[query] aggregation of split_multi collapses to one partition

[patrick-schultz]

For example,

mt = hl.split_multi_hts(mt, permit_shuffle=True)
mt = mt.annotate_cols(foo=hl.agg.sum(...))

performs the aggregation in one giant partition.

This is a known issue, and we are working on it. A workaround is to write the result of split_multi to disk before using it (which is a good idea regardless).

The cause appears to the interaction of several things:

• split_multi computes two separate pieces, then merges them together with a TableUnion
• The aggregation doesn't care how the mt is keyed, and we propagate that upstream
• Because of that, the TableUnion ends up unioning two unkeyed tables
• TableUnion, as currently implemented, can only produce a result with a strict partitioner. In the case of no key fields, that means one partition.