diff --git a/synapse/replication/slave/storage/client_ips.py b/synapse/replication/slave/storage/client_ips.py index 40ea78a35..4b0ea0cc0 100644 --- a/synapse/replication/slave/storage/client_ips.py +++ b/synapse/replication/slave/storage/client_ips.py @@ -15,7 +15,7 @@ from synapse.storage.database import DatabasePool from synapse.storage.databases.main.client_ips import LAST_SEEN_GRANULARITY -from synapse.util.caches.descriptors import DeferredCache +from synapse.util.caches.deferred_cache import DeferredCache from ._base import BaseSlavedStore diff --git a/synapse/storage/databases/main/client_ips.py b/synapse/storage/databases/main/client_ips.py index ad32701d3..9e66e6648 100644 --- a/synapse/storage/databases/main/client_ips.py +++ b/synapse/storage/databases/main/client_ips.py @@ -19,7 +19,7 @@ from typing import Dict, Optional, Tuple from synapse.metrics.background_process_metrics import wrap_as_background_process from synapse.storage._base import SQLBaseStore from synapse.storage.database import DatabasePool, make_tuple_comparison_clause -from synapse.util.caches.descriptors import DeferredCache +from synapse.util.caches.deferred_cache import DeferredCache logger = logging.getLogger(__name__) diff --git a/synapse/storage/databases/main/devices.py b/synapse/storage/databases/main/devices.py index d903155e8..e662a20d2 100644 --- a/synapse/storage/databases/main/devices.py +++ b/synapse/storage/databases/main/devices.py @@ -34,7 +34,8 @@ from synapse.storage.database import ( ) from synapse.types import Collection, JsonDict, get_verify_key_from_cross_signing_key from synapse.util import json_decoder, json_encoder -from synapse.util.caches.descriptors import DeferredCache, cached, cachedList +from synapse.util.caches.deferred_cache import DeferredCache +from synapse.util.caches.descriptors import cached, cachedList from synapse.util.iterutils import batch_iter from synapse.util.stringutils import shortstr diff --git a/synapse/storage/databases/main/events_worker.py b/synapse/storage/databases/main/events_worker.py index be7f60f2e..ff150f0be 100644 --- a/synapse/storage/databases/main/events_worker.py +++ b/synapse/storage/databases/main/events_worker.py @@ -42,7 +42,8 @@ from synapse.storage.database import DatabasePool from synapse.storage.engines import PostgresEngine from synapse.storage.util.id_generators import MultiWriterIdGenerator, StreamIdGenerator from synapse.types import Collection, get_domain_from_id -from synapse.util.caches.descriptors import DeferredCache, cached +from synapse.util.caches.deferred_cache import DeferredCache +from synapse.util.caches.descriptors import cached from synapse.util.iterutils import batch_iter from synapse.util.metrics import Measure diff --git a/synapse/util/caches/deferred_cache.py b/synapse/util/caches/deferred_cache.py new file mode 100644 index 000000000..f728cd2cf --- /dev/null +++ b/synapse/util/caches/deferred_cache.py @@ -0,0 +1,292 @@ +# -*- coding: utf-8 -*- +# Copyright 2015, 2016 OpenMarket Ltd +# Copyright 2018 New Vector Ltd +# Copyright 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. +# 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. + +import enum +import threading +from typing import Callable, Generic, Iterable, MutableMapping, Optional, TypeVar, cast + +from prometheus_client import Gauge + +from twisted.internet import defer + +from synapse.util.async_helpers import ObservableDeferred +from synapse.util.caches import register_cache +from synapse.util.caches.lrucache import LruCache +from synapse.util.caches.treecache import TreeCache, iterate_tree_cache_entry + +cache_pending_metric = Gauge( + "synapse_util_caches_cache_pending", + "Number of lookups currently pending for this cache", + ["name"], +) + + +KT = TypeVar("KT") +VT = TypeVar("VT") + + +class _Sentinel(enum.Enum): + # defining a sentinel in this way allows mypy to correctly handle the + # type of a dictionary lookup. + sentinel = object() + + +class DeferredCache(Generic[KT, VT]): + """Wraps an LruCache, adding support for Deferred results. + + It expects that each entry added with set() will be a Deferred; likewise get() + may return an ObservableDeferred. + """ + + __slots__ = ( + "cache", + "name", + "keylen", + "thread", + "metrics", + "_pending_deferred_cache", + ) + + 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. Ignored unless + `tree` is True. + 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` + """ + cache_type = TreeCache if tree else dict + + # _pending_deferred_cache maps from the key value to a `CacheEntry` object. + self._pending_deferred_cache = ( + cache_type() + ) # type: MutableMapping[KT, CacheEntry] + + # cache is used for completed results and maps to the result itself, rather than + # a Deferred. + 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, + apply_cache_factor_from_config=apply_cache_factor_from_config, + ) + + self.name = name + self.keylen = keylen + self.thread = None # type: Optional[threading.Thread] + self.metrics = register_cache( + "cache", + name, + self.cache, + 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) + + def _metrics_collection_callback(self): + cache_pending_metric.labels(self.name).set(len(self._pending_deferred_cache)) + + def check_thread(self): + expected_thread = self.thread + if expected_thread is None: + self.thread = threading.current_thread() + else: + if expected_thread is not threading.current_thread(): + raise ValueError( + "Cache objects can only be accessed from the main thread" + ) + + def get( + self, + key: KT, + default=_Sentinel.sentinel, + callback: Optional[Callable[[], None]] = None, + update_metrics: bool = True, + ): + """Looks the key up in the caches. + + Args: + key(tuple) + default: What is returned if key is not in the caches. If not + specified then function throws KeyError instead + callback(fn): Gets called when the entry in the cache is invalidated + update_metrics (bool): whether to update the cache hit rate metrics + + Returns: + Either an ObservableDeferred or the result itself + """ + callbacks = [callback] if callback else [] + val = self._pending_deferred_cache.get(key, _Sentinel.sentinel) + if val is not _Sentinel.sentinel: + val.callbacks.update(callbacks) + if update_metrics: + self.metrics.inc_hits() + return val.deferred + + val = self.cache.get(key, _Sentinel.sentinel, callbacks=callbacks) + if val is not _Sentinel.sentinel: + self.metrics.inc_hits() + return val + + if update_metrics: + self.metrics.inc_misses() + + if default is _Sentinel.sentinel: + raise KeyError() + else: + return default + + def set( + self, + key: KT, + value: defer.Deferred, + callback: Optional[Callable[[], None]] = None, + ) -> ObservableDeferred: + if not isinstance(value, defer.Deferred): + raise TypeError("not a Deferred") + + callbacks = [callback] if callback else [] + self.check_thread() + observable = ObservableDeferred(value, consumeErrors=True) + observer = observable.observe() + entry = CacheEntry(deferred=observable, callbacks=callbacks) + + existing_entry = self._pending_deferred_cache.pop(key, None) + if existing_entry: + existing_entry.invalidate() + + self._pending_deferred_cache[key] = entry + + def compare_and_pop(): + """Check if our entry is still the one in _pending_deferred_cache, and + if so, pop it. + + Returns true if the entries matched. + """ + existing_entry = self._pending_deferred_cache.pop(key, None) + if existing_entry is entry: + return True + + # oops, the _pending_deferred_cache has been updated since + # we started our query, so we are out of date. + # + # Better put back whatever we took out. (We do it this way + # round, rather than peeking into the _pending_deferred_cache + # and then removing on a match, to make the common case faster) + if existing_entry is not None: + self._pending_deferred_cache[key] = existing_entry + + return False + + def cb(result): + if compare_and_pop(): + self.cache.set(key, result, entry.callbacks) + else: + # we're not going to put this entry into the cache, so need + # to make sure that the invalidation callbacks are called. + # That was probably done when _pending_deferred_cache was + # updated, but it's possible that `set` was called without + # `invalidate` being previously called, in which case it may + # not have been. Either way, let's double-check now. + entry.invalidate() + + def eb(_fail): + compare_and_pop() + entry.invalidate() + + # once the deferred completes, we can move the entry from the + # _pending_deferred_cache to the real cache. + # + observer.addCallbacks(cb, eb) + return observable + + def prefill(self, key: KT, value: VT, callback: Callable[[], None] = None): + callbacks = [callback] if callback else [] + self.cache.set(key, value, callbacks=callbacks) + + def invalidate(self, key): + self.check_thread() + self.cache.pop(key, None) + + # if we have a pending lookup for this key, remove it from the + # _pending_deferred_cache, which will (a) stop it being returned + # for future queries and (b) stop it being persisted as a proper entry + # in self.cache. + entry = self._pending_deferred_cache.pop(key, None) + + # run the invalidation