 9. - clean: 17 of 17 on PG15 to PG19 in both twins. Ledger: 17 rows, 2 dated by noentry and 15 never. The num_values and num_rows mutations crash the backend and redden collateral checks, so the ledger cannot attribute them. suites_not_covered 125 -> 124, cluster_tests 2169 -> 2186, TESTS.md section 233. The ledger gate, guards (467), name grading and the self-test pass over 9bd1850. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…t by importing pyarrow in it CI failed the harness guards and both cluster-test jobs with: REFUSED test_parquet_count_bounds.py imports pyarrow outside require_module (line 83) The module fixture imported pyarrow to decide whether to build at all. A fixture cannot decline, so main's guard refuses the import. Each test now declines by the shell's skip name first, and only then asks for the fixture with request.getfixturevalue. The fixture no longer imports anything, and it never runs on a box without pyarrow. The check names and the count are unchanged. Measured: - Red, the parent: test_harness_deps "the guard half of the corpus runs without a database driver" fails with the REFUSED line above. - Green: test_harness_deps and test_capability_decline, 30 passed, no REFUSED line. The twin is 17/17 on PG17 and PG19; with pyarrow hidden it is 17 declined. - Probe: wanting 9 honest rows fails with "got = 8, want = 9", so the fixture still crafts and imports. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… no longer passes (#1341) test/pytest/test_parallel_degree.py ports all 6 checks of test/parallel_degree.sh by name. Three shell checks could pass with nothing measured; all three are fixed in both twins. - With the server stopped after ANALYZE, main's shell passed three of six checks. "identical data is granted the same parallel degree" compared "none" with "none"; it is now check_num. "both tables hold the same rows" compared "" with ""; it is now check_num. "the worker cap is not what decides either arm" read a missing worker count as 0; it now needs two numbers. With the edit, the fixed shell reds all six. - Three checks lose their premise: prefix under #1339. The columnar rows, the columnar partial path and the columnar degree under the cap are pgcolumnar's own work, and a one-line change reddens each. Measured, PG17, shell and twin: - server stopped after ANALYZE: 6 red in the fixed shell and in the twin (the twin drops its database instead); main's shell passed 3 of 6. - the #451 estimate never applied (2 workers against heap's 4): 1 red. The estimate over-asking (pages * 9): 1 red. Over-asking into the cap (pages * 243): 2 red. The partial path never added: 3 red (main's shell: 2, the cap check passed). An insert that drops one row: 1 red. Each red set is identical in both twins. - clean: 6 of 6 on PG15 to PG19 in both twins. Ledger: 6 rows, 4 dated by the five mutations and 2 never. suites_not_covered 108 -> 107, cluster_tests 2497 -> 2503, TESTS.md section 261. Census 3819 -> 3821 never, 2097 -> 2101 dated, 5922 rows. The ledger gate and the guards pass over 23f3d52. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…er touched no longer passes as parallel (#1341) test/pytest/test_parallel_flush_optin.py ports all 5 checks of test/parallel_flush_optin.sh by name. Two shell checks could pass with the property broken; both are fixed in both twins. - "with the GUC on, a wide-numeric flush dispatches parallel (not a silent serial fall-back)" read only the dispatch line, which says parallel before any worker is registered. With no worker ever registered, every column fell back to serial and main's shell passed 5 of 5. The check now also needs the split line ("N of M columns by K worker(s)") to show worker columns. - "opting in stores byte-identical chunks to the serial path" flushed eight copies of g, so a parallel assembly that stored each column under the next index wrote the same bytes, and main's shell passed 5 of 5. The eight columns now have eight different shapes, and the fingerprint adds the rows read back. - "premise: parallel_flush is off by default" loses its prefix under #1339: the boot default is pgcolumnar's own, and setting it to true reddens it. Measured, PG17, shell and twin: - no worker ever registered: 1 red (the ON check); main's shell 0. - each worker column stored under the next index: 1 red (the chunks check); main's shell 0. - the boot default set to true: 2 red. A content term added to the dispatch metric: 1 red. Each red set is identical in both twins. - clean: 5 of 5 on PG15 to PG19 in both twins. Ledger: 5 rows, 5 dated by the four mutations and 0 never. suites_not_covered 107 -> 106, cluster_tests 2503 -> 2508, TESTS.md section 262. Census 3821 -> 3821 never, 2101 -> 2106 dated, 5927 rows. The ledger gate and the guards pass over 7548c88. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

jdatcmd · 2026-10-22T01:06:37Z

Re-approval needed: rebased only, onto #1631's fixed head (d7f291c). This PR's own commit is unchanged; its guards job failed only on #1630/#1631's pyarrow imports. At the stack tip (bdaaa09): the guard's import scan finds 0 sites in 258 files, and the guard set is 474 passed.

OffgridwithJD

@jdatcmd Re-approving the rebased head 3db2eba0 (I approved 91488075).

Content unchanged: compared file by file against 91488075, this PR's own +/- lines match, and it's one commit over its new parent. Only the budget, expected_tests and TESTS.md differ, and those only restate counts.
Gates at the stack tip bdaaa098: guards 474 passed, collection 2670 = cluster_tests, census 3850 = the tsv's never count.

🤖 Generated with Claude Code

jdatcmd and others added 4 commits October 21, 2026 21:42

jdatcmd mentioned this pull request Oct 21, 2026

test: parquet_export_stats gains its pytest twin, and an export that folds valid dates to null no longer passes (#1341) #1644

Merged

OffgridwithJD approved these changes Oct 21, 2026

View reviewed changes

OffgridwithJD mentioned this pull request Oct 21, 2026

Missing tests: 87 confirmed behaviours no shell or pytest arm exercises (#1339 audit) #1341

Closed

87 tasks done

jdatcmd and others added 5 commits October 22, 2026 01:00

jdatcmd force-pushed the test/1341-parallel-flush-optin-twin branch from 9148807 to 3db2eba Compare October 22, 2026 01:06

OffgridwithJD approved these changes Oct 22, 2026

View reviewed changes

jdatcmd merged commit 09cc6c9 into main Oct 22, 2026
15 checks passed

Navigation Menu

Repository navigation

test: parallel_flush_optin gains its pytest twin, and a flush no worker touched no longer passes as parallel (#1341) - #1643

test: parallel_flush_optin gains its pytest twin, and a flush no worker touched no longer passes as parallel (#1341)#1643
jdatcmd merged 9 commits into
mainfrom
test/1341-parallel-flush-optin-twin

jdatcmd commented Oct 21, 2026

Uh oh!

OffgridwithJD left a comment

Uh oh!

jdatcmd commented Oct 22, 2026

Uh oh!

OffgridwithJD left a comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jdatcmd commented Oct 21, 2026

What was fixed

Main's shell, this shell and the twin under each mutant (PG17)

Gates (over 1d1e0850)

Logs

Uh oh!

OffgridwithJD left a comment

Choose a reason for hiding this comment

Uh oh!

jdatcmd commented Oct 22, 2026

Uh oh!

OffgridwithJD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Gates (over `1d1e0850`)