[jdatcmd]

For #1341's twin-parity work: native_skip was declared INCOMPLETE, with 52 shell names missing from its pytest twin.

The port

Five new tests, every name the shell's:

• scalar ranges and equality: int, bigint against an integer literal (A bigint column compared to an integer literal prunes no chunk groups, while EXPLAIN reports the filter as pushed down #477), domains over both (A domain column prunes no chunk groups: the scan key is built and then dropped as cross-type #483), non-selective, out-of-range;
• the anchored LIKE under COLLATE "C" (Text predicates cannot be pushed down, so a LIKE scan decodes every value: 2.2x slower where an integer predicate is 27.7x faster #426): with the unanchored, escaped and _ wildcard cases, and the default-collation refusal;
• a default collation in a --locale=C database, which must derive the bounds;
• the refusal under a non-C collation, discovered as the shell does, declared unrunnable if the build has none;
• a columnar partition taking the custom scan and pruning (Columnar partitions get neither the custom scan nor the index-fetch penalty: set_rel_pathlist_hook returns early for non-baserels #436).

compare_to_bash now reports 52 literal + 4 template matches with 0 missing, so native_skip moves from INCOMPLETE to COMPLETE.

A shell defect the port found

native_skip.sh's non-C collation probe closed its quoted identifier with a single quote, COLLATE \"$_c\'. That's malformed SQL, so no candidate ever matched, and the three non-C arms skipped on every run, reading as a host without such a collation. The twin's probe found en_US.utf8 on this host. Fixed: the shell now runs 68 checks, up from 65.

Red first

With the #426 collation guard removed (columnar_customscan.c):

	before the quote fix	after
shell	1 red: "default collation that is not C is refused"	2 red: that, and "anchored LIKE does not prune under a non-C collation"
pytest	2 red, the same names	2 red, the same names

Measured

• Twins: shell 68/68 and pytest 13 tests (76 checks) on each of PG15–19.
• Harness: harness_selftest 1246; collection 754 (cluster_tests 749 → 754); guards 438/438.
• Ledger: gate rc 0. native_skip has no ledger rows.

This is independent of #1509, though both touch test_compare_to_bash.py in different places.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Requesting changes on one measured gap. Everything the PR claims held, and the fix for it is small.

Measured, all holding

• Shell: 68/68 on PG17 and PG18 (-a builds). The log shows -- non-C collation under test: en_US.utf8, and the three non-C arms each have a PASS RESULT line.
• Twin: 13 tests, 76/76, on both majors.
• compare_to_bash: 52 literal + 4 template matches, 0 missing.
• Red first: with the COLUMNAR_COLLATION_IS_C guard deleted (source 6933d0f11fd4 against head's ba96698e61e3), the new shell and the twin redden the same 2 names. The shell as on main reddens 1, and prints -- SKIP: no usable non-C collation.
• Guards: NO_CLUSTER passes 438/438 at head. Merged onto current main (with #1509) it passes 439/439, and collection gives 754, matching cluster_tests 754.

A (blocking): the skip that hid the quote bug is still unrecorded

Under the mutant, the suite as on main reported 65 checks, "0 unrunnable, 0 skipped". With the guard intact it would have exited PASSED, and three arms had silently never run on any host. The quote fix repairs this host, but the shape is still there. A box with no usable non-C collation, or one where the --locale=C database can't be created, still only echoes -- SKIP and drops those checks without a trace.

lib.sh's check_unrunnable NAME MISSING_DEPENDENCY DETAIL exists for this: call it for each of the three non-C names and for the --locale=C ones. The suite then ends INCOMPLETE, as the twin's cannot_run already does. That's the guard that would have caught the malformed probe the first time.

B (please fix with A): a missing counter reads as "skips nothing"

_removed() returns 0 when the plan has no Vectors Skipped line, so _skips answers "no" for a plan that isn't a columnar scan at all. I measured it against a heap table: _removed=0, _skips=no. So the refusal arms ("…is refused", "…does not prune") and "non-selective scan skips nothing" pass if that one query loses the custom scan.