callbacks now, rather than waiting for the + # deferred to resolve. + if entry: + entry.invalidate() + + def invalidate_many(self, key: KT): + self.check_thread() + if not isinstance(key, tuple): + raise TypeError("The cache key must be a tuple not %r" % (type(key),)) + self.cache.del_multi(key) + + # if we have a pending lookup for this key, remove it from the + # _pending_deferred_cache, as above + entry_dict = self._pending_deferred_cache.pop(cast(KT, key), None) + if entry_dict is not None: + for entry in iterate_tree_cache_entry(entry_dict): + entry.invalidate() + + def invalidate_all(self): + self.check_thread() + self.cache.clear() + for entry in self._pending_deferred_cache.values(): + entry.invalidate() + self._pending_deferred_cache.clear() + + +class CacheEntry: + __slots__ = ["deferred", "callbacks", "invalidated"] + + def __init__( + self, deferred: ObservableDeferred, callbacks: Iterable[Callable[[], None]] + ): + self.deferred = deferred + self.callbacks = set(callbacks) + self.invalidated = False + + def invalidate(self): + if not self.invalidated: + self.invalidated = True + for callback in self.callbacks: + callback() + self.callbacks.clear() diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py index 7c9fe199b..1f4388680 100644 --- a/synapse/util/caches/descriptors.py +++ b/synapse/util/caches/descriptors.py @@ -13,44 +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 enum import functools import inspect import logging -import threading -from typing import ( - Any, - Callable, - Generic, - Iterable, - MutableMapping, - Optional, - Tuple, - TypeVar, - Union, - cast, -) +from typing import Any, Callable, Generic, Optional, Tuple, TypeVar, Union, cast from weakref import WeakValueDictionary -from prometheus_client import Gauge - 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.lrucache import LruCache -from synapse.util.caches.treecache import TreeCache, iterate_tree_cache_entry - -from . import register_cache +from synapse.util.caches.deferred_cache import DeferredCache logger = logging.getLogger(__name__) CacheKey = Union[Tuple, Any] F = TypeVar("F", bound=Callable[..., Any]) -KT = TypeVar("KT") -VT = TypeVar("VT") class _CachedFunction(Generic[F]): @@ -68,266 +48,6 @@ class _CachedFunction(Generic[F]): __call__ = None # type: F -cache_pending_metric = Gauge( - "synapse_util_caches_cache_pending", - "Number of lookups currently pending for this cache", - ["name"], -) - - -class _Sentinel(enum.Enum): - # defining a sentinel in this way allows mypy to correctly handle the - # type of a dictionary lookup. - sentinel = object() - - -class CacheEntry: - __slots__ = ["deferred", "callbacks", "invalidated"] - - def __init__( - self, deferred: ObservableDeferred, callbacks: Iterable[Callable[[], None]] - ): - self.deferred = deferred - self.callbacks = set(callbacks) - self.invalidated = False - - def invalidate(self): - if not self.invalidated: - self.invalidated = True - for callback in self.callbacks: - callback() - self.callbacks.clear() - - -class DeferredCache(Generic[KT, VT]): - """Wraps an LruCache, adding support for Deferred results. - - It expects that each entry added with set() will be a Deferred; likewise get() - may return an ObservableDeferred. - """ - - __slots__ = ( - "cache", - "name", - "keylen", - "thread", - "metrics", - "_pending_deferred_cache", - ) - - 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. Ignored unless - `tree` is True. - 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` - """ - cache_type = TreeCache if tree else dict - - # _pending_deferred_cache maps from the key value to a `CacheEntry` object. - self._pending_deferred_cache = ( - cache_type() - ) # type: MutableMapping[KT, CacheEntry] - - # cache is used for completed results and maps to the result itself, rather than - # a Deferred. - 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, - apply_cache_factor_from_config=apply_cache_factor_from_config, - ) - - self.name = name - self.keylen = keylen - self.thread = None # type: Optional[threading.Thread] - self.metrics = register_cache( - "cache", - name, - self.cache, - 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) - - def _metrics_collection_callback(self): - cache_pending_metric.labels(self.name).set(len(self._pending_deferred_cache)) - - def check_thread(self): - expected_thread = self.thread - if expected_thread is None: - self.thread = threading.current_thread() - else: - if expected_thread is not threading.current_thread(): - raise ValueError( - "Cache objects can only be accessed from the main