synapse-old/synapse/util/caches/descriptors.py

640 lines
22 KiB
Python
Raw Normal View History

# -*- coding: utf-8 -*-
2016-01-06 21:26:29 -07:00
# 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.
import logging
from synapse.util.async import ObservableDeferred
from synapse.util import unwrapFirstError, logcontext
2017-07-04 02:56:44 -06:00
from synapse.util.caches import CACHE_SIZE_FACTOR
from synapse.util.caches.lrucache import LruCache
from synapse.util.caches.treecache import TreeCache, iterate_tree_cache_entry
2017-05-04 07:18:46 -06:00
from synapse.util.stringutils import to_ascii
from . import register_cache
from twisted.internet import defer
from collections import namedtuple
import functools
import inspect
import threading
logger = logging.getLogger(__name__)
_CacheSentinel = object()
class CacheEntry(object):
__slots__ = [
"deferred", "sequence", "callbacks", "invalidated"
]
def __init__(self, deferred, sequence, callbacks):
self.deferred = deferred
self.sequence = sequence
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 Cache(object):
__slots__ = (
"cache",
"max_entries",
"name",
"keylen",
"sequence",
"thread",
"metrics",
"_pending_deferred_cache",
)
def __init__(self, name, max_entries=1000, keylen=1, tree=False, iterable=False):
2016-08-19 04:59:29 -06:00
cache_type = TreeCache if tree else dict
self._pending_deferred_cache = cache_type()
2016-08-19 04:59:29 -06:00
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,
2016-08-19 04:59:29 -06:00
)
self.name = name
self.keylen = keylen
self.sequence = 0
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:
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, default=_CacheSentinel, callback=None, update_metrics=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 a Deferred or the raw result
"""
callbacks = [callback] if callback else []
val = self._pending_deferred_cache.get(key, _CacheSentinel)
if val is not _CacheSentinel:
if val.sequence == self.sequence:
val.callbacks.update(callbacks)
if update_metrics:
self.metrics.inc_hits()
return val.deferred
val = self.cache.get(key, _CacheSentinel, callbacks=callbacks)
if val is not _CacheSentinel:
self.metrics.inc_hits()
return val
if update_metrics:
self.metrics.inc_misses()
if default is _CacheSentinel:
raise KeyError()
else:
return default
def set(self, key, value, callback=None):
callbacks = [callback] if callback else []
self.check_thread()
entry = CacheEntry(
deferred=value,
sequence=self.sequence,
callbacks=callbacks,
)
entry.callbacks.update(callbacks)
existing_entry = self._pending_deferred_cache.pop(key, None)
if existing_entry:
existing_entry.invalidate()
self._pending_deferred_cache[key] = entry
def shuffle(result):
if self.sequence == entry.sequence:
existing_entry = self._pending_deferred_cache.pop(key, None)
if existing_entry is entry:
self.cache.set(key, result, entry.callbacks)
else:
entry.invalidate()
else:
entry.invalidate()
return result
entry.deferred.addCallback(shuffle)
2016-08-19 04:18:26 -06:00
def prefill(self, key, value, callback=None):
callbacks = [callback] if callback else []
self.cache.set(key, value, callbacks=callbacks)
def invalidate(self, key):
self.check_thread()
# Increment the sequence number so that any SELECT statements that
# raced with the INSERT don't update the cache (SYN-369)
self.sequence += 1
entry = self._pending_deferred_cache.pop(key, None)
if entry:
entry.invalidate()
self.cache.pop(key, None)
def invalidate_many(self, key):
self.check_thread()
if not isinstance(key, tuple):
raise TypeError(
"The cache key must be a tuple not %r" % (type(key),)
)
self.sequence += 1
self.cache.del_multi(key)
2017-01-17 07:43:32 -07:00
entry_dict = self._pending_deferred_cache.pop(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.sequence += 1
self.cache.clear()
class _CacheDescriptorBase(object):
def __init__(self, orig, num_args, inlineCallbacks, cache_context=False):
self.orig = orig
if inlineCallbacks:
self.function_to_call = defer.inlineCallbacks(orig)
else:
self.function_to_call = orig
arg_spec = inspect.getargspec(orig)
all_args = arg_spec.args
if "cache_context" in all_args:
if not cache_context:
raise ValueError(
"Cannot have a 'cache_context' arg without setting"
" cache_context=True"
)
elif cache_context:
raise ValueError(
"Cannot have cache_context=True without having an arg"
" named `cache_context`"
)
if num_args is None:
num_args = len(all_args) - 1
if cache_context:
num_args -= 1
if len(all_args) < num_args + 1:
raise Exception(
"Not enough explicit positional arguments to key off for %r: "
"got %i args, but wanted %i. (@cached cannot key off *args or "
"**kwargs)"
% (orig.__name__, len(all_args), num_args)
)
self.num_args = num_args
2017-03-30 07:19:10 -06:00
# list of the names of the args used as the cache key
self.arg_names = all_args[1:num_args + 1]
2017-03-30 07:19:10 -06:00
# self.arg_defaults is a map of arg name to its default value for each
# argument that has a default value
if arg_spec.defaults:
self.arg_defaults = dict(zip(
all_args[-len(arg_spec.defaults):],
arg_spec.defaults
))
else:
self.arg_defaults = {}
if "cache_context" in self.arg_names:
raise Exception(
"cache_context arg cannot be included among the cache keys"
)
self.add_cache_context = cache_context
class CacheDescriptor(_CacheDescriptorBase):
""" A method decorator that applies a memoizing cache around the function.