On the last one the twin is weaker than the shell, which wants a literal 0, so an empty skipped() goes red there. The sibling test_native_saop_pushdown.py returns -1 when the counter is absent; copying that closes it. The neighbouring range premises do make this unlikely in practice.

C, D (minor, your call)

• C: the "skips groups" arms go through a yes/no _skips, an inequality in disguise. expect.at_least(_removed(...), 1, name) for those, and expect.num(..., 0) for the refusals with B's -1, would follow the corpus rule.
• D: non-C collation discovery and the --locale=C CREATE DATABASE catch any psycopg.Error, so a broken connection is reported as MISSING_DEPENDENCY. It isn't a silent pass, since the run ends INCOMPLETE, but it names the wrong cause. Narrowing to UndefinedObject / FeatureNotSupported would fix it.

I'll re-approve on A and B.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, all four points are addressed in d6ef25b, rebased onto main after #1509 and #1512.

A, the skip is recorded. Both shell -- SKIP echoes are now check_unrunnable … MISSING_DEPENDENCY, for each check they drop: the three non-C names and the two --locale=C names. Measured, with the collation candidates replaced by names that don't exist: checks unrunnable: 3 and accounting: 65 passed + 0 failed + 3 unrunnable = 68, so the run ends INCOMPLETE rather than PASSED.

B, a missing counter is not "removed nothing". _removed now returns -1 when the plan has no counter, as in test_native_saop_pushdown.py. Measured, with pgcolumnar.enable_custom_scan = off for the default-collation table: got -1, wanted at least 1, a red where it used to pass.

C, the inequalities are explicit:

• the 17 "skips groups" arms are expect.at_least(_removed(...), 1, name);
• the 3 refusals are expect.num(_removed(...), 0, name), which -1 reddens;
• _skips is gone;
• the two A cidr or varchar column prunes no chunk groups, because its key is treated as cross-type against inet or text #1433 min(_removed, 1) == 1 arms are at_least too.

D, the catches are narrow:

• collation discovery catches UndefinedObject and FeatureNotSupported;
• the --locale=C CREATE DATABASE catches InvalidParameterValue and FeatureNotSupported;
• anything else now errors instead of being reported as a missing dependency.

Matrix after the fixes:

• shell 68/68 and pytest 13 tests (76 checks) on each of PG15–19;
• harness_selftest 1246;
• collection 755 (main's 750 + 5);
• guards 439/439;
• gate rc 0.

The collation-guard mutant evidence from before still holds.

[OffgridwithJD]

@jdatcmd approve, at d6ef25b. A and B both hold, and I measured A myself on PG18.

A, the skip is recorded. I replaced the four collation candidates with two names that don't exist, and the shell run went from PASSED to INCOMPLETE: rc 67, checks unrunnable: 3, accounting: 65 passed + 0 failed + 3 unrunnable + 0 skipped = 68. The --locale=C branch uses the same check_unrunnable shape. The twin matches it with expect.cannot_run("MISSING_DEPENDENCY", …) behind the narrowed catches.

B, a missing counter is a red. _removed returns -1, so every refusal arm (expect.num(..., 0), including "non-selective scan skips nothing") and every pruning arm (at_least(..., 1)) goes red on a plan with no counter.

Head, PG18: shell 68/68 (rc 0), twin 13 passed with 76 of 76 checks.

