A simple statistics collector, written in Google Go.

First make sure you have the a working Go environment. See the install instructions. MetricsD will always compile on the release tag.

Then you will be able to clone and compile MetricsD:

git clone --recursive --branch release git://github.com/kpumuk/metricsd.git
cd metricsd
make install

By default MetricsD will be install to the /usr/local/metricsd directory. To change it, use DESTINATION environment variable with make install:

DESTINATION=/opt/metricsd make install

Please note: if you want to run make install with sudo, make sure that root user has all Go environment variables defined. Also you can use -E switch to preserve your current environment (needs setenv option set in sudoers):

sudo -E make install

Configuration is stored in JSON format, and you can find an example in metricsd.conf.example. Every config option could be overridden using command-line arguments. Following options available at the moment:

• Listen (-listen) — set the port (+optional address) to listen at. Default is "0.0.0.0:6311";
• DataDir (-data) — set the data directory. Default is "./data";
• LogLevel (-debug) — set the debug level, the lower - the more verbose (0-5). Default is 1;
• SliceInterval (-slice) — set the slice interval in seconds. Default is 10;
• WriteInterval (-write) — set the write interval in seconds. Default is 60;
• BatchWrites (-batch) — set the value indicating whether batch RRD updates should be used. Default is false;
• LookupDns (-lookup) — set the value indicating whether reverse DNS lookup should be performed for sources.

Another command-line options:

• -test — validate the configuration file and exit.
• -config — path to the configuration file.

MetricsD uses very simple UDP-based protocol for collecting metrics. Here is what it looks like:

1. metric:value — in this simplest case value will be collected in several RRD files; for each writer (see below) two files will be created: IP/metric-writer.rrd and all/metric-writer.rrd, where writer is a name of writer, IP — an IP address of the source host, metric — metric name.
2. source@metric:value — the same as previous, but instead of IP address of the source host, source will be used. If it's equal to all, no per-host RRD file will be created, only summary for all ones.
3. metric:value;source@metric:value — it's possible to send several metrics update in a single packet. Please note: you have to specify source for every metric (metrics without source will be saved to IP-based RRD files).
4. group$metric:value — metrics could be grouped in UI based on the group value.

Examples:

response_time:153
    all/response_time-quartiles.rrd, 10.0.0.1/response_time-quartiles.rrd,
    all/response_time-yesno.rrd, 10.0.0.1/response_time-yesno.rrd

app01@response_time:153
    all/response_time-quartiles.rrd, app01/response_time-quartiles.rrd,
    all/response_time-yesno.rrd, app01/response_time-yesno.rrd

all@response_time:153
    all/response_time-quartiles.rrd, all/response_time-yesno.rrd

app01@response_time:153;all@requests:-1
    all/response_time-quartiles.rrd, app01/response_time-quartiles.rrd,
    all/response_time-yesno.rrd, app01/response_time-yesno.rrd,
    all/requests-quartiles.rrd, all/requests-yesno.rrd

Writer is an implementation of a metrics aggregation algorithm. Each writer generates an RRD file with different (most probably) datasources and RRAs to store aggregated metrics.

There are three writers currently implemented:

1. count — calculates number of successful (value > 0) and failes (value < 0) events. Data sources: ok — number of successful events, fail — number of failed events.
2. quartiles — calculates quartiles for input data. Creates following data sources: q1 (first quartile), q2 (second quartile), q3 (third quartile), hi (max sample), lo (min sample), total (number of samples).
3. percentiles — calculates 90th and 95th percentiles for input data, along with mean value and standard deviation for values under the percentile. Creates following data sources: pct90 (90th percentile), pct90mean (mean of values under 90th percentile), pct90dev (standard deviation of values under 95th percentile), pct95 (95th percentile), pct95mean (mean of values under 95th percentile), pct95dev (standard deviation of values under 95th percentile).