summary refs log tree commit diff
path: root/synapse/util/caches
diff options
context:
space:
mode:
authorBrendan Abolivier <babolivier@matrix.org>2020-06-10 11:42:30 +0100
committerBrendan Abolivier <babolivier@matrix.org>2020-06-10 11:42:30 +0100
commitec0a7b9034806d6b2ba086bae58f5c6b0fd14672 (patch)
treef2af547b1342795e10548f8fb7a9cfc93e03df37 /synapse/util/caches
parentchangelog (diff)
parent1.15.0rc1 (diff)
downloadsynapse-ec0a7b9034806d6b2ba086bae58f5c6b0fd14672.tar.xz
Merge branch 'develop' into babolivier/mark_unread
Diffstat (limited to 'synapse/util/caches')
-rw-r--r--synapse/util/caches/__init__.py150
-rw-r--r--synapse/util/caches/descriptors.py115
-rw-r--r--synapse/util/caches/expiringcache.py29
-rw-r--r--synapse/util/caches/lrucache.py56
-rw-r--r--synapse/util/caches/response_cache.py4
-rw-r--r--synapse/util/caches/snapshot_cache.py94
-rw-r--r--synapse/util/caches/stream_change_cache.py158
-rw-r--r--synapse/util/caches/treecache.py4
-rw-r--r--synapse/util/caches/ttlcache.py2
9 files changed, 358 insertions, 254 deletions
diff --git a/synapse/util/caches/__init__.py b/synapse/util/caches/__init__.py
index b50e3503f0..dd356bf156 100644
--- a/synapse/util/caches/__init__.py
+++ b/synapse/util/caches/__init__.py
@@ -1,6 +1,6 @@
 # -*- coding: utf-8 -*-
 # Copyright 2015, 2016 OpenMarket Ltd
-# Copyright 2019 The Matrix.org Foundation C.I.C.
+# Copyright 2019, 2020 The Matrix.org Foundation C.I.C.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -15,34 +15,24 @@
 # limitations under the License.
 
 import logging
-import os
+from sys import intern
+from typing import Callable, Dict, Optional
 
-import six
-from six.moves import intern
+import attr
+from prometheus_client.core import Gauge
 
-from prometheus_client.core import REGISTRY, Gauge, GaugeMetricFamily
+from synapse.config.cache import add_resizable_cache
 
 logger = logging.getLogger(__name__)
 
-CACHE_SIZE_FACTOR = float(os.environ.get("SYNAPSE_CACHE_FACTOR", 0.5))
-
-
-def get_cache_factor_for(cache_name):
-    env_var = "SYNAPSE_CACHE_FACTOR_" + cache_name.upper()
-    factor = os.environ.get(env_var)
-    if factor:
-        return float(factor)
-
-    return CACHE_SIZE_FACTOR
-
-
 caches_by_name = {}
-collectors_by_name = {}
+collectors_by_name = {}  # type: Dict
 
 cache_size = Gauge("synapse_util_caches_cache:size", "", ["name"])
 cache_hits = Gauge("synapse_util_caches_cache:hits", "", ["name"])
 cache_evicted = Gauge("synapse_util_caches_cache:evicted_size", "", ["name"])
 cache_total = Gauge("synapse_util_caches_cache:total", "", ["name"])
+cache_max_size = Gauge("synapse_util_caches_cache_max_size", "", ["name"])
 
 response_cache_size = Gauge("synapse_util_caches_response_cache:size", "", ["name"])
 response_cache_hits = Gauge("synapse_util_caches_response_cache:hits", "", ["name"])
@@ -52,67 +42,82 @@ response_cache_evicted = Gauge(
 response_cache_total = Gauge("synapse_util_caches_response_cache:total", "", ["name"])
 
