Forbid allocate_empty_primary if Stale Copy Exists?

[original-brownbear]

Currently, it appears that if we have a stale copy of some unassigned shard on node_a, we can force allocate an empty primary on node_b and needless forgo the data that is still on node_a.

I think just like we added a check for the existence of stale data on a node for allocate_stale_primary we should add a check for the non-existence of any stale data before allowing allocate_empty_primary.
There isn't really a scenario in which it would be advantageous to allow forgoing a stale shard-copy by allocating an empty primary on a node when one could at least allocate a stale primary on another node is there?

[elasticmachine]

Pinging @elastic/es-distributed

[DaveCTurner]

What happens if the only remaining shard copy cannot be assigned because it is corrupt? I think in this case we might want to be able to ignore it and allocate an empty primary in its place.

Yet I have definitely seen people mistype allocate_stale_primary as allocate_empty_primary and this idea would have caught that.

[original-brownbear]

What happens if the only remaining shard copy cannot be assigned because it is corrupt? I think in this case we might want to be able to ignore it and allocate an empty primary in its place.

You mean, simply make allocate_stale_primary create an empty primary for corrupted shard copy? I'd be +1 to that idea :)