Exometer - Erlang instrumentation package

Version: Mar 19 2014 23:14:26

Authors: Ulf Wiger (ulf.wiger@feuerlabs.com), Magnus Feuer (magnus.feuer@feuerlabs.com).

The Exometer package allows for easy and efficient instrumentation of Erlang code, allowing crucial data on system performance to be exported to a wide variety of monitoring systems.

Exometer comes with a set of pre-defined monitor components, and can be expanded with custom components to handle new types of Metrics, as well as integration with additional external systems such as databases, load balancers, etc.

This document gives a high level overview of the Exometer system. For details, please see the documentation for individual modules, starting with exometer.

Table of Content

Concept and definitions
Built-in entries and probes
Built in Reporters
Instrumenting Erlang code
Configuring Exometer
Creating custom exometer entries
Creating custom probes
Creating custom reporter plugins

Concepts and Definitions

Exometer introduces a number of concepts and definitions used throughout the documentation and the code.

Metric

A metric is a specific measurement sampled inside an Erlang system and then reported to the Exometer system. An example metric would be "transactions_per_second", or "memory_usage".

Metrics are identified by a list of terms, such as given below:

[ xml_front_end, parser, file_size ]

A metric is created through a call by the code to be instrumented to exometer:new(). Once created, the metric can be updated through exometer:update(), or on its own initiative through the exometer_probe:sample behavior implementation.

Data Point

Each metric can consist of multiple data points, where each point has a specific value.

A typical example of data points would be a transactions_per_second (tps) metric, usually stored as a histogram covering the last couple of minutes of tps samples. Such a histogram would host multiple values, such as min, max, median, mean, 50_percentile, 75_percentile, etc.

It is up to the type of the metric, and the data probe backing that type (see below), to specify which data points are available under the given metric.

Metric Type

The type of a metric, specified when the metric is created through exometer:new(), determines which exometer_entry callback to use.

The link between the type and the entry to use is configured through the exometer_admin module, and its associated exometer defaults configuration data.

The metric type, in other words, is only used to map a metric to a configurable exometer_entry callback.

Entry Callback

An exometer entry callback will receive values reported to a metric through the exometer:update() call and compile it into one or more data points. The entry callback can either be a counter (implemented natively in exometer), or a more complex statistical analysis such as a uniform distribution or a regular histogram.

The various outputs from these entries are reported as data points under the given metric.

An entry can also interface external analytics packages. exometer_folsom, for example, integrates with the folsom_metrics package found at https://github.com/boundary/folsom.

Probe

Probes are a further specialization of exometer entries that run in their own Erlang processes and have their own state (like a gen_server). A probe is implemented through the exometer_probe behavior.

A probe can be used if independent monitoring is needed of, for example, /proc trees, network interfaces, and other subsystems that need periodic sampling. In these cases, the exometer_probe:probe_sample() call is invoked regularly by exometer, in the probe's own process, in order to extract data from the given subsystem and add it to the metric's data points.

Caching

Metric and data point values are read with the exometer:get_value() function. In the case of counters, this operation is very fast. With probes, the call results in a synchronous dialog with the probe process, and the cost of serving the request depends on the probe implementation and the nature of the metric being served.

If the cost of reading the value is so high that calling the function often would result in prohibitive load, it is possible to cache the value. This is done either explicitly from the probe itself (by calling exometer_cache:write()), or by specifying the option {cache, Lifetime} for the entry. If an entry has a non-zero cache lifetime specified, the get_value() call will try fetching the cached value before calling the actual entry and automatically caching the result.

Note that if {cache, Lifetime} is not specified, exometer:get_value() will neither read nor write to the cache. It is possible for the probe to periodically cache a value regardless of how the cache lifetime is set, and the probe may also explicitly read from the cache if it isn't done automatically.

Subscriptions and Reporters

The subscription concept, managed by exometer_report allows metrics and their data points to be sampled at given intervals and delivered to one or more recipients, which can be either an arbitrary process or a Reporter plugin.

Each subscription ties a specific metric-datapoint pair to a reporter and an interval (given in milliseconds). The reporter system will, at the given interval, send the current value of the data point to the subscribing reporter. The subscription, with all its parameters, is setup through a call to exometer_report:subscribe().

In the case of processes, subscribed-to values will be delivered as a message. Modules, which implement the exometer_report callback behavior, will receive the plugins as a callbacks within the exometer_report process.

Subscriptions can either be setup at runtime, through exometer_report:subscribe() calls, or statically through the exometer_report configuration data.