This caches deferreds, rather than the results themselves. Deferreds that
fail are removed from the cache.
The function is presumed to take zero or more arguments, which are used in
a tuple as the key for the cache. Hits are served directly from the cache;
misses use the function body to generate the value.
The wrapped function has an additional member, a callable called
"invalidate". This can be used to remove individual entries from the cache.
The wrapped function has another additional callable, called "prefill",
which can be used to insert values into the cache specifically, without
calling the calculation function.
2016-08-19 04:18:26 -06:00
Cached functions can be "chained" (i.e. a cached function can call other cached
functions and get appropriately invalidated when they called caches are
invalidated) by adding a special "cache_context" argument to the function
and passing that as a kwarg to all caches called. For example::
2016-08-19 08:02:38 -06:00
@cachedInlineCallbacks(cache_context=True)
2016-08-19 04:18:26 -06:00
def foo(self, key, cache_context):
2016-08-19 08:13:58 -06:00
r1 = yield self.bar1(key, on_invalidate=cache_context.invalidate)
r2 = yield self.bar2(key, on_invalidate=cache_context.invalidate)
2016-08-19 04:18:26 -06:00
defer.returnValue(r1 + r2)
Args:
num_args (int): number of positional arguments (excluding ``self`` and
``cache_context``) to use as cache keys. Defaults to all named
args of the function.
"""
def __init__(self, orig, max_entries=1000, num_args=None, tree=False,
inlineCallbacks=False, cache_context=False, iterable=False):
super(CacheDescriptor, self).__init__(
orig, num_args=num_args, inlineCallbacks=inlineCallbacks,
cache_context=cache_context)
max_entries = int(max_entries * CACHE_SIZE_FACTOR)
self.max_entries = max_entries
self.tree = tree
self.iterable = iterable
def __get__(self, obj, objtype=None):
cache = Cache(
name=self.orig.__name__,
max_entries=self.max_entries,
keylen=self.num_args,
tree=self.tree,
iterable=self.iterable,
)
2017-05-04 07:18:46 -06:00
def get_cache_key_gen(args, kwargs):
"""Given some args/kwargs return a generator that resolves into
the cache_key.
We loop through each arg name, looking up if its in the `kwargs`,
otherwise using the next argument in `args`. If there are no more
args then we try looking the arg name up in the defaults
"""
pos = 0
for nm in self.arg_names:
if nm in kwargs:
yield kwargs[nm]
elif pos < len(args):
yield args[pos]
pos += 1
else:
yield self.arg_defaults[nm]
