[jdatcmd]

Fixes #1435, found by @OffgridwithJD reviewing #1429.

The bug: a wrong answer

On a columnar table, an ungrouped aggregate ignored a WHERE that references no column:

query	columnar before	heap
SELECT count(*), min(id), max(id) FROM t WHERE (SELECT false)	every row's aggregates	0 rows
... WHERE now() < '2000-01-01'	every row's aggregates	0 rows
... WHERE current_user = 'x'	every row's aggregates	0 rows

Cause: a column-free WHERE is a pseudoconstant (gating) qual, and extract_actual_clauses(..., false) leaves it out of the node's quals. With no other qual, needsScan was false and the node answered from the zone maps. The node's own gating refusal sits in the scan-fold branch, which this case never reached.

The fix

In src/columnar_vector.c, the ungrouped node refuses a gating qual before needsScan is decided, as the grouped node does. The query goes to the ordinary plan, which applies the one-time filter. A query with no WHERE, or with a WHERE on a column, is unchanged.

Tests (red first)

test/native_agg.sh and a new test_a_where_with_no_column_in_it_is_applied in test_native_agg.py:

• each of the three gating WHEREs gives heap's answer;
• a premise that the same aggregate with no WHERE is the vectorized node, so the arms are aimed at it;
• a premise that heap's answer to a false gate is 0/-/-, so the comparison can't pass on two wrong answers.

On main the three arms are red in both twins: got [10000/1/10000] want [0/-/-] in the shell, and got '5000/1/5000' in the twin.

Evidence

• Green: native_agg 40/40 and the twin 15 passed, on each of PG15 to PG19, with no build warnings. On PG17, native_groupagg 94/94, int8_agg_int128 34/34 and batch_fold_explain 7/7.
• Ledger: native_agg gains 5 rows. The three gating arms are dated by the tree before the fix; the two premises arrive never. The census is 3380 + 5 = 3385 rows, 2546 never, 839 dated, matching the tsv. The gate gives rc=0, cluster_tests is 601 by collection, the guard set passes 438/438, and docs_style passes.
• Two commits: the tests (red on main), then the fix, its bookkeeping and the CHANGELOG entry.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Approved at 01640fd.

Red first, then green, on PG18:

tree	native_agg.sh	test_native_agg.py
bc40b67 (the tests only)	37/40: the three gating arms red, got [10000/1/10000] want [0/-/-]	1 failed, 1 passed: got '5000/1/5000' want '0/-/-'
01640fd (the fix)	40/40	2 passed

The .so differs between the two runs (9c1993a0 and 58d3e7f6).

Probes against heap on the fixed build, 200000 rows, with count/min/max/sum compared each time:

• Gates: a false gate, a true gate (200000/1/200000/20000100000 on both sides) and a STABLE current_setting gate.
• A gate beside a column qual: v > 10 AND (SELECT false), and the same with a true now() gate.
• Other plan shapes: GROUP BY under a false gate (0 groups), HAVING (SELECT false) (0 rows), and a prepared WHERE $1 with false, both custom and forced-generic (0).
• Settings: enable_ungrouped_vector_agg = on, and forced parallel with 4 workers, for false and true gates and GROUP BY. The parallel plan is Gather -> Partial Aggregate -> Result (One-Time Filter) -> Parallel Custom Scan.

All agree with heap. A query with no WHERE still plans Columnar Vectorized Aggregates: 3, so the refusal is no wider than the gate. Returning before any ungrouped path is built is the right place for it.

Found while probing, a separate bug, on main: #1437. Any join with a columnar fact and a gating qual fails with pgcolumnar runtime filter expected a core Hash Join, true gates and EXPLAIN included. enable_join_runtime_filter = off avoids it. It doesn't block this PR, which doesn't touch that path.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, this is now 79f514a, stacked on #1426 (996413c), which is on #1429, so the three merge in order without restating.

• One real conflict: columnar_vector.c, where test: row-level security holds on every columnar read path (#1341) #1429's leak check and this fix both go right before the ungrouped node builds its quals. I kept both, the gating refusal first.
• On the combined tree (PG17): native_agg 40/40, rls_query_paths 24/24, native_groupagg 94/94, type_matrix 41/41, and the twins pass.
• Bookkeeping: the census is 3450 / 2583 / 867 by the tsv, cluster_tests is 603, and the gate gives rc=0. The two commits are kept, so red first still shows.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Approved at 79f514a, stacked on #1426 (996413c).

• The patch: identical to 01640fd's apart from one blank line.
• The conflict: resolved as it should be. In columnar_vector.c the gating refusal comes first and #1429's pgcolumnar_quals_could_leak check follows it, and both just return.
• The combined tree on PG17: native_agg 40/40 and rls_query_paths 24/24 in the shell, and both twins (test_native_agg.py and test_rls_query_paths.py) pass.

My PG18 red-first runs and probes from the 01640fd approval still cover the fix.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, restated as 3e0a7c0 onto #1426's new head (c1441b4), which added the heap-not-empty arms. Only the budget conflicted, and it's re-derived (3466 rows, 2598 never, 868 dated, matching the tsv). The patch is identical to 79f514a, which you approved, outside the budget and expected_tests.txt. cluster_tests is 603 by collection, and the gate gives rc=0.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Approved at 3e0a7c0, on #1426's c1441b4.

• The patch: the src, the tests and the CHANGELOG are identical to 79f514a's, and only the budget was re-derived.
• Runs: my PG17 run of the combined tree and the PG18 red-first runs still cover it, since c1441b4 only adds type_matrix arms.

🤖 Generated with Claude Code