 
-def register_cache(cache_type, cache_name, cache, collect_callback=None):
-    """Register a cache object for metric collection.
+@attr.s
+class CacheMetric(object):
+
+    _cache = attr.ib()
+    _cache_type = attr.ib(type=str)
+    _cache_name = attr.ib(type=str)
+    _collect_callback = attr.ib(type=Optional[Callable])
+
+    hits = attr.ib(default=0)
+    misses = attr.ib(default=0)
+    evicted_size = attr.ib(default=0)
+
+    def inc_hits(self):
+        self.hits += 1
+
+    def inc_misses(self):
+        self.misses += 1
+
+    def inc_evictions(self, size=1):
+        self.evicted_size += size
+
+    def describe(self):
+        return []
+
+    def collect(self):
+        try:
+            if self._cache_type == "response_cache":
+                response_cache_size.labels(self._cache_name).set(len(self._cache))
+                response_cache_hits.labels(self._cache_name).set(self.hits)
+                response_cache_evicted.labels(self._cache_name).set(self.evicted_size)
+                response_cache_total.labels(self._cache_name).set(
+                    self.hits + self.misses
+                )
+            else:
+                cache_size.labels(self._cache_name).set(len(self._cache))
+                cache_hits.labels(self._cache_name).set(self.hits)
+                cache_evicted.labels(self._cache_name).set(self.evicted_size)
+                cache_total.labels(self._cache_name).set(self.hits + self.misses)
+                if getattr(self._cache, "max_size", None):
+                    cache_max_size.labels(self._cache_name).set(self._cache.max_size)
+            if self._collect_callback:
+                self._collect_callback()
+        except Exception as e:
+            logger.warning("Error calculating metrics for %s: %s", self._cache_name, e)
+            raise
+
+
+def register_cache(
+    cache_type: str,
+    cache_name: str,
+    cache,
+    collect_callback: Optional[Callable] = None,
+    resizable: bool = True,
+    resize_callback: Optional[Callable] = None,
+) -> CacheMetric:
+    """Register a cache object for metric collection and resizing.
 
     Args:
-        cache_type (str):
-        cache_name (str): name of the cache
-        cache (object): cache itself
-        collect_callback (callable|None): if not None, a function which is called during
-            metric collection to update additional metrics.
+        cache_type
+        cache_name: name of the cache
+        cache: cache itself
+        collect_callback: If given, a function which is called during metric
+            collection to update additional metrics.
+        resizable: Whether this cache supports being resized.
+        resize_callback: A function which can be called to resize the cache.
 
     Returns:
         CacheMetric: an object which provides inc_{hits,misses,evictions} methods
     """
+    if resizable:
+        if not resize_callback:
+            resize_callback = getattr(cache, "set_cache_factor")
+        add_resizable_cache(cache_name, resize_callback)
 
-    # Check if the metric is already registered. Unregister it, if so.
-    # This usually happens during tests, as at runtime these caches are
-    # effectively singletons.
+    metric = CacheMetric(cache, cache_type, cache_name, collect_callback)
     metric_name = "cache_%s_%s" % (cache_type, cache_name)
-    if metric_name in collectors_by_name.keys():
-        REGISTRY.unregister(collectors_by_name[metric_name])
-
-    class CacheMetric(object):
-
-        hits = 0
-        misses = 0
-        evicted_size = 0
-
-        def inc_hits(self):
-            self.hits += 1
-
-        def inc_misses(self):
-            self.misses += 1
-
-        def inc_evictions(self, size=1):
-            self.evicted_size += size
-
-        def describe(self):
-            return []
-
-        def collect(self):
-            try:
-                if cache_type == "response_cache":
-                    response_cache_size.labels(cache_name).set(len(cache))
-                    response_cache_hits.labels(cache_name).set(self.hits)
-                    response_cache_evicted.labels(cache_name).set(self.evicted_size)
-                    response_cache_total.labels(cache_name).set(self.hits + self.misses)
-                else:
-                    cache_size.labels(cache_name).set(len(cache))
-                    cache_hits.labels(cache_name).set(self.hits)
-                    cache_evicted.labels(cache_name).set(self.evicted_size)
-                    cache_total.labels(cache_name).set(self.hits + self.misses)
-                if collect_callback:
-                    collect_callback()
-            except Exception as e:
-                logger.warn("Error calculating metrics for %s: %s", cache_name, e)
-                raise
-
-            yield GaugeMetricFamily("__unused", "")
-
-    metric = CacheMetric()
-    REGISTRY.register(metric)
     caches_by_name[cache_name] = cache
     collectors_by_name[metric_name] = metric
     return metric
@@ -147,9 +152,6 @@ def intern_string(string):
         return None
 
     try:
-        if six.PY2:
-            string = string.encode("ascii")
-
         return intern(string)
     except UnicodeEncodeError:
         return string
diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
index 43f66ec4be..cd48262420 100644
--- a/synapse/util/caches/descriptors.py
+++ b/synapse/util/caches/descriptors.py
@@ -13,22 +13,24 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 import functools
 import inspect
 import logging
 import threading
-from collections import namedtuple
+from typing import Any, Tuple, Union, cast
+from weakref import WeakValueDictionary
 
 from six import itervalues
 
 from prometheus_client import Gauge
+from typing_extensions import Protocol
 
 from twisted.internet import defer
 
 from synapse.logging.context import make_deferred_yieldable, preserve_fn
 from synapse.util import unwrapFirstError
 from synapse.util.async_helpers import ObservableDeferred
-from synapse.util.caches import get_cache_factor_for
 from synapse.util.caches.lrucache import LruCache
 from synapse.util.caches.treecache import TreeCache, iterate_tree_cache_entry
 
