synapse-product/synapse/metrics/__init__.py

221 lines
6.2 KiB
Python
Raw Normal View History

# -*- coding: utf-8 -*-
2016-01-06 23:26:29 -05:00
# 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.
# Because otherwise 'resource' collides with synapse.metrics.resource
from __future__ import absolute_import
import logging
from resource import getrusage, RUSAGE_SELF
2015-08-13 06:38:59 -04:00
import functools
import os
import stat
2015-08-13 06:38:59 -04:00
import time
import gc
2015-08-13 06:38:59 -04:00
from twisted.internet import reactor
from .metric import (
CounterMetric, CallbackMetric, DistributionMetric, CacheMetric
)
logger = logging.getLogger(__name__)
# We'll keep all the available metrics in a single toplevel dict, one shared
# for the entire process. We don't currently support per-HomeServer instances
# of metrics, because in practice any one python VM will host only one
# HomeServer anyway. This makes a lot of implementation neater
all_metrics = {}
class Metrics(object):
""" A single Metrics object gives a (mutable) slice view of the all_metrics
dict, allowing callers to easily register new metrics that are namespaced
nicely."""
def __init__(self, name):
self.name_prefix = name
def _register(self, metric_class, name, *args, **kwargs):
full_name = "%s_%s" % (self.name_prefix, name)
metric = metric_class(full_name, *args, **kwargs)
all_metrics[full_name] = metric
return metric
def register_counter(self, *args, **kwargs):
return self._register(CounterMetric, *args, **kwargs)
def register_callback(self, *args, **kwargs):
return self._register(CallbackMetric, *args, **kwargs)
def register_distribution(self, *args, **kwargs):
return self._register(DistributionMetric, *args, **kwargs)
def register_cache(self, *args, **kwargs):
return self._register(CacheMetric, *args, **kwargs)
def get_metrics_for(pkg_name):
""" Returns a Metrics instance for conveniently creating metrics
namespaced with the given name prefix. """
# Convert a "package.name" to "package_name" because Prometheus doesn't
# let us use . in metric names
return Metrics(pkg_name.replace(".", "_"))
def render_all():
strs = []
# TODO(paul): Internal hack
update_resource_metrics()
for name in sorted(all_metrics.keys()):
try:
strs += all_metrics[name].render()
2015-03-12 12:45:05 -04:00
except Exception:
strs += ["# FAILED to render %s" % name]
logger.exception("Failed to render %s metric", name)
2015-03-12 12:24:38 -04:00
strs.append("") # to generate a final CRLF
return "\n".join(strs)
# Now register some standard process-wide state metrics, to give indications of
# process resource usage
rusage = None
2015-03-12 12:24:38 -04:00
def update_resource_metrics():
global rusage
rusage = getrusage(RUSAGE_SELF)
resource_metrics = get_metrics_for("process.resource")
# msecs
resource_metrics.register_callback("utime", lambda: rusage.ru_utime * 1000)
resource_metrics.register_callback("stime", lambda: rusage.ru_stime * 1000)
# kilobytes
resource_metrics.register_callback("maxrss", lambda: rusage.ru_maxrss * 1024)
TYPES = {
stat.S_IFSOCK: "SOCK",
stat.S_IFLNK: "LNK",
stat.S_IFREG: "REG",
stat.S_IFBLK: "BLK",
stat.S_IFDIR: "DIR",
stat.S_IFCHR: "CHR",
stat.S_IFIFO: "FIFO",
}
2015-04-01 14:17:38 -04:00
def _process_fds():
counts = {(k,): 0 for k in TYPES.values()}
counts[("other",)] = 0
# Not every OS will have a /proc/self/fd directory
if not os.path.exists("/proc/self/fd"):
return counts
for fd in os.listdir("/proc/self/fd"):
try:
s = os.stat("/proc/self/fd/%s" % (fd))
fmt = stat.S_IFMT(s.st_mode)
if fmt in TYPES:
t = TYPES[fmt]
else:
t = "other"
counts[(t,)] += 1
except OSError:
# the dirh itself used by listdir() is usually missing by now
pass
return counts
get_metrics_for("process").register_callback("fds", _process_fds, labels=["type"])
2015-08-13 06:38:59 -04:00
reactor_metrics = get_metrics_for("reactor")
tick_time = reactor_metrics.register_distribution("tick_time")
pending_calls_metric = reactor_metrics.register_distribution("pending_calls")
gc_time = reactor_metrics.register_distribution("gc_time")
2015-08-13 06:38:59 -04:00
def runUntilCurrentTimer(func):
@functools.wraps(func)
def f(*args, **kwargs):
2015-08-14 10:42:52 -04:00
now = reactor.seconds()
num_pending = 0
# _newTimedCalls is one long list of *all* pending calls. Below loop
# is based off of impl of reactor.runUntilCurrent
2015-08-18 06:47:00 -04:00
for delayed_call in reactor._newTimedCalls:
if delayed_call.time > now:
2015-08-14 10:42:52 -04:00
break
2015-08-18 06:47:00 -04:00
if delayed_call.delayed_time > 0:
2015-08-14 10:42:52 -04:00
continue
num_pending += 1
num_pending += len(reactor.threadCallQueue)
2015-08-13 06:41:57 -04:00
start = time.time() * 1000
2015-08-13 06:38:59 -04:00
ret = func(*args, **kwargs)
end = time.time() * 1000
tick_time.inc_by(end - start)
2015-08-14 10:42:52 -04:00
pending_calls_metric.inc_by(num_pending)
2016-05-13 11:31:08 -04:00
# Check if we need to do a manual GC (since its been disabled), and do
# one if necessary.
threshold = gc.get_threshold()
counts = gc.get_count()
start = time.time() * 1000
for i in [2, 1, 0]:
if threshold[i] < counts[i]:
logger.info("Collecting gc %d", i)
gc.collect(i)
end = time.time() * 1000
gc_time.inc_by(end - start)
2015-08-13 06:38:59 -04:00
return ret
return f
try:
# Ensure the reactor has all the attributes we expect
reactor.runUntilCurrent
reactor._newTimedCalls
reactor.threadCallQueue
2015-08-13 06:38:59 -04:00
# runUntilCurrent is called when we have pending calls. It is called once
# per iteratation after fd polling.
reactor.runUntilCurrent = runUntilCurrentTimer(reactor.runUntilCurrent)
2016-05-13 11:31:08 -04:00
# We manually run the GC each reactor tick so that we can get some metrics
# about time spent doing GC,
gc.disable()
except AttributeError:
pass