Well, sure, it doesn't _require_ it. But the alternative is higher query latency. You have to collect a filtered set of the join column from table A, then ship them over to the servers responsible for table B. If your interconnect has high latency or low bandwidth, this can be painful - particularly if the intermediate set of keys is very large.

Hence why the datastore simply disallows this - yes, it's _possible_ to make joins work on larger datasets. But it's not possible to make _arbitrary_ joins work well on larger datasets.