diff --git a/changelog.d/4521.feature b/changelog.d/4521.feature
new file mode 100644
index 000000000..bda713adf
--- /dev/null
+++ b/changelog.d/4521.feature
@@ -0,0 +1 @@
+Implement MSC1708 (.well-known routing for server-server federation)
\ No newline at end of file
diff --git a/synapse/http/federation/matrix_federation_agent.py b/synapse/http/federation/matrix_federation_agent.py
index eeb3665f7..26649e70b 100644
--- a/synapse/http/federation/matrix_federation_agent.py
+++ b/synapse/http/federation/matrix_federation_agent.py
@@ -298,16 +298,18 @@ class MatrixFederationAgent(object):
             response = yield make_deferred_yieldable(
                 self._well_known_agent.request(b"GET", uri),
             )
+            body = yield make_deferred_yieldable(readBody(response))
+            if response.code != 200:
+                raise Exception("Non-200 response %s" % (response.code, ))
         except Exception as e:
-            logger.info("Connection error fetching %s: %s", uri_str, e)
-            self._well_known_cache.set(server_name, None, WELL_KNOWN_INVALID_CACHE_PERIOD)
-            defer.returnValue(None)
+            logger.info("Error fetching %s: %s", uri_str, e)
 
-        body = yield make_deferred_yieldable(readBody(response))
+            # add some randomness to the TTL to avoid a stampeding herd every hour
+            # after startup
+            cache_period = WELL_KNOWN_INVALID_CACHE_PERIOD
+            cache_period += random.uniform(0, WELL_KNOWN_DEFAULT_CACHE_PERIOD_JITTER)
 
-        if response.code != 200:
-            logger.info("Error response %i from %s", response.code, uri_str)
-            self._well_known_cache.set(server_name, None, WELL_KNOWN_INVALID_CACHE_PERIOD)
+            self._well_known_cache.set(server_name, None, cache_period)
             defer.returnValue(None)
 
         try:
@@ -328,8 +330,8 @@ class MatrixFederationAgent(object):
         )
         if cache_period is None:
             cache_period = WELL_KNOWN_DEFAULT_CACHE_PERIOD
-            # add some randomness to the TTL to avoid a stampeding herd every hour after
-            # startup
+            # add some randomness to the TTL to avoid a stampeding herd every 24 hours
+            # after startup
             cache_period += random.uniform(0, WELL_KNOWN_DEFAULT_CACHE_PERIOD_JITTER)
         else:
             cache_period = min(cache_period, WELL_KNOWN_MAX_CACHE_PERIOD)