@@ -36,6 +38,20 @@ from . import register_cache
 
 logger = logging.getLogger(__name__)
 
+CacheKey = Union[Tuple, Any]
+
+
+class _CachedFunction(Protocol):
+    invalidate = None  # type: Any
+    invalidate_all = None  # type: Any
+    invalidate_many = None  # type: Any
+    prefill = None  # type: Any
+    cache = None  # type: Any
+    num_args = None  # type: Any
+
+    def __name__(self):
+        ...
+
 
 cache_pending_metric = Gauge(
     "synapse_util_caches_cache_pending",
@@ -65,7 +81,6 @@ class CacheEntry(object):
 class Cache(object):
     __slots__ = (
         "cache",
-        "max_entries",
         "name",
         "keylen",
         "thread",
@@ -73,7 +88,29 @@ class Cache(object):
         "_pending_deferred_cache",
     )
 
-    def __init__(self, name, max_entries=1000, keylen=1, tree=False, iterable=False):
+    def __init__(
+        self,
+        name: str,
+        max_entries: int = 1000,
+        keylen: int = 1,
+        tree: bool = False,
+        iterable: bool = False,
+        apply_cache_factor_from_config: bool = True,
+    ):
+        """
+        Args:
+            name: The name of the cache
+            max_entries: Maximum amount of entries that the cache will hold
+            keylen: The length of the tuple used as the cache key
+            tree: Use a TreeCache instead of a dict as the underlying cache type
+            iterable: If True, count each item in the cached object as an entry,
+                rather than each cached object
+            apply_cache_factor_from_config: Whether cache factors specified in the
+                config file affect `max_entries`
+
+        Returns:
+            Cache
+        """
         cache_type = TreeCache if tree else dict
         self._pending_deferred_cache = cache_type()
 
@@ -83,6 +120,7 @@ class Cache(object):
             cache_type=cache_type,
             size_callback=(lambda d: len(d)) if iterable else None,
             evicted_callback=self._on_evicted,
+            apply_cache_factor_from_config=apply_cache_factor_from_config,
         )
 
         self.name = name
@@ -95,6 +133,10 @@ class Cache(object):
             collect_callback=self._metrics_collection_callback,
         )
 
+    @property
+    def max_entries(self):
+        return self.cache.max_size
+
     def _on_evicted(self, evicted_count):
         self.metrics.inc_evictions(evicted_count)
 
@@ -245,7 +287,9 @@ class Cache(object):
 
 
 class _CacheDescriptorBase(object):
-    def __init__(self, orig, num_args, inlineCallbacks, cache_context=False):
+    def __init__(
+        self, orig: _CachedFunction, num_args, inlineCallbacks, cache_context=False
+    ):
         self.orig = orig
 
         if inlineCallbacks:
@@ -253,7 +297,7 @@ class _CacheDescriptorBase(object):
         else:
             self.function_to_call = orig
 
-        arg_spec = inspect.getargspec(orig)
+        arg_spec = inspect.getfullargspec(orig)
         all_args = arg_spec.args
 
         if "cache_context" in all_args:
@@ -352,13 +396,11 @@ class CacheDescriptor(_CacheDescriptorBase):
             cache_context=cache_context,
         )
 
-        max_entries = int(max_entries * get_cache_factor_for(orig.__name__))
-
         self.max_entries = max_entries
         self.tree = tree
         self.iterable = iterable
 
-    def __get__(self, obj, objtype=None):
+    def __get__(self, obj, owner):
         cache = Cache(
             name=self.orig.__name__,
             max_entries=self.max_entries,
@@ -404,7 +446,7 @@ class CacheDescriptor(_CacheDescriptorBase):
                 return tuple(get_cache_key_gen(args, kwargs))
 
