On Wed, Sep 9, 2026 at 7:09 PM Jeevan Chalke
<[email protected]> wrote:
> Note up front: the reproducer below (case 1) only crashes on REL_18_STABLE
> and earlier. On REL_19_STABLE, add_partial_path()'s dominance check was
> rewritten to compare startup cost as well as total cost, and that
> incidentally avoids the specific comparison that discards the shared path for
> this data. The underlying bug is still present on REL_19_STABLE; I just
> haven't found data that gets the new cost comparison to hit it.

I found a query that reproduces it on v19 and master:

set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set min_parallel_table_scan_size = 0;
set min_parallel_index_scan_size = 0;
set max_parallel_workers_per_gather = 1;

explain (costs off)
select unique1, generate_series(1, 5000000) from tenk1 where unique2 < 100
union
select 1, 2;
ERROR:  unrecognized node type: 3301232

The trick here is the SRF with a large fan-out: it adds the same large
per-row cost to both partial index scans, whose startup costs are
already equal, so their total costs end up within the fuzz factor.
When grouping_planner() re-submits them to final_rel, the sorted one
evicts the other and frees it, while the Gather built over it is still
final_rel's cheapest path.

So this bug is real on all branches, and I think we should fix it.

I think the parent_rel check approach in your 0001 works.  The rule it
states is that add_path() and add_partial_path() free only paths built
for the rel they are building, which is what the freeing was always
relying on.

One consequence: postgres_fdw's add_foreign_ordered_paths() and
add_foreign_final_paths() build their upper paths with the input rel
as parent, since postgresGetForeignPlan() deparses from that rel.
Under the check those paths are never freed when they lose a
comparison.  That is a node or two per query, so maybe it's ok.

The alternatives are a flat copy of Path, and reference counting (and
maybe more).  Copying also fixes the bug, so maybe it would be a
workable solution too.  Reference counting is more appealing in
principle, but I think it's far too large a change to back-patch.

Thoughts?

- Richard


Reply via email to