Add specific metric to time long-running `/messages` requests (#13533)

changelog.d/13533.misc

@@ -0,0 +1 @@
+Track HTTP response times over 10 seconds from `/messages` (`synapse_room_message_list_rest_servlet_response_time_seconds`).

synapse/rest/client/room.py

@@ -19,6 +19,8 @@ import re
 from typing import TYPE_CHECKING, Awaitable, Dict, List, Optional, Tuple
 from urllib import parse as urlparse
 
+from prometheus_client.core import Histogram
+
 from twisted.web.server import Request
 
 from synapse import event_auth
@@ -60,6 +62,35 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+# This is an extra metric on top of `synapse_http_server_response_time_seconds`
+# which times the same sort of thing but this one allows us to see values
+# greater than 10s. We use a separate dedicated histogram with its own buckets
+# so that we don't increase the cardinality of the general one because it's
+# multiplied across hundreds of servlets.
+messsages_response_timer = Histogram(
+    "synapse_room_message_list_rest_servlet_response_time_seconds",
+    "sec",
+    [],
+    buckets=(
+        0.005,
+        0.01,
+        0.025,
+        0.05,
+        0.1,
+        0.25,
+        0.5,
+        1.0,
+        2.5,
+        5.0,
+        10.0,
+        30.0,
+        60.0,
+        120.0,
+        180.0,
+        "+Inf",
+    ),
+)
+
 
 class TransactionRestServlet(RestServlet):
     def __init__(self, hs: "HomeServer"):
@@ -560,6 +591,7 @@ class RoomMessageListRestServlet(RestServlet):
         self.auth = hs.get_auth()
         self.store = hs.get_datastores().main
 
+    @messsages_response_timer.time()
     async def on_GET(
         self, request: SynapseRequest, room_id: str
     ) -> Tuple[int, JsonDict]: