I just want to add that if a filter aims to accept or reject groups of consecutive key-value pairs where each group constitutes a logical “record” or “document” then the boundary keys that could be used between independent filter instances are not necessarily also FDB shard boundaries.
In the scatter-gather model using colocated layer instances, this is easy to work around because you can adjust the given begin and end keys of each query to ensure they are at the beginning and end of a logical record, respectively. The colocated layer will still query mostly local data but possibly also a small amount of remote data.
This is not so easy to address in the model where the storage server is calling some external filter on the result set composed of only local kv pairs.