IIRC this doesn't break tests because its only hit on reconnection, or something.
Basically, when a process needs to fetch missing updates for the `typing` stream it needs to query the writer instance via HTTP (as we don't write typing notifications to the DB), the problem was that the endpoint (`streams`) was only registered on master and specifically not on the typing writer worker.
This changes the replication protocol so that the server does not send down `RDATA` for rows that happened before the client connected. Instead, the server will send a `POSITION` and clients then query the database (or master out of band) to get up to date.
When the `/keys/query` API is hit on client_reader worker Synapse may
decide that it needs to resync some remote deivces. Usually this happens
on master, and then gets cached. However, that fails on workers and so
it falls back to fetching devices from remotes directly, which may in
turn fail if the remote is down.
This allows registration to be handled by a worker, though the actual
write to the database still happens on master.
Note: due to the in-memory session map all registration requests must be
handled by the same worker.
