You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
No test reaches the grouped batch fold with deletes: a fold that ignored the delete mask passes every suite tried #1675
Every grouped query in that suite has a text key or an expression key. Those take the row path, so the batch fold never runs there. Deleted rows are therefore never presented to it.
What is unknown
Whether another suite (for example join_vector_agg, the native agg suites, or native_agg_deletes) reaches the grouped batch fold with deleted rows. That has to be checked by running the same mutant across them before a test is written.
Next step
Run the mutant over every suite that groups on a columnar table, and record which, if any, goes red.
If none does: add an arm that groups on a plain integer key, deletes rows, and compares with core's grouped aggregate (enable_group_vectorization=off). Show it is red under the mutant and green on main, in both twins.
Closing: the gap is covered elsewhere. I ran the same mutant (the grouped batch fold ignores the delete mask, mut-gnodelete.patch) across all 25 suites that group on a columnar table, on PG16 at main. Two catch it:
native_groupagg_batch, 5 failed: "deletes: answers match the heap mirror", "toggle-differential: plain", "toggle-differential: filtered", "toggle-differential: two keys", "parallel: and they equal the heap mirror";
native_groupagg, 1 failed: "oracle exact: default cap runs the high-cardinality key".
The other 23 pass under it, parallel_vector_agg among them, which is what #1673 measured. So no suite needs a new arm. parallel_vector_agg simply doesn't reach that fold, and its PR body already says so.
Activity
jdatcmd commentedon Oct 22, 2026
Closing: the gap is covered elsewhere. I ran the same mutant (the grouped batch fold ignores the delete mask,
mut-gnodelete.patch) across all 25 suites that group on a columnar table, on PG16 at main. Two catch it:native_groupagg_batch, 5 failed: "deletes: answers match the heap mirror", "toggle-differential: plain", "toggle-differential: filtered", "toggle-differential: two keys", "parallel: and they equal the heap mirror";native_groupagg, 1 failed: "oracle exact: default cap runs the high-cardinality key".The other 23 pass under it,
parallel_vector_aggamong them, which is what #1673 measured. So no suite needs a new arm.parallel_vector_aggsimply doesn't reach that fold, and its PR body already says so.