Fix catchup-on-reconnect for the Federation Stream (#7374)

looks like we managed to break this during the refactorathon.

synapse/replication/tcp/resource.py

@@ -80,7 +80,7 @@ class ReplicationStreamer(object):
             for stream in STREAMS_MAP.values():
                 if stream == FederationStream and hs.config.send_federation:
                     # We only support federation stream if federation sending
-                    # hase been disabled on the master.
+                    # has been disabled on the master.
                     continue
 
                 self.streams.append(stream(hs))

synapse/replication/tcp/streams/_base.py

@@ -104,7 +104,8 @@ class Stream(object):
         implemented by subclasses.
 
         current_token_function is called to get the current token of the underlying
-        stream.
+        stream. It is only meaningful on the process that is the source of the
+        replication stream (ie, usually the master).
 
         update_function is called to get updates for this stream between a pair of
         stream tokens. See the UpdateFunction type definition for more info.

synapse/replication/tcp/streams/federation.py

@@ -15,7 +15,7 @@
 # limitations under the License.
 from collections import namedtuple
 
-from synapse.replication.tcp.streams._base import Stream, db_query_to_update_function
+from synapse.replication.tcp.streams._base import Stream, make_http_update_function
 
 
 class FederationStream(Stream):
@@ -35,21 +35,33 @@ class FederationStream(Stream):
     ROW_TYPE = FederationStreamRow
 
     def __init__(self, hs):
-        # Not all synapse instances will have a federation sender instance,
-        # whether that's a `FederationSender` or a `FederationRemoteSendQueue`,
-        # so we stub the stream out when that is the case.
-        if hs.config.worker_app is None or hs.should_send_federation():
+        if hs.config.worker_app is None:
+            # master process: get updates from the FederationRemoteSendQueue.
+            # (if the master is configured to send federation itself, federation_sender
+            # will be a real FederationSender, which has stubs for current_token and
+            # get_replication_rows.)
             federation_sender = hs.get_federation_sender()
             current_token = federation_sender.get_current_token
-            update_function = db_query_to_update_function(
-                federation_sender.get_replication_rows
-            )
+            update_function = federation_sender.get_replication_rows
+
+        elif hs.should_send_federation():
+            # federation sender: Query master process
+            update_function = make_http_update_function(hs, self.NAME)
+            current_token = self._stub_current_token
+
         else:
-            current_token = lambda: 0
+            # other worker: stub out the update function (we're not interested in
+            # any updates so when we get a POSITION we do nothing)
             update_function = self._stub_update_function
+            current_token = self._stub_current_token
 
         super().__init__(hs.get_instance_name(), current_token, update_function)
 
+    @staticmethod
+    def _stub_current_token():
+        # dummy current-token method for use on workers
+        return 0
+
     @staticmethod
     async def _stub_update_function(instance_name, from_token, upto_token, limit):
         return [], upto_token, False