         @functools.wraps(self.orig)
-        def wrapped(*args, **kwargs):
+        def _wrapped(*args, **kwargs):
             # If we're passed a cache_context then we'll want to call its invalidate()
             # whenever we are invalidated
             invalidate_callback = kwargs.pop("on_invalidate", None)
@@ -414,7 +456,7 @@ class CacheDescriptor(_CacheDescriptorBase):
             # Add our own `cache_context` to argument list if the wrapped function
             # has asked for one
             if self.add_cache_context:
-                kwargs["cache_context"] = _CacheContext(cache, cache_key)
+                kwargs["cache_context"] = _CacheContext.get_instance(cache, cache_key)
 
             try:
                 cached_result_d = cache.get(cache_key, callback=invalidate_callback)
@@ -422,7 +464,7 @@ class CacheDescriptor(_CacheDescriptorBase):
                 if isinstance(cached_result_d, ObservableDeferred):
                     observer = cached_result_d.observe()
                 else:
-                    observer = cached_result_d
+                    observer = defer.succeed(cached_result_d)
 
             except KeyError:
                 ret = defer.maybeDeferred(
@@ -440,6 +482,8 @@ class CacheDescriptor(_CacheDescriptorBase):
 
             return make_deferred_yieldable(observer)
 
+        wrapped = cast(_CachedFunction, _wrapped)
+
         if self.num_args == 1:
             wrapped.invalidate = lambda key: cache.invalidate(key[0])
             wrapped.prefill = lambda key, val: cache.prefill(key[0], val)
@@ -464,9 +508,8 @@ class CacheListDescriptor(_CacheDescriptorBase):
     Given a list of keys it looks in the cache to find any hits, then passes
     the list of missing keys to the wrapped function.
 
-    Once wrapped, the function returns either a Deferred which resolves to
-    the list of results, or (if all results were cached), just the list of
-    results.
+    Once wrapped, the function returns a Deferred which resolves to the list
+    of results.
     """
 
     def __init__(
@@ -600,21 +643,45 @@ class CacheListDescriptor(_CacheDescriptorBase):
                 )
                 return make_deferred_yieldable(d)
             else:
-                return results
+                return defer.succeed(results)
 
         obj.__dict__[self.orig.__name__] = wrapped
 
         return wrapped
 
 
-class _CacheContext(namedtuple("_CacheContext", ("cache", "key"))):
-    # We rely on _CacheContext implementing __eq__ and __hash__ sensibly,
-    # which namedtuple does for us (i.e. two _CacheContext are the same if
-    # their caches and keys match). This is important in particular to
-    # dedupe when we add callbacks to lru cache nodes, otherwise the number
-    # of callbacks would grow.
-    def invalidate(self):
-        self.cache.invalidate(self.key)
+class _CacheContext:
+    """Holds cache information from the cached function higher in the calling order.
+
+    Can be used to invalidate the higher level cache entry if something changes
+    on a lower level.
+    """
+
+    _cache_context_objects = (
+        WeakValueDictionary()
+    )  # type: WeakValueDictionary[Tuple[Cache, CacheKey], _CacheContext]
+
+    def __init__(self, cache, cache_key):  # type: (Cache, CacheKey) -> None
+        self._cache = cache
+        self._cache_key = cache_key
+
+    def invalidate(self):  # type: () -> None
+        """Invalidates the cache entry referred to by the context."""
+        self._cache.invalidate(self._cache_key)
+
+    @classmethod
+    def get_instance(cls, cache, cache_key):  # type: (Cache, CacheKey) -> _CacheContext
+        """Returns an instance constructed with the given arguments.
+
+        A new instance is only created if none already exists.
+        """
+
+        # We make sure there are no identical _CacheContext instances. This is
+        # important in particular to dedupe when we add callbacks to lru cache
+        # nodes, otherwise the number of callbacks would grow.
+        return cls._cache_context_objects.setdefault(
+            (cache, cache_key), cls(cache, cache_key)
+        )
 
 
 def cached(
diff --git a/synapse/util/caches/expiringcache.py b/synapse/util/caches/expiringcache.py
index cddf1ed515..2726b67b6d 100644
--- a/synapse/util/caches/expiringcache.py
+++ b/synapse/util/caches/expiringcache.py
@@ -18,6 +18,7 @@ from collections import OrderedDict
 
 from six import iteritems, itervalues
 
+from synapse.config import cache as cache_config
 from synapse.metrics.background_process_metrics import run_as_background_process
 from synapse.util.caches import register_cache
 
@@ -51,15 +52,16 @@ class ExpiringCache(object):
                 an item on access. Defaults to False.
             iterable (bool): If true, the size is calculated by summing the
                 sizes of all entries, rather than the number of entries.
-
         """
         self._cache_name = cache_name
 
+        self._original_max_size = max_len
+
+        self._max_size = int(max_len * cache_config.properties.default_factor_size)
+
         self._clock = clock
 
-        self._max_len = max_len
         self._expiry_ms = expiry_ms
-
         self._reset_expiry_on_get = reset_expiry_on_get
 
         self._cache = OrderedDict()
@@ -82,9 +84,11 @@ class ExpiringCache(object):
     def __setitem__(self, key, value):
         now = self._clock.time_msec()
         self._cache[key] = _CacheEntry(now, value)
+        self.evict()
 