thread" - ) - - def get( - self, - key: KT, - default=_Sentinel.sentinel, - callback: Optional[Callable[[], None]] = None, - update_metrics: bool = True, - ): - """Looks the key up in the caches. - - Args: - key(tuple) - default: What is returned if key is not in the caches. If not - specified then function throws KeyError instead - callback(fn): Gets called when the entry in the cache is invalidated - update_metrics (bool): whether to update the cache hit rate metrics - - Returns: - Either an ObservableDeferred or the result itself - """ - callbacks = [callback] if callback else [] - val = self._pending_deferred_cache.get(key, _Sentinel.sentinel) - if val is not _Sentinel.sentinel: - val.callbacks.update(callbacks) - if update_metrics: - self.metrics.inc_hits() - return val.deferred - - val = self.cache.get(key, _Sentinel.sentinel, callbacks=callbacks) - if val is not _Sentinel.sentinel: - self.metrics.inc_hits() - return val - - if update_metrics: - self.metrics.inc_misses() - - if default is _Sentinel.sentinel: - raise KeyError() - else: - return default - - def set( - self, - key: KT, - value: defer.Deferred, - callback: Optional[Callable[[], None]] = None, - ) -> ObservableDeferred: - if not isinstance(value, defer.Deferred): - raise TypeError("not a Deferred") - - callbacks = [callback] if callback else [] - self.check_thread() - observable = ObservableDeferred(value, consumeErrors=True) - observer = observable.observe() - entry = CacheEntry(deferred=observable, callbacks=callbacks) - - existing_entry = self._pending_deferred_cache.pop(key, None) - if existing_entry: - existing_entry.invalidate() - - self._pending_deferred_cache[key] = entry - - def compare_and_pop(): - """Check if our entry is still the one in _pending_deferred_cache, and - if so, pop it. - - Returns true if the entries matched. - """ - existing_entry = self._pending_deferred_cache.pop(key, None) - if existing_entry is entry: - return True - - # oops, the _pending_deferred_cache has been updated since - # we started our query, so we are out of date. - # - # Better put back whatever we took out. (We do it this way - # round, rather than peeking into the _pending_deferred_cache - # and then removing on a match, to make the common case faster) - if existing_entry is not None: - self._pending_deferred_cache[key] = existing_entry - - return False - - def cb(result): - if compare_and_pop(): - self.cache.set(key, result, entry.callbacks) - else: - # we're not going to put this entry into the cache, so need - # to make sure that the invalidation callbacks are called. - # That was probably done when _pending_deferred_cache was - # updated, but it's possible that `set` was called without - # `invalidate` being previously called, in which case it may - # not have been. Either way, let's double-check now. - entry.invalidate() - - def eb(_fail): - compare_and_pop() - entry.invalidate() - - # once the deferred completes, we can move the entry from the - # _pending_deferred_cache to the real cache. - # - observer.addCallbacks(cb, eb) - return observable - - def prefill(self, key: KT, value: VT, callback: Callable[[], None] = None): - callbacks = [callback] if callback else [] - self.cache.set(key, value, callbacks=callbacks) - - def invalidate(self, key): - self.check_thread() - self.cache.pop(key, None) - - # if we have a pending lookup for this key, remove it from the - # _pending_deferred_cache, which will (a) stop it being returned - # for future queries and (b) stop it being persisted as a proper entry - # in self.cache. - entry = self._pending_deferred_cache.pop(key, None) - - # run the invalidation callbacks now, rather than waiting for the - # deferred to resolve. - if entry: - entry.invalidate() - - def invalidate_many(self, key: KT): - self.check_thread() - if not isinstance(key, tuple): - raise TypeError("The cache key must be a tuple not %r" % (type(key),)) - self.cache.del_multi(key) - - # if we have a pending lookup for this key, remove it from the - # _pending_deferred_cache, as above - entry_dict = self._pending_deferred_cache.pop(cast(KT, key), None) - if entry_dict is not None: - for entry in iterate_tree_cache_entry(entry_dict): - entry.invalidate() - - def invalidate_all(self): - self.check_thread() - self.cache.clear() - for entry in self._pending_deferred_cache.values(): - entry.invalidate() - self._pending_deferred_cache.clear() - - class _CacheDescriptorBase: def __init__(self, orig: _CachedFunction, num_args, cache_context=False): self.orig = orig diff --git a/tests/storage/test__base.py b/tests/storage/test__base.py