JoinSelection Rule to choose physical join implementation: HashJoin(Partitioned or CollectLeft) or SortMergeJoin base on Stats

[mingmwang]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]
(This section helps Arrow developers understand the context and why for this feature, in addition to the what)

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

[mingmwang]

@alamb @andygrove @isidentical @Dandandan @yahoNanJing
I plan to add a new physical rule for physical join implementation selection based on stats.

The basic idea is if one of side(left or right) is small enough(threshold on size or row count, for example 10M), will choose HashJoin:CollectLeft.
Else if either side exceed that threshold but still smaller than hash join threshold(estimated_size/partition_count <= 128M for example), will choose HashJoin: Partitioned.
Otherwise choose SortMergeJoin.

Please share your thoughts.

[Dandandan]

Sounds like a good plan.

For hash join, probably needs some benchmarking to figure out good defaults and avoid performance degradation. CollectLeft limits the amount of parellization on the left side: building the hash table is relatively expensive and is done (at least currently) in a single thread. In quite a few cases it might be more beneficial to do a (local) hash repartitioning which is relatively cheap.
It also depends on the size of the probe/right side: if that's e.g. >100x as big as the left side it might be beneficial to avoid the hash repartitioning on the right side by switching to CollectLeft.

[alamb]

I agree with @Dandandan that doing some benchmarking on sort merge join is likely a good idea. I don't think we have much at the moment as it is never used.

What I think would be an ideal solution, though maybe harder to implement, is to NOT decide between has or sort merge join at plan time, but to decide at runtime

So that would involve starting out using HashJoin but if the hash table spilled (or exceeded some size threshold) sort/spill it and then switch to sort-merge-join