+    def evict(self):
         # Evict if there are now too many items
-        while self._max_len and len(self) > self._max_len:
+        while self._max_size and len(self) > self._max_size:
             _key, value = self._cache.popitem(last=False)
             if self.iterable:
                 self.metrics.inc_evictions(len(value.value))
@@ -170,6 +174,23 @@ class ExpiringCache(object):
         else:
             return len(self._cache)
 
+    def set_cache_factor(self, factor: float) -> bool:
+        """
+        Set the cache factor for this individual cache.
+
+        This will trigger a resize if it changes, which may require evicting
+        items from the cache.
+
+        Returns:
+            bool: Whether the cache changed size or not.
+        """
+        new_size = int(self._original_max_size * factor)
+        if new_size != self._max_size:
+            self._max_size = new_size
+            self.evict()
+            return True
+        return False
+
 
 class _CacheEntry(object):
     __slots__ = ["time", "value"]
diff --git a/synapse/util/caches/lrucache.py b/synapse/util/caches/lrucache.py
index 1536cb64f3..df4ea5901d 100644
--- a/synapse/util/caches/lrucache.py
+++ b/synapse/util/caches/lrucache.py
@@ -13,10 +13,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import threading
 from functools import wraps
+from typing import Callable, Optional, Type, Union
 
+from synapse.config import cache as cache_config
 from synapse.util.caches.treecache import TreeCache
 
 
@@ -52,17 +53,18 @@ class LruCache(object):
 
     def __init__(
         self,
-        max_size,
-        keylen=1,
-        cache_type=dict,
-        size_callback=None,
-        evicted_callback=None,
+        max_size: int,
+        keylen: int = 1,
+        cache_type: Type[Union[dict, TreeCache]] = dict,
+        size_callback: Optional[Callable] = None,
+        evicted_callback: Optional[Callable] = None,
+        apply_cache_factor_from_config: bool = True,
     ):
         """
         Args:
-            max_size (int):
+            max_size: The maximum amount of entries the cache can hold
 
-            keylen (int):
+            keylen: The length of the tuple used as the cache key
 
             cache_type (type):
                 type of underlying cache to be used. Typically one of dict
@@ -73,9 +75,24 @@ class LruCache(object):
             evicted_callback (func(int)|None):
                 if not None, called on eviction with the size of the evicted
                 entry
+
+            apply_cache_factor_from_config (bool): If true, `max_size` will be
+                multiplied by a cache factor derived from the homeserver config
         """
         cache = cache_type()
         self.cache = cache  # Used for introspection.
+        self.apply_cache_factor_from_config = apply_cache_factor_from_config
+
+        # Save the original max size, and apply the default size factor.
+        self._original_max_size = max_size
+        # We previously didn't apply the cache factor here, and as such some caches were
+        # not affected by the global cache factor. Add an option here to disable applying
+        # the cache factor when a cache is created
+        if apply_cache_factor_from_config:
+            self.max_size = int(max_size * cache_config.properties.default_factor_size)
+        else:
+            self.max_size = int(max_size)
+
         list_root = _Node(None, None, None, None)
         list_root.next_node = list_root
         list_root.prev_node = list_root
@@ -83,7 +100,7 @@ class LruCache(object):
         lock = threading.Lock()
 
         def evict():
-            while cache_len() > max_size:
+            while cache_len() > self.max_size:
                 todelete = list_root.prev_node
                 evicted_len = delete_node(todelete)
                 cache.pop(todelete.key, None)
@@ -236,6 +253,7 @@ class LruCache(object):
             return key in cache
 
         self.sentinel = object()
+        self._on_resize = evict
         self.get = cache_get
         self.set = cache_set
         self.setdefault = cache_set_default
@@ -266,3 +284,23 @@ class LruCache(object):
 
     def __contains__(self, key):
         return self.contains(key)
+
+    def set_cache_factor(self, factor: float) -> bool:
+        """
+        Set the cache factor for this individual cache.
+
+        This will trigger a resize if it changes, which may require evicting
+        items from the cache.
+
+        Returns:
+            bool: Whether the cache changed size or not.
+        """
+        if not self.apply_cache_factor_from_config:
+            return False
+
+        new_size = int(self._original_max_size * factor)
+        if new_size != self.max_size:
+            self.max_size = new_size
+            self._on_resize()
+            return True
+        return False
diff --git a/synapse/util/caches/response_cache.py b/synapse/util/caches/response_cache.py
index 82d3eefe0e..a6c60888e5 100644
--- a/synapse/util/caches/response_cache.py
+++ b/synapse/util/caches/response_cache.py
@@ -38,7 +38,7 @@ class ResponseCache(object):
         self.timeout_sec = timeout_ms / 1000.0
 