2017-05-04 07:18:46 -06:00
# By default our cache key is a tuple, but if there is only one item
# then don't bother wrapping in a tuple. This is to save memory.
if self.num_args == 1:
nm = self.arg_names[0]
def get_cache_key(args, kwargs):
if nm in kwargs:
return kwargs[nm]
elif len(args):
return args[0]
else:
return self.arg_defaults[nm]
else:
def get_cache_key(args, kwargs):
return tuple(get_cache_key_gen(args, kwargs))
@functools.wraps(self.orig)
def wrapped(*args, **kwargs):
2016-08-19 08:02:38 -06:00
# If we're passed a cache_context then we'll want to call its invalidate()
# whenever we are invalidated
2016-08-19 08:13:58 -06:00
invalidate_callback = kwargs.pop("on_invalidate", None)
2016-08-19 04:18:26 -06:00
2017-05-04 07:18:46 -06:00
cache_key = get_cache_key(args, kwargs)
2016-08-19 04:18:26 -06:00
# 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)
2016-08-19 04:18:26 -06:00
try:
2016-08-19 08:13:58 -06:00
cached_result_d = cache.get(cache_key, callback=invalidate_callback)
if isinstance(cached_result_d, ObservableDeferred):
observer = cached_result_d.observe()
else:
observer = cached_result_d
except KeyError:
ret = defer.maybeDeferred(
logcontext.preserve_fn(self.function_to_call),
obj, *args, **kwargs
)
def onErr(f):
cache.invalidate(cache_key)
return f
ret.addErrback(onErr)
2017-05-04 07:18:46 -06:00
# If our cache_key is a string, try to convert to ascii to save
# a bit of space in large caches
if isinstance(cache_key, basestring):
cache_key = to_ascii(cache_key)
result_d = ObservableDeferred(ret, consumeErrors=True)
cache.set(cache_key, result_d, callback=invalidate_callback)
observer = result_d.observe()
if isinstance(observer, defer.Deferred):
return logcontext.make_deferred_yieldable(observer)
else:
return observer
2017-05-04 07:18:46 -06:00
if self.num_args == 1:
wrapped.invalidate = lambda key: cache.invalidate(key[0])
wrapped.prefill = lambda key, val: cache.prefill(key[0], val)
else:
wrapped.invalidate = cache.invalidate
wrapped.invalidate_all = cache.invalidate_all
wrapped.invalidate_many = cache.invalidate_many
wrapped.prefill = cache.prefill
wrapped.invalidate_all = cache.invalidate_all
wrapped.cache = cache
wrapped.num_args = self.num_args
obj.__dict__[self.orig.__name__] = wrapped
return wrapped
class CacheListDescriptor(_CacheDescriptorBase):
"""Wraps an existing cache to support bulk fetching of keys.
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.
"""
def __init__(self, orig, cached_method_name, list_name, num_args=None,
inlineCallbacks=False):
"""
Args:
orig (function)
cached_method_name (str): The name of the chached method.
list_name (str): Name of the argument which is the bulk lookup list
num_args (int): number of positional arguments (excluding ``self``,
but including list_name) to use as cache keys. Defaults to all
named args of the function.
inlineCallbacks (bool): Whether orig is a generator that should
be wrapped by defer.inlineCallbacks
"""
super(CacheListDescriptor, self).__init__(
orig, num_args=num_args, inlineCallbacks=inlineCallbacks)
self.list_name = list_name
self.list_pos = self.arg_names.index(self.list_name)
self.cached_method_name = cached_method_name
self.sentinel = object()
if self.list_name not in self.arg_names:
raise Exception(
"Couldn't see arguments %r for %r."
% (self.list_name, cached_method_name,)
)
def __get__(self, obj, objtype=None):
cached_method = getattr(obj, self.cached_method_name)
cache = cached_method.cache
num_args = cached_method.num_args
@functools.wraps(self.orig)
def wrapped(*args, **kwargs):
2016-08-19 08:13:58 -06:00
# 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)
2016-08-19 04:18:26 -06:00
arg_dict = inspect.getcallargs(self.orig, obj, *args, **kwargs)
keyargs = [arg_dict[arg_nm] for arg_nm in self.arg_names]
list_args = arg_dict[self.list_name]
# cached is a dict arg -> deferred, where deferred results in a
# 2-tuple (`arg`, `result`)
2016-06-01 11:01:22 -06:00
results = {}
cached_defers = {}
missing = []