Built-in entries and probes

There are a number of built-in entries and probes shipped with the Exometer package, as described below:

counter (exometer native)

The counter is implemented directly in exometer to provide simple counters. A call to exometer:update() will add the provided value to the counter.

The counter can be reset to zero through exometer:reset().

The available data points under a metric using the counter entry are value and ms_since_reset.

fast_counter (exometer native)

A fast counter implements the counter functionality, through the trace_info system, yielding a speed increase of about 3.5 in comparison to the regular counter.

The tradeoff is that running tracing and/or debugging may interfere with the counter functionality.

A call to exometer:update() will add the provided value to the counter.

The counter can be reset to zero through exometer:reset().

The available data points under a metric using the fast_counter entry are value and ms_since_reset.

exometer_histogram (probe)

The histogram probe stores a given number of updates, provided through exometer:update(), in a histogram. The histogram maintains a log derived from all values received during a configurable time span and provides min, max, median, mean, and percentile analysis data points for the stored data.

In order to save memory, the histogram is divided into equal-sized time slots, where each slot spans a settable interval. All values received during a time slot will be averaged into a single value to be stored in the histogram once the time slot expires. The averaging function (which can be replaced by the caller), allows for high-frequency update metrics to have their resolution traded against resource consumption.

exometer_uniform (probe)

The uniform probe provides a uniform sample over a pool of values provided through exometer:update(). When the pool reaches its configurable max size, existing values will be replaced at random to make space for new values. Much like exometer_histogram, the uniform probe provides min, max, median, mean, and percentile analysis data points for the stored data.

exometer_spiral (probe)

The spiral probe maintains the total sum of all values stored in its histogram. The histogram has a configurable time span, all values provided to the probe, through exometer:update(), within that time span will be summed up and reported. If, for example, the histogram covers 60 seconds, the spiral probe will report the sum of all values reported during the last minute.

The grand total of all values received during the lifetime of the probe is also available.

exometer_folsom [entry]

The folsom entry integrates with the folsom metrics package provided by the boundary repo at github. Updated values sent to the folsom entry can be forwarded to folsom's counter, histogram, duration, meter, and spiral.

Folsom integration is provided as a backup. New code using Exometer should use the native probes that duplicate folsom.

exometer_function [entry]

The function entry allows for a simple caller-supplied function to be invoked in order to retrieve non-exometer data. The exometer_function:get_value() function will invoke a Module:Function(DataPoints) call, where Module and Function are provided by the caller.

The function entry provides an easy way of integrating an external system without having to write a complete entry.

Built in Reporters

Exometer ships with some built-in reporters which can be used to forward updated metrics and their data points to external systems. They can also serve as templates for custom-developed reporters.

exometer_report_graphite

The graphite reporter uses the TCP/IP protocol to forward subscribed-to metrics and data points to a graphite server, such as the one provided by http://hostedgraphite.com. When the graphite reporter receives a metric-datapoint value (subscribed to through exometer_report:subscriber()), the reporter will immediately forward the key-value pair to the graphite server.

exometer_report_collectd

The collectd reporter communicates with a local collectd process through its unix socket protocol. All subscribed-to metric-datapoint values received by the reporter are immediately forwarded to collectd. Once a value has been forwarded, the reporter continuously refreshes the value toward collectd at a configurable interval in order to keep it from expiring inside collectd.

If the collectd connection is lost, the reporter will attempt to reconnect to it at a configurable interval.

All metrics reported to collectd will be have identifiers formatted as follows:

HostName/PluginName-PluginInstance/Type-Metric_DataPoint

HostName

Host name of the entry.

Configurable through the hostname application environment parameter.

Default is the value returned by netadm:localhost().
PluginName

The collectd plugin name.

Configurable through the plugin_name application environment parameter.

Default is exometer.
PluginInstance

The instance ID to use for the plugin.

Configurable through the plugin_instance application environment parameter.

Default is the erlang node name in the left hand side of the value returned by node().
Type

Type assigned to the reported value.

The type is looked up through the type_map.

The given metric and data points are used as a key in a list format, such as [ db, cache, hits, median ]. The type that is resolved from the metric/data point will be used as the Type component in the collectd identifier. Please see types.db(5) for a list of available collectd types.

Default for Type is 'gauge'.
Metric

The name of the metric. The atoms in the metric list will be converted to a string separated by _. Thus [ db, cache, hits ] will be converted to db_cache_hits.
DataPoint

The data point of the given metric. Will be added to the end of the metrics string.