         self._name = name
-        self._metrics = register_cache("response_cache", name, self)
+        self._metrics = register_cache("response_cache", name, self, resizable=False)
 
     def size(self):
         return len(self.pending_result_cache)
@@ -144,7 +144,7 @@ class ResponseCache(object):
         """
         result = self.get(key)
         if not result:
-            logger.info(
+            logger.debug(
                 "[%s]: no cached result for [%s], calculating new one", self._name, key
             )
             d = run_in_background(callback, *args, **kwargs)
diff --git a/synapse/util/caches/snapshot_cache.py b/synapse/util/caches/snapshot_cache.py
deleted file mode 100644
index 8318db8d2c..0000000000
--- a/synapse/util/caches/snapshot_cache.py
+++ /dev/null
@@ -1,94 +0,0 @@
-# -*- coding: utf-8 -*-
-# Copyright 2015, 2016 OpenMarket Ltd
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from synapse.util.async_helpers import ObservableDeferred
-
-
-class SnapshotCache(object):
-    """Cache for snapshots like the response of /initialSync.
-    The response of initialSync only has to be a recent snapshot of the
-    server state. It shouldn't matter to clients if it is a few minutes out
-    of date.
-
-    This caches a deferred response. Until the deferred completes it will be
-    returned from the cache. This means that if the client retries the request
-    while the response is still being computed, that original response will be
-    used rather than trying to compute a new response.
-
-    Once the deferred completes it will removed from the cache after 5 minutes.
-    We delay removing it from the cache because a client retrying its request
-    could race with us finishing computing the response.
-
-    Rather than tracking precisely how long something has been in the cache we
-    keep two generations of completed responses. Every 5 minutes discard the
-    old generation, move the new generation to the old generation, and set the
-    new generation to be empty. This means that a result will be in the cache
-    somewhere between 5 and 10 minutes.
-    """
-
-    DURATION_MS = 5 * 60 * 1000  # Cache results for 5 minutes.
-
-    def __init__(self):
-        self.pending_result_cache = {}  # Request that haven't finished yet.
-        self.prev_result_cache = {}  # The older requests that have finished.
-        self.next_result_cache = {}  # The newer requests that have finished.
-        self.time_last_rotated_ms = 0
-
-    def rotate(self, time_now_ms):
-        # Rotate once if the cache duration has passed since the last rotation.
-        if time_now_ms - self.time_last_rotated_ms >= self.DURATION_MS:
-            self.prev_result_cache = self.next_result_cache
-            self.next_result_cache = {}
-            self.time_last_rotated_ms += self.DURATION_MS
-
-        # Rotate again if the cache duration has passed twice since the last
-        # rotation.
-        if time_now_ms - self.time_last_rotated_ms >= self.DURATION_MS:
-            self.prev_result_cache = self.next_result_cache
-            self.next_result_cache = {}
-            self.time_last_rotated_ms = time_now_ms
-
-    def get(self, time_now_ms, key):
-        self.rotate(time_now_ms)
-        # This cache is intended to deduplicate requests, so we expect it to be
-        # missed most of the time. So we just lookup the key in all of the
-        # dictionaries rather than trying to short circuit the lookup if the
-        # key is found.
-        result = self.prev_result_cache.get(key)
-        result = self.next_result_cache.get(key, result)
-        result = self.pending_result_cache.get(key, result)
-        if result is not None:
-            return result.observe()
-        else:
-            return None
-
-    def set(self, time_now_ms, key, deferred):
-        self.rotate(time_now_ms)
-
-        result = ObservableDeferred(deferred)
-
-        self.pending_result_cache[key] = result
-
-        def shuffle_along(r):
-            # When the deferred completes we shuffle it along to the first
-            # generation of the result cache. So that it will eventually
-            # expire from the rotation of that cache.
-            self.next_result_cache[key] = result
-            self.pending_result_cache.pop(key, None)
-            return r
-
-        result.addBoth(shuffle_along)
-
-        return result.observe()
diff --git a/synapse/util/caches/stream_change_cache.py b/synapse/util/caches/stream_change_cache.py
index 235f64049c..2a161bf244 100644
--- a/synapse/util/caches/stream_change_cache.py
+++ b/synapse/util/caches/stream_change_cache.py
@@ -14,17 +14,23 @@
 # limitations under the License.
 
 import logging
+import math
+from typing import Dict, FrozenSet, List, Mapping, Optional, Set, Union
 
 from six import integer_types
 
 from sortedcontainers import SortedDict
 
+from synapse.types import Collection
 from synapse.util import caches
 
 logger = logging.getLogger(__name__)
 
+# for now, assume all entities in the cache are strings
+EntityType = str
 
-class StreamChangeCache(object):
+
+class StreamChangeCache:
     """Keeps track of the stream positions of the latest change in a set of entities.
 
