Not all optimal tpch queries plans are really optimal

[MBkkt]

Intro

Right now, according to 1314950

Queries TPC-H q1, 3, 4, 6, 12, 14, 16 are optimal, but actually some of them are not.

Queries

1. q1, q6, q12, q14 -- are ok
2. q3 -- ideally should have Velox operator that allows to combine last hashjoin (lineitem x (orders x customer)) with aggregation after
3. q4 -- we should use right semi filter join instead of right semi project join + filter
4. q16 -- we should use left anti join instead of left semi project join + filter not

Conclusion

q3 doesn't looks like a bug, just improvement that can be done. First in Velox, then in Axiom.

q4, q16 looks like same issue, we should try to avoid semi project when it's possible.

Questions

Why doesn't Velox have right anti join?

CC @mbasmanova @pashandor789

[mbasmanova]

Why doesn't Velox have right anti join?

I don't remember off the top of my head. FYI, ANTI joins are discussed in https://facebookincubator.github.io/velox/develop/anti-join.html