| info: | Bucky Statsd and Collectd server for Graphite |
|---|
Bucky is a small server for collecting and translating metrics for Graphite. It can current collect metric data from CollectD daemons and from StatsD clients.
You can install with easy_install or pip as per normal modus operandi:
$ easy_install bucky # or $ pip install bucky
After installing, you can run Bucky like:
$ bucky
Bucky will try to install PyCrypto which requires the python-dev package to be installed.
By default, Bucky will open a CollectD UDP socket on 127.0.0.1:25826, a StatsD socket on 127.0.0.1:8125 as well as attempt to connect to a local Graphite (Carbon) daemon on 127.0.0.1:2003.
These are all optional as illustrated below. You can also disable the CollectD or StatsD servers completely if you so desire.
If the py-setproctitle module is installed Bucky will use it to set user readable process names. This will make the child processes of Bucky easier to identify. Please note that this is completely optional.
To install py-setproctitle run:
$ easy_install setproctitle # or $ pip install setproctitle
The astute observer will notice that Bucky has no flags for daemonization. This is quite on purpose. The recommended way to run Bucky in production is via runit. There's an example service directory in Bucky's source repository.
Bucky supports Python 3. However this support is still very young and we would like to hear from you if you are running Bucky on Python3 and help us improve the support in real production environments.
Bucky has support for logging error messages to Sentry via the Python Raven client.
To install raven run:
$ pip install raven # or $ easy_install raven
Next enable Sentry in Bucky's configuration file.
The command line options are limited to controlling the network parameters. If you want to configure some of the more intricate workings you'll need to use a config file. Here's the bucky -h output:
Usage: main.py [OPTIONS] [CONFIG_FILE]
Options:
--debug Put server into dry-run debug mode where output
goes to stdout instead of carbon. [False].
--metricsd-ip=IP IP address to bind for the MetricsD UDP socket
[127.0.0.1]
--metricsd-port=INT Port to bind for the MetricsD UDP socket [23632]
--disable-metricsd Disable the MetricsD UDP server
--collectd-ip=IP IP address to bind for the CollectD UDP socket
[127.0.0.1]
--collectd-port=INT Port to bind for the CollectD UDP socket [25826]
--collectd-types=FILE
Path to the collectd types.db file, can be specified
multiple times
--disable-collectd Disable the CollectD UDP server
--statsd-ip=IP IP address to bind for the StatsD UDP socket
[127.0.0.1]
--statsd-port=INT Port to bind for the StatsD UDP socket [8125]
--disable-statsd Disable the StatsD server
--graphite-ip=IP IP address of the Graphite/Carbon server [127.0.0.1]
--graphite-port=INT Port of the Graphite/Carbon server [2003]
--enable-influxdb Enable InfluxDB client
--full-trace Display full error if config file fails to load
--log-level=NAME Logging output verbosity [INFO]
--version show program's version number and exit
-h, --help show this help message and exit
The configuration file is a normal Python file that defines a number of variables. Most of command line options can also be specified in this file (remove the "--" prefix and replace "-" with "_") but if specified in both places, the command line takes priority. The defaults as a config file:
# Debug mode sends output only to stdout and not to carbon/graphite.
debug = False
# Valid log levels: DEBUG, INFO, WARNING, ERROR, CRITICAL
log_level = "INFO"