2017-05-22 08:12:19 -06:00
# If the cache takes a single arg then that is used as the key,
# otherwise a tuple is used.
if num_args == 1:
def cache_get(arg):
return cache.get(arg, callback=invalidate_callback)
else:
key = list(keyargs)
def cache_get(arg):
key[self.list_pos] = arg
return cache.get(tuple(key), callback=invalidate_callback)
for arg in list_args:
try:
2017-05-22 08:12:19 -06:00
res = cache_get(arg)
if not isinstance(res, ObservableDeferred):
results[arg] = res
elif not res.has_succeeded():
2016-06-01 11:01:22 -06:00
res = res.observe()
res.addCallback(lambda r, arg: (arg, r), arg)
cached_defers[arg] = res
else:
results[arg] = res.get_result()
except KeyError:
missing.append(arg)
if missing:
args_to_call = dict(arg_dict)
args_to_call[self.list_name] = missing
ret_d = defer.maybeDeferred(
logcontext.preserve_fn(self.function_to_call),
**args_to_call
)
ret_d = ObservableDeferred(ret_d)
# We need to create deferreds for each arg in the list so that
# we can insert the new deferred into the cache.
for arg in missing:
observer = ret_d.observe()
observer.addCallback(lambda r, arg: r.get(arg, None), arg)
observer = ObservableDeferred(observer)
if num_args == 1:
cache.set(
arg, observer,
callback=invalidate_callback
)
def invalidate(f, key):
cache.invalidate(key)
return f
observer.addErrback(invalidate, arg)
else:
key = list(keyargs)
key[self.list_pos] = arg
cache.set(
tuple(key), observer,
callback=invalidate_callback
)
def invalidate(f, key):
cache.invalidate(key)
return f
observer.addErrback(invalidate, tuple(key))
res = observer.observe()
res.addCallback(lambda r, arg: (arg, r), arg)
2016-06-01 11:01:22 -06:00
cached_defers[arg] = res
if cached_defers:
2016-06-03 04:47:07 -06:00
def update_results_dict(res):
results.update(res)
return results
return logcontext.make_deferred_yieldable(defer.gatherResults(
2016-06-01 11:01:22 -06:00
cached_defers.values(),
consumeErrors=True,
2016-06-03 04:47:07 -06:00
).addCallback(update_results_dict).addErrback(
2016-06-01 11:01:22 -06:00
unwrapFirstError
2016-06-03 04:47:07 -06:00
))
2016-06-01 11:01:22 -06:00
else:
return results
obj.__dict__[self.orig.__name__] = wrapped
return wrapped
class _CacheContext(namedtuple("_CacheContext", ("cache", "key"))):
2017-02-02 06:50:22 -07:00
# We rely on _CacheContext implementing __eq__ and __hash__ sensibly,
2017-02-02 07:47:15 -07:00
# 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.
2016-08-19 04:18:26 -06:00
def invalidate(self):
self.cache.invalidate(self.key)
def cached(max_entries=1000, num_args=None, tree=False, cache_context=False,
iterable=False):
return lambda orig: CacheDescriptor(
orig,
max_entries=max_entries,
num_args=num_args,
tree=tree,
2016-08-19 08:02:38 -06:00
cache_context=cache_context,
iterable=iterable,
)
def cachedInlineCallbacks(max_entries=1000, num_args=None, tree=False,
cache_context=False, iterable=False):
return lambda orig: CacheDescriptor(
orig,
max_entries=max_entries,
num_args=num_args,
tree=tree,
inlineCallbacks=True,
2016-08-19 08:02:38 -06:00
cache_context=cache_context,
iterable=iterable,
)
def cachedList(cached_method_name, list_name, num_args=None, inlineCallbacks=False):
2015-08-12 09:42:46 -06:00
"""Creates a descriptor that wraps a function in a `CacheListDescriptor`.
Used to do batch lookups for an already created cache. A single argument
is specified as a list that is iterated through to lookup keys in the
original cache. A new list consisting of the keys that weren't in the cache
get passed to the original function, the result of which is stored in the
cache.
Args:
cache (Cache): The underlying cache to use.
list_name (str): The name of the argument that is the list to use to
do batch lookups in the cache.
num_args (int): Number of arguments to use as the key in the cache
(including list_name). Defaults to all named parameters.
2015-08-12 09:42:46 -06:00
inlineCallbacks (bool): Should the function be wrapped in an
`defer.inlineCallbacks`?
Example:
class Example(object):
@cached(num_args=2)
def do_something(self, first_arg):
...
@cachedList(do_something.cache, list_name="second_args", num_args=2)
def batch_do_something(self, first_arg, second_args):
...
"""
return lambda orig: CacheListDescriptor(
orig,
cached_method_name=cached_method_name,
list_name=list_name,
num_args=num_args,
inlineCallbacks=inlineCallbacks,
)