Make LruCache register its own metrics (#8561)

rather than have everything that instantiates an LruCache manage metrics
separately, have LruCache do it itself.

1 files changed, 35 insertions, 11 deletions

diff --git a/synapse/util/caches/lrucache.py b/synapse/util/caches/lrucache.py
index 33eae2b7c4..e4804f79e0 100644
--- a/synapse/util/caches/lrucache.py
+++ b/synapse/util/caches/lrucache.py
@@ -18,6 +18,7 @@ from functools import wraps
 from typing import Callable, Optional, Type, Union
 
 from synapse.config import cache as cache_config
+from synapse.util.caches import CacheMetric, register_cache
 from synapse.util.caches.treecache import TreeCache
 
 
@@ -43,27 +44,29 @@ class _Node:
 
 class LruCache:
     """
-    Least-recently-used cache.
+    Least-recently-used cache, supporting prometheus metrics and invalidation callbacks.
+
     Supports del_multi only if cache_type=TreeCache
     If cache_type=TreeCache, all keys must be tuples.
-
-    Can also set callbacks on objects when getting/setting which are fired
-    when that key gets invalidated/evicted.
     """
 
     def __init__(
         self,
         max_size: int,
+        cache_name: Optional[str] = None,
         keylen: int = 1,
         cache_type: Type[Union[dict, TreeCache]] = dict,
         size_callback: Optional[Callable] = None,
-        evicted_callback: Optional[Callable] = None,
+        metrics_collection_callback: Optional[Callable[[], None]] = None,
         apply_cache_factor_from_config: bool = True,
     ):
         """
         Args:
             max_size: The maximum amount of entries the cache can hold
 
+            cache_name: The name of this cache, for the prometheus metrics. If unset,
+                no metrics will be reported on this cache.
+
             keylen: The length of the tuple used as the cache key. Ignored unless
                 cache_type is `TreeCache`.
 
@@ -73,9 +76,13 @@ class LruCache:
 
             size_callback (func(V) -> int | None):
 
-            evicted_callback (func(int)|None):
-                if not None, called on eviction with the size of the evicted
-                entry
+            metrics_collection_callback:
+                metrics collection callback. This is called early in the metrics
+                collection process, before any of the metrics registered with the
+                prometheus Registry are collected, so can be used to update any dynamic
+                metrics.
+
+                Ignored if cache_name is None.
 
             apply_cache_factor_from_config (bool): If true, `max_size` will be
                 multiplied by a cache factor derived from the homeserver config
@@ -94,6 +101,19 @@ class LruCache:
         else:
             self.max_size = int(max_size)
 
+        if cache_name is not None:
+            metrics = register_cache(
+                "lru_cache",
+                cache_name,
+                self,
+                collect_callback=metrics_collection_callback,
+            )  # type: Optional[CacheMetric]
+        else:
+            metrics = None
+
+        # this is exposed for access from outside this class
+        self.metrics = metrics
+
         list_root = _Node(None, None, None, None)
         list_root.next_node = list_root
         list_root.prev_node = list_root
@@ -105,8 +125,8 @@ class LruCache:
                 todelete = list_root.prev_node
                 evicted_len = delete_node(todelete)
                 cache.pop(todelete.key, None)
-                if evicted_callback:
-                    evicted_callback(evicted_len)
+                if metrics:
+                    metrics.inc_evictions(evicted_len)
 
         def synchronized(f):
             @wraps(f)
@@ -169,13 +189,17 @@ class LruCache:
             return deleted_len
 
         @synchronized
-        def cache_get(key, default=None, callbacks=[]):
+        def cache_get(key, default=None, callbacks=[], update_metrics=True):
             node = cache.get(key, None)
             if node is not None:
                 move_node_to_front(node)
                 node.callbacks.update(callbacks)
+                if update_metrics and metrics:
+                    metrics.inc_hits()
                 return node.value
             else:
+                if update_metrics and metrics:
+                    metrics.inc_misses()
                 return default
 
         @synchronized