# The size of the input queue 0 is no max size.
# this is useful to set if you are concern about bucky consuming to
# much memory if the downstream systems are offline
max_sample_queue = 0
# Whether to print the entire stack trace for errors encountered
# when loading the config file
full_trace = False
# Basic metricsd conifguration
metricsd_ip = "127.0.0.1"
metricsd_port = 23632
metricsd_enabled = True
# The default interval between flushes of metric data to Graphite
metricsd_default_interval = 10.0
# You can specify the frequency of flushes to Graphite based on
# the metric name used for each metric. These are specified as
# regular expressions. An entry in this list should be a 3-tuple
# that is: (regexp, frequency, priority)
#
# The regexp is applied with the match method. Frequency should be
# in seconds. Priority is used to break ties when a metric name
# matches more than one handler. (The largest priority wins)
metricsd_handlers = []
# Basic collectd configuration
collectd_ip = "127.0.0.1"
collectd_port = 25826
collectd_enabled = True
# A list of file names for collectd types.db
# files.
collectd_types = []
# A mapping of plugin names to converter callables. These are
# explained in more detail in the README.
collectd_converters = {}
# Whether to load converters from entry points. The entry point
# used to define converters is 'bucky.collectd.converters'.
collectd_use_entry_points = True
# If a collectd metric is received with a value of type counter when
# our types.db define it as derive, or vice versa, don't raise an
# exception and assume the server's types.db is correct.
# Types counter and derive are very similar. Also, it's common
# for different versions/installations of collectd in 'clients'
# to have a bit different definitions for the same metrics
# (counter/derive conflict).
collectd_counter_eq_derive = False
# CollectD server can also run using multiple worker subprocesses.
# Incoming packets are routed to workers based on source IP.
collectd_workers = 1
# Cryptographic settings for collectd. Security level 1 requires
# signed packets, level 2 requires encrypted communication.
# Auth file should contain lines in the form 'user: password'
collectd_security_level = 0
collectd_auth_file = None
# Basic statsd configuration
statsd_ip = "127.0.0.1"
statsd_port = 8125
statsd_enabled = True
# How often stats should be flushed to Graphite.
statsd_flush_time = 10.0
# If the legacy namespace is enabled, the statsd backend uses the
# default prefixes except for counters, which are stored directly
# in stats.NAME for the rate and stats_counts.NAME for the
# absolute count. If legacy names are disabled, the prefixes are
# configurable, and counters are stored under
# stats.counters.{rate,count} by default. Any prefix can be set
# to None to skip it.
statsd_legacy_namespace = True
statsd_global_prefix = "stats"
statsd_prefix_counter = "counters"
statsd_prefix_timer = "timers"
statsd_prefix_gauge = "gauges"
# Timer thresholds
# Used to compute percentile values like:
# - stats.timers.my.awesome.timer.mean_90
statsd_percentile_thresholds = [90]
# Timer metrics
# These will enable or disable computing and futher dispatching metrics for specified type.
# Corresponding percentile metric will be disabled as well.
statsd_timer_mean = True
statsd_timer_upper = True
statsd_timer_lower = True
statsd_timer_count = True
statsd_timer_count_ps = True
statsd_timer_sum = True
statsd_timer_sum_squares = True
statsd_timer_median = True
statsd_timer_std = True
# Basic Graphite configuration
graphite_ip = "127.0.0.1"
graphite_port = 2003
# If the Graphite connection fails these numbers define how it
# will reconnect. The max reconnects applies each time a
# disconnect is encountered and the reconnect delay is the time
# in seconds between connection attempts. Setting max reconnects
# to a negative number removes the limit. The backoff factor
# determines how much the reconnect delay will be multiplied with
# each reconnect round. It can be limited with a maximum after which
# the delay will not be multiplied anymore.
graphite_max_reconnects = 3
graphite_reconnect_delay = 5
graphite_backoff_factor = 1.5
graphite_backoff_max = 60
influxdb_enabled = False
influxdb_hosts = [
"127.0.0.1:8089"
]
# Configuration for sending metrics to Graphite via the pickle
# interface. Be sure to edit graphite_port to match the settings
# on your Graphite cache/relay.
graphite_pickle_enabled = False
graphite_pickle_buffer_size = 500
# Bucky provides these settings to allow the system wide
# configuration of how metric names are processed before
# sending to Graphite.
#
# Prefix and postfix allow to tag all values with some value.
name_prefix = None
name_postfix = None
# The replacement character is used to munge any '.' characters
# in name components because it is special to Graphite. Setting
# this to None will prevent this step.
name_replace_char = '_'
# Optionally strip duplicates in path components. For instance
# a.a.b.c.c.b would be rewritten as a.b.c.b
name_strip_duplicates = True
# Bucky reverses hostname components to improve the locality
# of metric values in Graphite. For instance, "node.company.tld"
# would be rewritten as "tld.company.node". This setting allows
# for the specification of hostname components that should
# be stripped from hostnames. For instance, if "company.tld"
# were specified, the previous example would end up as "node".
name_host_trim = []
# processor is a callable that takes a (host, name, val, time)
# tuple as input and is expected to return a tuple of the same
# structure to forward the sample to the clients, or None to
# drop it. processor_drop_on_error specifies if the sample is
# dropped or forwarded to clients in case an exception is
# raised by the processor callable.
processor = None
processor_drop_on_error = False
You should only need to add something like this to your collectd.conf:
LoadPlugin "network" <Plugin "network"> Server "127.0.0.1" "25826" </Plugin>
Obviously, you'll want to match up the IP addresses and ports and make sure that your firewall's are configured to allow UDP packets through.
Just point your StatsD clients at Bucky's IP/Port and you should be good to go.
TODO
Make sure that your InfluxDB server(s) have a UDP listener enabled, like so:
[[udp]] enabled = true bind-address = ":8089" database = "mydatabase"
Bucky will periodically resolve all hostnames in the influxdb_hosts list and fan out metrics to all resolved endpoints. Thus providing replication as well as hot swapping.
CollectD metrics aren't exactly directly translatable to Graphite metric names. The default translator attempts to make a best guess but this can result in slightly less than pretty Graphite trees.
For this reason, Bucky has configurable converters. These are keyed off the CollectD plugin name. The input to these functions is a representation of the CollectD metric that looks like such:
{
'host': 'toroid.local',
'interval': 10.0,
'plugin': 'memory',
'plugin_instance': '',
'time': 1320970329.175534,
'type': 'memory',
'type_instance': 'inactive',
'value': 823009280.0,
'value_name': 'value',
'value_type': 1
}
The result of this function should be a list of strings that represent part of the Graphite metric name or None to drop sample entirely. For instance, if a converter returned ["foo", "bar"], the final metric name will end up as: $prefix.$hostname.foo.bar.$postfix.
An example builtin converter looks like such:
# This might be how you define a converter in
# your config file
class MemoryConverter(object):
PRIORITY = 0
def __call__(self, sample):
return ["memory", sample["type_instance"]]
collectd_converters = {"memory": MemoryConverter()}
Converters can either be declared and/or imported in the optional config file, or they can be autodiscovered via entry points. The entry point that is searched is "bucky.collectd.converters". The entry point name should be the CollectD plugin name.
collectd_converters in config file should be a mapping of collectd plugin name to converter instance. The default catch-all converter (used when no special converter is defined for a plugin) can be overidden by specifying _default as the plugin name.
Converters also have a notion of priority in order to resolve conflicts. This is merely a property on the callable named "PRIORITY" and larger priorities are preferred. I don't imagine this will need to be used very often, but its there just in case.
A Processor is a process that recieves samples as they are parsed by the servers and performs actions on them before handing them over to the clients.
If a callable is defined in the processor configuration variable, a Processor process will aply this callable to the sample recieved (host, name, val, time) and expects back a tuple of the same structure to forward to clients, or None to drop the sample.
This makes it easy to add all sorts of custom filtering and modification on samples.
This might be how you define a processor in your config file:
import time
def timediff(host, name, val, timestamp):
"""Drop samples with large time offset
Drop samples that are more than 2 minutes in the future
or more than 5 minutes in the past.
"""
future = 120 # 2 minutes
past = 300 # 5 minutes
now = time.time()
if timestamp > now + future or timestamp < now - past:
return None
return host, name, val, timestamp
processor = timediff