     Typically the entity will be a room or user id.
@@ -34,19 +40,52 @@ class StreamChangeCache(object):
     old then the cache will simply return all given entities.
     """
 
-    def __init__(self, name, current_stream_pos, max_size=10000, prefilled_cache=None):
-        self._max_size = int(max_size * caches.CACHE_SIZE_FACTOR)
-        self._entity_to_key = {}
-        self._cache = SortedDict()
+    def __init__(
+        self,
+        name: str,
+        current_stream_pos: int,
+        max_size=10000,
+        prefilled_cache: Optional[Mapping[EntityType, int]] = None,
+    ):
+        self._original_max_size = max_size
+        self._max_size = math.floor(max_size)
+        self._entity_to_key = {}  # type: Dict[EntityType, int]
+
+        # map from stream id to the a set of entities which changed at that stream id.
+        self._cache = SortedDict()  # type: SortedDict[int, Set[EntityType]]
+
+        # the earliest stream_pos for which we can reliably answer
+        # get_all_entities_changed. In other words, one less than the earliest
+        # stream_pos for which we know _cache is valid.
+        #
         self._earliest_known_stream_pos = current_stream_pos
         self.name = name
-        self.metrics = caches.register_cache("cache", self.name, self._cache)
+        self.metrics = caches.register_cache(
+            "cache", self.name, self._cache, resize_callback=self.set_cache_factor
+        )
 
         if prefilled_cache:
             for entity, stream_pos in prefilled_cache.items():
                 self.entity_has_changed(entity, stream_pos)
 
-    def has_entity_changed(self, entity, stream_pos):
+    def set_cache_factor(self, factor: float) -> bool:
+        """
+        Set the cache factor for this individual cache.
+
+        This will trigger a resize if it changes, which may require evicting
+        items from the cache.
+
+        Returns:
+            bool: Whether the cache changed size or not.
+        """
+        new_size = math.floor(self._original_max_size * factor)
+        if new_size != self._max_size:
+            self.max_size = new_size
+            self._evict()
+            return True
+        return False
+
+    def has_entity_changed(self, entity: EntityType, stream_pos: int) -> bool:
         """Returns True if the entity may have been updated since stream_pos
         """
         assert type(stream_pos) in integer_types
@@ -67,22 +106,27 @@ class StreamChangeCache(object):
         self.metrics.inc_hits()
         return False
 
-    def get_entities_changed(self, entities, stream_pos):
+    def get_entities_changed(
+        self, entities: Collection[EntityType], stream_pos: int
+    ) -> Union[Set[EntityType], FrozenSet[EntityType]]:
         """
         Returns subset of entities that have had new things since the given
         position.  Entities unknown to the cache will be returned.  If the
         position is too old it will just return the given list.
         """
-        assert type(stream_pos) is int
-
-        if stream_pos >= self._earliest_known_stream_pos:
-            changed_entities = {
-                self._cache[k]
-                for k in self._cache.islice(start=self._cache.bisect_right(stream_pos))
-            }
-
-            result = changed_entities.intersection(entities)
-
+        changed_entities = self.get_all_entities_changed(stream_pos)
+        if changed_entities is not None:
+            # We now do an intersection, trying to do so in the most efficient
+            # way possible (some of these sets are *large*). First check in the
+            # given iterable is already set that we can reuse, otherwise we
+            # create a set of the *smallest* of the two iterables and call
+            # `intersection(..)` on it (this can be twice as fast as the reverse).
+            if isinstance(entities, (set, frozenset)):
+                result = entities.intersection(changed_entities)
+            elif len(changed_entities) < len(entities):
+                result = set(changed_entities).intersection(entities)
+            else:
+                result = set(entities).intersection(changed_entities)
             self.metrics.inc_hits()
         else:
             result = set(entities)
@@ -90,13 +134,13 @@ class StreamChangeCache(object):
 
         return result
 
-    def has_any_entity_changed(self, stream_pos):
+    def has_any_entity_changed(self, stream_pos: int) -> bool:
         """Returns if any entity has changed
         """
         assert type(stream_pos) is int
 
         if not self._cache:
-            # If we have no cache, nothing can have changed.
+            # If the cache is empty, nothing can have changed.
             return False
 
         if stream_pos >= self._earliest_known_stream_pos:
@@ -106,42 +150,66 @@ class StreamChangeCache(object):
             self.metrics.inc_misses()
             return True
 
