4 files changed, 50 insertions, 9 deletions
diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
index af65bfe7b8..bf3a66eae4 100644
--- a/synapse/util/caches/descriptors.py
+++ b/synapse/util/caches/descriptors.py
@@ -75,6 +75,7 @@ class Cache(object):
         self.cache = LruCache(
             max_size=max_entries, keylen=keylen, cache_type=cache_type,
             size_callback=(lambda d: len(d)) if iterable else None,
+            evicted_callback=self._on_evicted,
         )
 
         self.name = name
@@ -83,6 +84,9 @@ class Cache(object):
         self.thread = None
         self.metrics = register_cache(name, self.cache)
 
+    def _on_evicted(self, evicted_count):
+        self.metrics.inc_evictions(evicted_count)
+
     def check_thread(self):
         expected_thread = self.thread
         if expected_thread is None:
diff --git a/synapse/util/caches/dictionary_cache.py b/synapse/util/caches/dictionary_cache.py
index d4105822b3..1709e8b429 100644
--- a/synapse/util/caches/dictionary_cache.py
+++ b/synapse/util/caches/dictionary_cache.py
@@ -132,9 +132,13 @@ class DictionaryCache(object):
                 self._update_or_insert(key, value, known_absent)
 
     def _update_or_insert(self, key, value, known_absent):
-        entry = self.cache.setdefault(key, DictionaryEntry(False, set(), {}))
+        # We pop and reinsert as we need to tell the cache the size may have
+        # changed
+
+        entry = self.cache.pop(key, DictionaryEntry(False, set(), {}))
         entry.value.update(value)
         entry.known_absent.update(known_absent)
+        self.cache[key] = entry
 
     def _insert(self, key, value, known_absent):
         self.cache[key] = DictionaryEntry(True, known_absent, value)
diff --git a/synapse/util/caches/expiringcache.py b/synapse/util/caches/expiringcache.py
index 6ad53a6390..0aa103eecb 100644
--- a/synapse/util/caches/expiringcache.py
+++ b/synapse/util/caches/expiringcache.py
@@ -79,7 +79,11 @@ class ExpiringCache(object):
         while self._max_len and len(self) > self._max_len:
             _key, value = self._cache.popitem(last=False)
             if self.iterable:
-                self._size_estimate -= len(value.value)
+                removed_len = len(value.value)
+                self.metrics.inc_evictions(removed_len)
+                self._size_estimate -= removed_len
+            else:
+                self.metrics.inc_evictions()
 
     def __getitem__(self, key):
         try:
diff --git a/synapse/util/caches/lrucache.py b/synapse/util/caches/lrucache.py
index cf5fbb679c..1c5a982094 100644
--- a/synapse/util/caches/lrucache.py
+++ b/synapse/util/caches/lrucache.py
@@ -49,7 +49,24 @@ class LruCache(object):
     Can also set callbacks on objects when getting/setting which are fired
     when that key gets invalidated/evicted.
     """
-    def __init__(self, max_size, keylen=1, cache_type=dict, size_callback=None):
+    def __init__(self, max_size, keylen=1, cache_type=dict, size_callback=None,
+                 evicted_callback=None):
+        """
+        Args:
+            max_size (int):
+
+            keylen (int):
+
+            cache_type (type):
+                type of underlying cache to be used. Typically one of dict
+                or TreeCache.
+
+            size_callback (func(V) -> int | None):
+
+            evicted_callback (func(int)|None):
+                if not None, called on eviction with the size of the evicted
+                entry
+        """
         cache = cache_type()
         self.cache = cache  # Used for introspection.
         list_root = _Node(None, None, None, None)
@@ -61,8 +78,10 @@ class LruCache(object):
         def evict():
             while cache_len() > max_size:
                 todelete = list_root.prev_node
-                delete_node(todelete)
+                evicted_len = delete_node(todelete)
                 cache.pop(todelete.key, None)
+                if evicted_callback:
+                    evicted_callback(evicted_len)
 
         def synchronized(f):
             @wraps(f)
@@ -111,12 +130,15 @@ class LruCache(object):
             prev_node.next_node = next_node
             next_node.prev_node = prev_node
 
+            deleted_len = 1
             if size_callback:
-                cached_cache_len[0] -= size_callback(node.value)
+                deleted_len = size_callback(node.value)
+                cached_cache_len[0] -= deleted_len
 
             for cb in node.callbacks:
                 cb()
             node.callbacks.clear()
+            return deleted_len
 
         @synchronized
         def cache_get(key, default=None, callbacks=[]):
@@ -132,14 +154,21 @@ class LruCache(object):
         def cache_set(key, value, callbacks=[]):
             node = cache.get(key, None)
             if node is not None:
-                if value != node.value:
+                # We sometimes store large objects, e.g. dicts, which cause
+                # the inequality check to take a long time. So let's only do
+                # the check if we have some callbacks to call.
+                if node.callbacks and value != node.value:
                     for cb in node.callbacks:
                         cb()
                     node.callbacks.clear()
 
-                    if size_callback:
-                        cached_cache_len[0] -= size_callback(node.value)
-                        cached_cache_len[0] += size_callback(value)
+                # We don't bother to protect this by value != node.value as
+                # generally size_callback will be cheap compared with equality
+                # checks. (For example, taking the size of two dicts is quicker
+                # than comparing them for equality.)
+                if size_callback:
+                    cached_cache_len[0] -= size_callback(node.value)
+                    cached_cache_len[0] += size_callback(value)
 
                 node.callbacks.update(callbacks)