Adjust subscription logic a bit (#416)

Co-authored-by: Martin Hjelmare <[email protected]>

matter_server/server/device_controller.py

@@ -7,6 +7,7 @@
 from datetime import datetime
 from functools import partial
 import logging
+import random
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Type, TypeVar, cast
 
 from chip.ChipDeviceCtrl import CommissionableNode
@@ -564,16 +565,11 @@ async def _subscribe_node(self, node_id: int) -> None:
                 # Wildcard endpoint, specific cluster
                 attr_subscriptions.append(cluster)
 
-        if len(attr_subscriptions) > 50:
-            # prevent memory overload on node and fallback to wildcard sub if too many
-            # individual subscriptions (e.g. bridges)
-            attr_subscriptions = "*"  # type: ignore[assignment]
-
-        if not node.attribute_subscriptions:
-            # temp fix for backwards compatbility with HA releases below 2023.7
-            # fallback to wildcard subscriptions if we have no explicit
-            # node subscriptions defined.
-            # TODO: remove this after HA release 2023.8
+        if len(attr_subscriptions) > 9:
+            # strictly taken a matter device can only handle 9 individual subscriptions
+            # (3 subscriptions of 3 paths per fabric)
+            # although the device can probably handle more, we play it safe and opt for
+            # wildcard as soon as we have more than 9 paths to watch for.
             attr_subscriptions = "*"  # type: ignore[assignment]
 
         # check if we already have an subscription for this node,
@@ -591,21 +587,31 @@ async def _subscribe_node(self, node_id: int) -> None:
         self._attr_subscriptions[node_id] = attr_subscriptions
         async with node_lock:
             node_logger.debug("Setting up attributes and events subscription.")
+            # Use a report interval of 0, X which means we want to receive state changes
+            # as soon as possible (the 0 as floor) but we want to receive a report
+            # at least once every X seconds, this is also used to detect the node is still alive.
+            # A resubscription will be initiated automatically by the sdk
+            # if there was no report within the interval.
+            # NOTE 1: The report interval ceiling is subject to a lot of discussion
+            # as setting it too low causes a lot of (unneeded) traffic and causes network
+            # congestion as well as drains batteries on sleeping devices.
+            # Preferred would be to set the interval as high as possible ( 30 mins or even 1 hour)
+            # but that would also mean that detecting that a device is offline would be delayed
+            # by that amount of time as the interval ceiling also meant as liveness detection.
+            # For now we settle on (more or less) 10 minutes but we might need to increase this
+            # even more in the future.
+            # see also: https://github.com/project-chip/connectedhomeip/issues/29804
+            # NOTE 2: We randomize the interval a bit to prevent all nodes reporting
+            # at the exact same time, causing congestion.
+            interval_floor = 0
+            interval_ceiling = random.randint(500, 700)
             sub: Attribute.SubscriptionTransaction = await self.chip_controller.Read(
                 nodeid=node_id,
-                # In order to prevent network congestion due to wildcard subscriptions on all nodes,
-                # we keep a list of attributes we are explicitly interested in.
                 attributes=attr_subscriptions,
                 # simply subscribe to urgent device events only (e.g. button press etc.)
-                # non urgent events are disagnostic reports etc. for which we have no usecase (yet).
+                # non urgent events are diagnostic reports etc. for which we have no usecase (yet).
                 events=[("*", 1)],
-                # Use a report interval of 0, 300 which means we want to receive state changes
-                # as soon as possible (the 0 as floor) but we want to receive a report
-                # at least once every 5 minutes (300 as ceiling).
-                # This is also used to detect the node is still alive.
-                # A resubscription will be initiated automatically by the sdk
-                # if there was no report within the interval.
-                reportInterval=(0, 300),
+                reportInterval=(interval_floor, interval_ceiling),
                 # Use fabricfiltered as False to detect changes made by other controllers
                 # and to be able to provide a list of all fabrics attached to the device
                 fabricFiltered=False,
@@ -717,15 +723,17 @@ def resubscription_attempted(
             if nextResubscribeIntervalMsec / 1000 > MAX_POLL_INTERVAL:
                 # workaround to handle devices that are unplugged
                 # from power for a longer period of time
-                # cancel subscription and add this node to our node polling job
-                # TODO: fix this once OerationalNodeDiscovery is available:
+                # where the sdk is extending the poll timeout at every attempt
+                # until even 1,5 hours which is way too long.
+                # instead a device back alive should be detected using mDNS,
+                # which is not yet implemented in the core sdk.
+                # For now, we just override the timeout.
+                # NOTE 1: fix this once OperationalNodeDiscovery is available:
                 # https://github.com/project-chip/connectedhomeip/pull/26718
-                sub.Shutdown()
-                self._subscriptions.pop(node_id)
-                assert self.server.loop
-                self.server.loop.create_task(
-                    self._check_interview_and_subscription(node_id, MAX_POLL_INTERVAL)
-                )
+                # https://github.com/project-chip/connectedhomeip/issues/29663
+                # NOTE 2: We could also just implement zeroconf/mdns ourselves
+                # to listen for the announcements.
+                sub.OverrideLivenessTimeoutMs(MAX_POLL_INTERVAL * 1000)
 
         def resubscription_succeeded(
             transaction: Attribute.SubscriptionTransaction,