-    def get_all_entities_changed(self, stream_pos):
-        """Returns all entites that have had new things since the given
+    def get_all_entities_changed(self, stream_pos: int) -> Optional[List[EntityType]]:
+        """Returns all entities that have had new things since the given
         position. If the position is too old it will return None.
+
+        Returns the entities in the order that they were changed.
         """
         assert type(stream_pos) is int
 
-        if stream_pos >= self._earliest_known_stream_pos:
-            return [
-                self._cache[k]
-                for k in self._cache.islice(start=self._cache.bisect_right(stream_pos))
-            ]
-        else:
+        if stream_pos < self._earliest_known_stream_pos:
             return None
 
-    def entity_has_changed(self, entity, stream_pos):
+        changed_entities = []  # type: List[EntityType]
+
+        for k in self._cache.islice(start=self._cache.bisect_right(stream_pos)):
+            changed_entities.extend(self._cache[k])
+        return changed_entities
+
+    def entity_has_changed(self, entity: EntityType, stream_pos: int) -> None:
         """Informs the cache that the entity has been changed at the given
         position.
         """
         assert type(stream_pos) is int
 
-        if stream_pos > self._earliest_known_stream_pos:
-            old_pos = self._entity_to_key.get(entity, None)
-            if old_pos is not None:
-                stream_pos = max(stream_pos, old_pos)
-                self._cache.pop(old_pos, None)
-            self._cache[stream_pos] = entity
-            self._entity_to_key[entity] = stream_pos
-
-            while len(self._cache) > self._max_size:
-                k, r = self._cache.popitem(0)
-                self._earliest_known_stream_pos = max(
-                    k, self._earliest_known_stream_pos
-                )
-                self._entity_to_key.pop(r, None)
-
-    def get_max_pos_of_last_change(self, entity):
+        if stream_pos <= self._earliest_known_stream_pos:
+            return
+
+        old_pos = self._entity_to_key.get(entity, None)
+        if old_pos is not None:
+            if old_pos >= stream_pos:
+                # nothing to do
+                return
+            e = self._cache[old_pos]
+            e.remove(entity)
+            if not e:
+                # cache at this point is now empty
+                del self._cache[old_pos]
+
+        e1 = self._cache.get(stream_pos)
+        if e1 is None:
+            e1 = self._cache[stream_pos] = set()
+        e1.add(entity)
+        self._entity_to_key[entity] = stream_pos
+        self._evict()
+
+        # if the cache is too big, remove entries
+        while len(self._cache) > self._max_size:
+            k, r = self._cache.popitem(0)
+            self._earliest_known_stream_pos = max(k, self._earliest_known_stream_pos)
+            for entity in r:
+                del self._entity_to_key[entity]
+
+    def _evict(self):
+        while len(self._cache) > self._max_size:
+            k, r = self._cache.popitem(0)
+            self._earliest_known_stream_pos = max(k, self._earliest_known_stream_pos)
+            for entity in r:
+                self._entity_to_key.pop(entity, None)
+
+    def get_max_pos_of_last_change(self, entity: EntityType) -> int:
+
         """Returns an upper bound of the stream id of the last change to an
         entity.
         """
diff --git a/synapse/util/caches/treecache.py b/synapse/util/caches/treecache.py
index 9a72218d85..2ea4e4e911 100644
--- a/synapse/util/caches/treecache.py
+++ b/synapse/util/caches/treecache.py
@@ -1,3 +1,5 @@
+from typing import Dict
+
 from six import itervalues
 
 SENTINEL = object()
@@ -12,7 +14,7 @@ class TreeCache(object):
 
     def __init__(self):
         self.size = 0
-        self.root = {}
+        self.root = {}  # type: Dict
 
     def __setitem__(self, key, value):
         return self.set(key, value)
diff --git a/synapse/util/caches/ttlcache.py b/synapse/util/caches/ttlcache.py
index 99646c7cf0..6437aa907e 100644
--- a/synapse/util/caches/ttlcache.py
+++ b/synapse/util/caches/ttlcache.py
@@ -38,7 +38,7 @@ class TTLCache(object):
 
         self._timer = timer
 
-        self._metrics = register_cache("ttl", cache_name, self)
+        self._metrics = register_cache("ttl", cache_name, self, resizable=False)
 
     def set(self, key, value, ttl):
         """Add/update an entry in the cache