Repository navigation
q5 from LDBC SNB complex14 set regressed #1369
Description
uv pip install -i https://test.pypi.org/simple/ ladybug==0.22.0.dev20261017
python query_complex14.py 5
Query 5:
MATCH (person:Person {ID: $personId})-[:knows*1..2]-(friend:Person)
WHERE NOT person = friend
WITH DISTINCT friend
MATCH (friend)<-[membership:hasMember]-(forum:Forum)
WHERE membership.joinDate > TIMESTAMP($minDate)
WITH forum, COLLECT(friend) AS friends
OPTIONAL MATCH (author:Person)<-[:postHasCreator]-(post:Post)<-[:containerOf]-(forum)
WHERE author IN friends
WITH forum, COUNT(post) AS postCount
RETURN forum.title AS forumName, postCount
ORDER BY postCount DESC, forum.ID ASC
LIMIT 20;
shape: (20, 2)
...
Completed 1 query(ies) in 1.14s
Same query regressed to 2.34s on ladybug==0.22.0.dev20261021
Cc: @zahariash
Activity
adsharma commentedon Oct 23, 2026
hyperdimensionalcomputingai/graph-benchmark-ldbc#17 has the queries. @prrao87 prefers a separate repo.
zahariash commentedon Oct 23, 2026
Looking into it.
zahariash commentedon Oct 24, 2026
This is the lookup join from #1285, which landed between the two nightlies. On a build of the 10-21 nightly's commit, turning off only the lookup join gives back the 10-17 plan and instruction count for q5, so nothing else in that range is involved.
The cap checks the outer side's estimate before flattening it. Here that is 73 rows, while the flattened estimate is 11,895 and the actual probe count is 1,569,282. #1370 checks the flattened estimate instead, which puts q5 back on the hash join and keeps the lookups in q1, q3, q7 and q17.