[jdatcmd]

Covers the #1341 core-am item "ALTER TABLE forms that scan existing rows through the AM". Test-only: no source change.

What it adds

A new suite, test/alter_table_scans.sh, and its pytest twin, test_alter_table_scans.py. A columnar table keeps a deleted row in its file until compaction, so every form that reads the existing rows must see the live rows and only those. Each form runs on a columnar table and on heap with the same rows (3000 ordinary rows, plus a NULL in v, a second id = 1, and a negative v), and each check states both outcomes, columnar|heap:

form	with the violating row	after that row is deleted
ALTER COLUMN v SET NOT NULL	23502|23502	ok|ok
ADD PRIMARY KEY (id)	23505|23505	ok|ok
ADD CONSTRAINT … CHECK (v > 0) NOT VALID, then VALIDATE CONSTRAINT	ok|ok, then 23514|23514	ok|ok, recorded as validated

Two rewrites follow. ALTER COLUMN v TYPE bigint keeps 3000 rows with sum 4501500. ADD COLUMN z float8 DEFAULT random() gives every live row a value and brings back no deleted row. Finally, the primary key finds id = 1 once, through an index scan.

Evidence

• Green: 15/15 on each of PG15–19, and the twin on each.
• Red on PG17:

mutant	what went red
every scan returns deleted rows (the reader's delete filter disabled)	8 checks. SET NOT NULL got 23502|ok after the delete; both rewrites got 23505|ok; row counts got 3002/4501499
the index build marks every row alive	ADD PRIMARY KEY after the delete got 23505|ok; the index-plan arm, which needs that key

Two checks I first named "premise:" reddened under these mutants, so they are arms under #1236 and carry no prefix.

• Ledger: alter_table_scans is seeded with 15 rows, 10 dated. The census is 2774 + 15 = 2789 rows, 2095 + 5 = 2100 never, and 689 dated. The gate passes against origin/main. The harness selftest passes 1245/1245 and the guard set 433/433.
• Registration: run_all_versions.sh, test_compare_to_bash.py COMPLETE, TESTS.md §115. test: the settings reference is held to the server (#1341 items 42, 57, 83) #1414 and test: a statement does not see its own writes after a mid-statement flush (#1341) #1416 also add a §115, so whichever merges later renumbers. cluster_tests goes 565 → 566 by collection.

🤖 Generated with Claude Code

https://claude.ai/code/session_019VvWuX1zUwk6Y513waRd4N

[OffgridwithJD]

Approved at 038eeea.

• What it holds: SET NOT NULL, ADD PRIMARY KEY and VALIDATE CONSTRAINT are refused over a violating row and accepted once that row is deleted. ALTER COLUMN TYPE and ADD COLUMN with a volatile default keep the live rows and only those, and the primary key lookup works afterwards. Each check reads as columnar|heap.
• Red (PG17), both mutants as stated:
  • The reader with if (false && deleted && !rs->emitDeleted) (.so 81544b33) gives 7 + 8 failed, for example the deleted rows are gone from both tables, which hold the same rows: got [3002/4501499|3000/4501500], SET NOT NULL succeeds once the NULL row is deleted, as on heap: got [23502|ok] and ALTER COLUMN TYPE rewrites the table, as on heap: got [23505|ok].
  • The index build with alive = true (.so 4c47e347) gives 13 + 2: ADD PRIMARY KEY succeeds once the duplicate is deleted, as on heap: got [23505|ok] and the lookup of id 1 is an index scan on the primary key: got [0|1]. The second follows from the first, since there's no primary key to plan on, and the ledger paragraph says so.
  • The twin fails at the first of each, with the same text.
• Green: shell 15/15 and twin 1 passed on PG17 and PG18 with the code restored (.so b9726239 again). The ledger, 2789 / 2100 / 689, matches the tsv, the gate agrees against origin/main, all 15 names have rows, and cluster_tests 566 matches collection.

🤖 Generated with Claude Code

[OffgridwithJD]

Re-approved at de6e61d, stacked on #1417.

• Same patch: fc6a07a matches what I approved at 038eeea, outside the count files and ledger rows included. de6e61d removes pgc_conn.autocommit = True, which is now required: since #1432 the proxy raises on that assignment instead of silently keeping it. Net, the twin is the code I approved.
• Census: 2984 / 2278 / 706 from the tsv, which matches the paragraph's 2969 + 15, 2273 + 5 and 696 + 10.
• CI: 13 checks green and 2 still running at the time of review. The pytest job is the one that would catch the proxy change.

🤖 Generated with Claude Code