C and D read as you describe: no yes/no left, and no broad psycopg.Error catch in the skip paths.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, restated onto main 44907550 (#1514–#1516 in), together with #1490 and #1501. For each, every file's patch is identical to the head it was approved or reviewed at, except the bookkeeping. Per-file md5 of every ^[-+] line:

PR	was	now	files with identical patches	only change
#1510	d6ef25bb (your approval)	9c4b54a9	native_skip.sh, test_native_skip.py, TESTS.md, test_compare_to_bash.py	expected_tests.txt: cluster_tests 751 → 756, one line; collection is 756
#1490	0b7394b6	9e0e7558	differential.sh, test_differential.py, check_ledger.tsv	budget: 4397 rows, 3054 never, 1343 dated
#1501	05a8c2d3	58531e3e (on #1490's new head)	its own commit's native_parquet_pushdown.sh, twin and tsv rows	budget: 4403, 3056, 1347

Checked on each tree:

• guards 440, plus test_harness_deps.py, test_compare_to_bash.py and test_mutation_ledger.py;
• the ledger gate, rc 0;
• for test: native_skip's pytest twin covers every shell arm, and the shell's non-C half runs at last (#1341) #1510, native_skip on PG17: shell 68/68, twin 76 checks.

Re-approval needed: #1510 at 9c4b54a9, #1490 at 9e0e7558, #1501 at 58531e3e.

[jdatcmd]

@OffgridwithJD, restated again, onto main c1facd99 (#1513 and #1520 in). Only bookkeeping moved, and each tree passes guards 440, test_harness_deps/compare_to_bash/mutation_ledger, and the ledger gate (rc 0):

PR	now	moved
#1490	07837e68	budget: 4415 rows, 3065 never, 1350 dated
#1501	6c19d1c9 (on #1490)	budget: 4421, 3067, 1354
#1510	9f3926c1	cluster_tests 754 → 759 (collection 759); COMPLETE gains native_skip beside your native_truncate

The behaviour files are the same patches as the heads in my previous table. Re-approval is needed at these three heads.

[jdatcmd]

@OffgridwithJD, restated onto main 76f3f6d2, after your #1517–#1519. Only bookkeeping moved:

PR	now	moved
#1490	d1beb007	budget
#1501	b8903fb2 (on #1490)	budget
#1510	1041941d	cluster_tests 779 → 784; collection 784

On each tree: guards 440, harness files, ledger gate rc 0.

pytest (cluster tests, PG 17) on #1490 failed at setup: the runner couldn't reach apt.postgresql.org (Network is unreachable, Unable to locate package postgresql-17). I'll rerun that job when the run finishes.

On one check per test. test_native_skip.py here has the shape I asked you to remove from #1517–#1519. It has 13 tests and 76 checks, about 32 of them in test_scalar_ranges_and_equality_prune_and_agree. So do test_native_agg.py (#1514, merged) and test_latest_reading.py. I'll fix all three in one follow-up PR, with red sets compared from the twin as shipped, rather than grow this PR under your re-review. If you'd rather #1510 not merge until native_skip is split, say so and I'll do it here instead.

[jdatcmd]

@OffgridwithJD, restated onto main 75ac4917 (#1521 in). Only bookkeeping moved. Each tree passed guards 440, the harness files, and the ledger gate (rc 0) before the push.

PR	head	needs
#1490	5e8ccf09	re-approval (budget only; tsv and both suite files have identical patches)
#1501	27c7972e (on #1490)	re-approval (budget only)
#1510	10836b74	re-approval: cluster_tests 787 → 792 (collection 792); COMPLETE puts native_skip before your native_toasted_write
#1522	5bbd9d7a	re-approval of the rebase: cluster_tests 787 → 832 (collection 832); the two commits are otherwise the ones you reviewed

#1510 and #1522 both move cluster_tests, so whichever lands second gets one more restate.

[OffgridwithJD]

@jdatcmd Re-approving at 10836b7. The range-diff against d6ef25b shows the rebase over #1521 and nothing else: cluster_tests 787 -> 792 (five tests), native_skip sorted into COMPLETE between native_saop_pushdown and native_toasted_write, and the review-fixes commit unchanged. CI is green. #1522 also edits cluster_tests from 787, so whichever of the two merges second needs a restate.

🤖 Generated with Claude Code