index 00adcab7b..2598dbe0a 100644 --- a/tests/storage/test__base.py +++ b/tests/storage/test__base.py @@ -20,7 +20,8 @@ from mock import Mock from twisted.internet import defer from synapse.util.async_helpers import ObservableDeferred -from synapse.util.caches.descriptors import DeferredCache, cached +from synapse.util.caches.deferred_cache import DeferredCache +from synapse.util.caches.descriptors import cached from tests import unittest diff --git a/tests/test_metrics.py b/tests/test_metrics.py index 1c03a52f7..759e4cd04 100644 --- a/tests/test_metrics.py +++ b/tests/test_metrics.py @@ -15,7 +15,7 @@ # limitations under the License. from synapse.metrics import REGISTRY, InFlightGauge, generate_latest -from synapse.util.caches.descriptors import DeferredCache +from synapse.util.caches.deferred_cache import DeferredCache from tests import unittest diff --git a/tests/util/caches/test_deferred_cache.py b/tests/util/caches/test_deferred_cache.py new file mode 100644 index 000000000..9b6acdfc4 --- /dev/null +++ b/tests/util/caches/test_deferred_cache.py @@ -0,0 +1,64 @@ +# -*- coding: utf-8 -*- +# Copyright 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. +# 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. + +import unittest +from functools import partial + +from twisted.internet import defer + +import synapse.util.caches.deferred_cache + + +class DeferredCacheTestCase(unittest.TestCase): + def test_invalidate_all(self): + cache = synapse.util.caches.deferred_cache.DeferredCache("testcache") + + callback_record = [False, False] + + def record_callback(idx): + callback_record[idx] = True + + # add a couple of pending entries + d1 = defer.Deferred() + cache.set("key1", d1, partial(record_callback, 0)) + + d2 = defer.Deferred() + cache.set("key2", d2, partial(record_callback, 1)) + + # lookup should return observable deferreds + self.assertFalse(cache.get("key1").has_called()) + self.assertFalse(cache.get("key2").has_called()) + + # let one of the lookups complete + d2.callback("result2") + + # for now at least, the cache will return real results rather than an + # observabledeferred + self.assertEqual(cache.get("key2"), "result2") + + # now do the invalidation + cache.invalidate_all() + + # lookup should return none + self.assertIsNone(cache.get("key1", None)) + self.assertIsNone(cache.get("key2", None)) + + # both callbacks should have been callbacked + self.assertTrue(callback_record[0], "Invalidation callback for key1 not called") + self.assertTrue(callback_record[1], "Invalidation callback for key2 not called") + + # letting the other lookup complete should do nothing + d1.callback("result1") + self.assertIsNone(cache.get("key1", None)) diff --git a/tests/util/caches/test_descriptors.py b/tests/util/caches/test_descriptors.py index bd870b4a3..3d1f96086 100644 --- a/tests/util/caches/test_descriptors.py +++ b/tests/util/caches/test_descriptors.py @@ -14,7 +14,6 @@ # See the License for the specific language governing permissions and # limitations under the License. import logging -from functools import partial import mock @@ -42,49 +41,6 @@ def run_on_reactor(): return make_deferred_yieldable(d) -class DeferredCacheTestCase(unittest.TestCase): - def test_invalidate_all(self): - cache = descriptors.DeferredCache("testcache") - - callback_record = [False, False] - - def record_callback(idx): - callback_record[idx] = True - - # add a couple of pending entries - d1 = defer.Deferred() - cache.set("key1", d1, partial(record_callback, 0)) - - d2 = defer.Deferred() - cache.set("key2", d2, partial(record_callback, 1)) - - # lookup should return observable deferreds - self.assertFalse(cache.get("key1").has_called()) - self.assertFalse(cache.get("key2").has_called()) - - # let one of the lookups complete - d2.callback("result2") - - # for now at least, the cache will return real results rather than an - # observabledeferred - self.assertEqual(cache.get("key2"), "result2") - - # now do the invalidation - cache.invalidate_all() - - # lookup should return none - self.assertIsNone(cache.get("key1", None)) - self.assertIsNone(cache.get("key2", None)) - - # both callbacks should have been callbacked - self.assertTrue(callback_record[0], "Invalidation callback for key1 not called") - self.assertTrue(callback_record[1], "Invalidation callback for key2 not called") - - # letting the other lookup complete should do nothing - d1.callback("result1") - self.assertIsNone(cache.get("key1", None)) - - class DescriptorTestCase(unittest.TestCase): @defer.inlineCallbacks def test_cache(self):