Delete stale non-e2e devices for users, take 2 (#14595)

This should help reduce the number of devices e.g. simple bots the repeatedly login rack up.

We only delete non-e2e devices as they should be safe to delete, whereas if we delete e2e devices for a user we may accidentally break their ability to receive e2e keys for a message.

2 files changed, 108 insertions, 2 deletions

diff --git a/synapse/handlers/device.py b/synapse/handlers/device.py
index d4750a32e6..7674c187ef 100644
--- a/synapse/handlers/device.py
+++ b/synapse/handlers/device.py
@@ -52,6 +52,7 @@ from synapse.util import stringutils
 from synapse.util.async_helpers import Linearizer
 from synapse.util.caches.expiringcache import ExpiringCache
 from synapse.util.cancellation import cancellable
+from synapse.util.iterutils import batch_iter
 from synapse.util.metrics import measure_func
 from synapse.util.retryutils import NotRetryingDestination
 
@@ -421,6 +422,9 @@ class DeviceHandler(DeviceWorkerHandler):
 
         self._check_device_name_length(initial_device_display_name)
 
+        # Prune the user's device list if they already have a lot of devices.
+        await self._prune_too_many_devices(user_id)
+
         if device_id is not None:
             new_device = await self.store.store_device(
                 user_id=user_id,
@@ -452,6 +456,31 @@ class DeviceHandler(DeviceWorkerHandler):
 
         raise errors.StoreError(500, "Couldn't generate a device ID.")
 
+    async def _prune_too_many_devices(self, user_id: str) -> None:
+        """Delete any excess old devices this user may have."""
+        device_ids = await self.store.check_too_many_devices_for_user(user_id)
+        if not device_ids:
+            return
+
+        # We don't want to block and try and delete tonnes of devices at once,
+        # so we cap the number of devices we delete synchronously.
+        first_batch, remaining_device_ids = device_ids[:10], device_ids[10:]
+        await self.delete_devices(user_id, first_batch)
+
+        if not remaining_device_ids:
+            return
+
+        # Now spawn a background loop that deletes the rest.
+        async def _prune_too_many_devices_loop() -> None:
+            for batch in batch_iter(remaining_device_ids, 10):
+                await self.delete_devices(user_id, batch)
+
+                await self.clock.sleep(1)
+
+        run_as_background_process(
+            "_prune_too_many_devices_loop", _prune_too_many_devices_loop
+        )
+
     async def _delete_stale_devices(self) -> None:
         """Background task that deletes devices which haven't been accessed for more than
         a configured time period.
@@ -481,7 +510,7 @@ class DeviceHandler(DeviceWorkerHandler):
             device_ids = [d for d in device_ids if d != except_device_id]
         await self.delete_devices(user_id, device_ids)
 
-    async def delete_devices(self, user_id: str, device_ids: List[str]) -> None:
+    async def delete_devices(self, user_id: str, device_ids: Collection[str]) -> None:
         """Delete several devices
 
         Args:
diff --git a/synapse/storage/databases/main/devices.py b/synapse/storage/databases/main/devices.py
index a5bb4d404e..08ccd46a2b 100644
--- a/synapse/storage/databases/main/devices.py
+++ b/synapse/storage/databases/main/devices.py
@@ -1569,6 +1569,72 @@ class DeviceBackgroundUpdateStore(SQLBaseStore):
 
         return rows
 
+    async def check_too_many_devices_for_user(self, user_id: str) -> List[str]:
+        """Check if the user has a lot of devices, and if so return the set of
+        devices we can prune.
+
+        This does *not* return hidden devices or devices with E2E keys.
+        """
+
+        num_devices = await self.db_pool.simple_select_one_onecol(
+            table="devices",
+            keyvalues={"user_id": user_id, "hidden": False},
+            retcol="COALESCE(COUNT(*), 0)",
+            desc="count_devices",
+        )
+
+        # We let users have up to ten devices without pruning.
+        if num_devices <= 10:
+            return []
+
+        # We prune everything older than N days.
+        max_last_seen = self._clock.time_msec() - 14 * 24 * 60 * 60 * 1000
+
+        if num_devices > 50:
+            # If the user has more than 50 devices, then we chose a last seen
+            # that ensures we keep at most 50 devices.
+            sql = """
+                SELECT last_seen FROM devices
+                LEFT JOIN e2e_device_keys_json USING (user_id, device_id)
+                WHERE
+                    user_id = ?
+                    AND NOT hidden
+                    AND last_seen IS NOT NULL
+                    AND key_json IS NULL
+                ORDER BY last_seen DESC
+                LIMIT 1
+                OFFSET 50
+            """
+
+            rows = await self.db_pool.execute(
+                "check_too_many_devices_for_user_last_seen", None, sql, (user_id,)
+            )
+            if rows:
+                max_last_seen = max(rows[0][0], max_last_seen)
+
+        # Now fetch the devices to delete.
+        sql = """
+            SELECT DISTINCT device_id FROM devices
+            LEFT JOIN e2e_device_keys_json USING (user_id, device_id)
+            WHERE
+                user_id = ?
+                AND NOT hidden
+                AND last_seen < ?
+                AND key_json IS NULL
+            ORDER BY last_seen
+        """
+
+        def check_too_many_devices_for_user_txn(
+            txn: LoggingTransaction,
+        ) -> List[str]:
+            txn.execute(sql, (user_id, max_last_seen))
+            return [device_id for device_id, in txn]
+
+        return await self.db_pool.runInteraction(
+            "check_too_many_devices_for_user",
+            check_too_many_devices_for_user_txn,
+        )
+
 
 class DeviceStore(DeviceWorkerStore, DeviceBackgroundUpdateStore):
     # Because we have write access, this will be a StreamIdGenerator
@@ -1627,6 +1693,7 @@ class DeviceStore(DeviceWorkerStore, DeviceBackgroundUpdateStore):
                 values={},
                 insertion_values={
                     "display_name": initial_device_display_name,
+                    "last_seen": self._clock.time_msec(),
                     "hidden": False,
                 },
                 desc="store_device",
@@ -1672,7 +1739,15 @@ class DeviceStore(DeviceWorkerStore, DeviceBackgroundUpdateStore):
             )
             raise StoreError(500, "Problem storing device.")
 
-    async def delete_devices(self, user_id: str, device_ids: List[str]) -> None:
+    @cached(max_entries=0)
+    async def delete_device(self, user_id: str, device_id: str) -> None:
+        raise NotImplementedError()
+
+    # Note: sometimes deleting rows out of `device_inbox` can take a long time,
+    # so we use a cache so that we deduplicate in flight requests to delete
+    # devices.
+    @cachedList(cached_method_name="delete_device", list_name="device_ids")
+    async def delete_devices(self, user_id: str, device_ids: Collection[str]) -> dict:
         """Deletes several devices.
 
         Args:
@@ -1709,6 +1784,8 @@ class DeviceStore(DeviceWorkerStore, DeviceBackgroundUpdateStore):
         for device_id in device_ids:
             self.device_id_exists_cache.invalidate((user_id, device_id))
 
+        return {}
+
     async def update_device(
         self, user_id: str, device_id: str, new_display_name: Optional[str] = None
     ) -> None: