===============================
Deprication
This project is no longer supported by Spil Games and has been adopted by DB-Art <https://github.com/db-art/>
_.
The new repository can be found here:
MySQL-StatsD @ DB-Art <https://github.com/db-art/mysql-statsd>
_
mysql-statsd
Daemon that gathers statistics from MySQL and sends them to statsd.
Usage / Installation
Install mysql_statsd through pip(pip is a python package manager,
please don't use sudo!):
::
pip install mysql_statsd
If all went well, you'll now have a new executable called mysql_statsd
in your path.
Running mysql_statsd
::
$ mysql_statsd --config /etc/mysql-statsd.conf
Assuming you placed a config file in /etc/ named mysql-statsd.conf
See our example
configuration <https://github.com/spilgames/mysql-statsd/blob/master/docs/mysql-statsd.conf>
__
or read below about how to configure
Running the above command will start mysql_statsd in deamon mode. If
you wish to see it's output, then run the command with -f / --foreground
Usage
::
$ mysql_statsd --help
usage: mysql_statsd.py [-h] [-c FILE] [-d] [-f]
optional arguments:
-h, --help show this help message and exit
-c FILE, --config FILE
Configuration file
-d, --debug Prints statsd metrics next to sending them
--dry-run Print the output that would be sent to statsd without
actually sending data somewhere
-f, --foreground Dont fork main program
At the moment there is also a deamon script <https://github.com/spilgames/mysql-statsd/blob/master/docs/mysql_statsd>
_
for this package
You're more than welcome to help us improve it!
Platforms
We would love to support many other kinds of database servers, but
currently we're supporting these:
Both MySQL versions supported with Percona flavour as well as vanilla.
Todo:
Support for the following platforms
- Mysql 5.6
- MariaDB
We're looking forward to your pull request for other platforms
Development installation
------------------------
To install package, setup a `python virtual
environment <http://docs.python-guide.org/en/latest/dev/virtualenvs/>`_
Install the requirements(once the virtual environment is active):
::
pip install -r requirements.txt
*NOTE: MySQL-Python package needs mysql\_config command to be in your
path.*
There are future plans to replace the mysql-python package with
`PyMySQL <https://github.com/PyMySQL/PyMySQL>`_
After that you're able to run the script through
::
$ python mysql_statsd/mysql_statsd.py
Coding standards
----------------
We like to stick with the python standard way of working:
`PEP-8 <http://legacy.python.org/dev/peps/pep-0008/>`_
Configuration
=============
The configuration consists out of four sections:
- daemon specific (log/pidfiles)
- statsd (host, port, prefixes)
- mysql (connecitons, queries, etc)
- metrics (metrics to be stored including their type)
Daemon
------
The daemon section allows you to set the paths to your log and pic files
Statsd
------
The Statsd section allows you to configure the prefix and hostname of the
metrics. In our example the prefix has been set to mysql and the hostname
is included. This will log the status.com_select metric to:
mysql.<hostname>.status.com_select
You can use any prefix that is necessary in your environment.
MySQL
-----
The MySQL section allows you to configure the credentials of your mysql host
(preferrably on localhost) and the queries + timings for the metrics.
The queries and timings are configured through the stats_types configurable,
so take for instance following example:
::
stats_types = status, innodb
This will execute both the query_status and query_innodb on the MySQL server.
The frequency can then be controlled through the time (in milliseconds) set in
the interval_status and interval_innodb.
The complete configuration would be:
::
stats_types = status, innodb
query_status = SHOW GLOBAL STATUS
interval_status = 1000
query_innodb = SHOW ENGINE INNODB STATUS
interval_innodb = 10000
A special case is the query_commit: as the connection opened by mysql_statsd
will be kept open and auto commit is turned off by default the status
variables are not updated if your server is set to REPEATABLE_READ transaction
isolation. Also most probably your history_list will skyrocket and your
ibdata files will grow fast enough to drain all available diskspace. So when
in doubt about your transaction isolation: do include the query_commit!
Now here is the interesting part of mysql_statsd: if you wish to keep track
of your own application data inside your application database you *could*
create your own custom query this way. So for example:
::
stats_types = myapp
query_myapp = SELECT some_metric_name, some_metric_value FROM myapp.metric_table WHERE metric_ts >= DATE_SUB(NOW(), interval 1 MINUTE)
interval_myapp = 60000
This will query your application database every 60 seconds, fetch all the
metrics that have changed since then and send them through StatsD.
Obviously you need to whitelist them via the metrics section below.
Metrics
-------
The metrics section is basically a whitelisting of all metrics you wish to
send to Graphite via StatsD. Currently there is no possibilty to whitelist all
possible metrics, but there is a special case where we do allow wildcarding:
for the bufferpool\_* we whitelist all bufferpools with that specific metric.
Don't worry if you haven't configured multiple bufferpools: the output will
be omitted by InnoDB and also not parsed by the preprocessor.
Important to know about the metrics is that you will have to specify what type
they are. By default Graphite stores all metric equaly but treats them
differently per type:
- Gauge (g for gauge)
- Rate (r for raw, d for delta)
- Timer (t for timer)
Gauges are sticky values (like the spedometer in your car). Rates are the
number of units that need to be translated to units per second. Timers are
the time it took to perform a certain task.
An ever increasing value like the com\_select can be sent various ways. If you
wish to retain the absolute value of the com_select it is advised to configure
it as a gauge. However if you are going to use it as a rate (queries per
second) it is no use storing it as a rate in the first place and then later
on calculate the integral of the gauge to get the rate. It would be far more
accurate to store it as a rate in the first place.
Keep in mind that sending the com\_select value as a raw value is in this case
a bad habit: StatsD will average out the collected metrics per second, so
sending within a 10 second timeframe 10 times a value of 1,000,000 will average
out to the expected 1,000,000. However as the processing of metrics also takes
a bit of time the chance of missing one beat is relatively high and you end up
sending only 9 times the value, hence averaging out to 900,000 once in a while.
The best way to configure the com_select to a rate is by defining it as a delta.
The delta metric will remember the metric as it was during the previous run and
will only send the difference of the two values.
Media:
======
Art gave a talk about this tool at Percona London 2013:
http://www.percona.com/live/mysql-conference-2013/sessions/mysql-performance-monitoring-using-statsd-and-graphite
Contributors
------------
spil-jasper
thijsdezoete
art-spilgames
bnkr
History
-------
0.1.5 (2013-08-30)
++++++++++++++++++
* Support socket config
* Add innodb preprocessor update
0.1.1 (2013-08-30)
++++++++++++++++++
* Preparing package for sdist releases
0.1.0 (2013-08-30)
++++++++++++++++++
* First release on PyPI.