Please see Configuring collectd reporter for details on the application environment parameters listed above.

exometer_report_snmp

The SNMP reporter enables the export of metrics and their datapoints to SNMP managers. The export needs to be enabled for each metric through their options. Moreover, SNMP notifications can be created using the options to send periodic reports on datapoints to SNMP managers. All SNMP protocol handling is done by the snmp application shipped with Erlang/OTP. Thus, the snmp application needs to be started and the local SNMP master agent needs to be configured correctly for SNMP export to work properly.

To configure SNMP export for a single metric use these options:

{snmp, disabled} (default)

Disables SNMP export for the metric. Same as not specifying the option at all.
{snmp, []}

Enables SNMP export for the metric. No subscriptions are setup.
{snmp, [{Datapint, Interval}]}

Enables SNMP export for the metric.

Subscriptions are setup for the given Datapoint/Interval pairs.

Each subscription report will be forwarded to SNMP mangers as notifications.
{snmp, [{Datapint, Interval, Extra}]}

Same as above, but using an addition extra identification for the subscriptions.

Allow the creation ofmultiple subscriptions for a single datapoint.

Please see Configuring collectd reporter for details on how to configure the SNMP reporter.

Instrumenting Erlang code

The code using Exometer needs to be instrumented in order to setup and use metrics reporting.

Exometer Start

The system using Exometer must start the exometer application prior to using it:

application:start(exometer).

Once started, the default mapping between metrics and the entries is loaded from the configuration data:

exometer_admin:preset_defaults().

See Configuring Exometer for details on configuration data format.

Creating metrics

A metric, can be created throuh a call to

exometer:new(Name, Type)

Name is a list of atoms, uniquely identifying the metric created. The type of the metric, specified by Type will be mapped to an exometer entry through the table maintained by exometer_admin Please see the Configuring type - entry maps for details.

The resolved entry to use will determine the data points available under the given metric.

Deleting metrics

A metric previously created with exometer:new() can be deleted by exometer:delete().

All subscriptions to the deleted metrics will be cancelled.

Setting metric values

A created metric can have its value updated through the exometer:update() function:

exometer:update(Name, Value)

The Name parameter is the same atom list provided to a previous exometer:new() call. The Value is an arbitrarty element that is forwarded to the exometer:update() function of the entry/probe that the metric is mapped to.

The receiving entry/probe will process the provided value and modify its data points accordingly.

Retrieving metric values

Exometer-using code can at any time retrieve the data point values associated with a previously created metric. In order to find out which data points are available for a metric, the following call can be used:

exometer:info(Name, datapoints)

The Name parameter is the same atom list provided to a previous exometer:new() call. The call will return a list of data point atoms that can then be provided to exometer:get_value() to retrieve their actual value:

exometer:get_value(Name, DataPoint)

The Name paramer identifies the metric, and DataPoints identifies the data points (returned from the previous info() call) to retrieve the value for.

If no DataPoints are provided, the values of a default list of data points, determined by the backing entry / probe, will be returned.

Setting up subscriptions

A subscription can either be statically configured, or dynamically setup from within the code using Exometer. For details on statically configured subscriptions, please see Configuring static subscriptions.

A dynamic subscription can be setup with the following call:

exometer_report:subscribe(Recipient, Metric, DataPoint, Inteval)

Recipient is the name of a reporter.

Set metric options

Each created metric can have options setup for it through the following call:

exometer:setopts(Name, Options)

The Name paramer identifies the metric to set the options for, and Options is a proplist ([{ Key, Value },...]) with the options to be set.

Exometer looks up the the backing entry that hosts the metric with the given Name, and will invoke the entry's setopts/4 function to set the actual options. Please see thesetopts/4 function for the various entries for details.

Configuring Exometer

Exometer defaults can be changed either through OTP application environment variables or through the use of Basho's cuttlefish (https://github.com/basho/cuttlefish).

Configuring type - entry maps

The dynamic method of configuring defaults for exometer entries is:

exometer_admin:set_default(NamePattern, Type, Default)

Where NamePattern is a list of terms describing what is essentially a name prefix with optional wildcards ('_'). A pattern that matches any legal name is ['_'].

Type is an atom defining a type of metric. The types already known to exometer, counter, fast_counter, ticker, uniform, histogram, spiral, netlink, and probe may be redefined, but other types can be described as well.

Default is either an #exometer_entry{} record (unlikely), or a list of {Key, Value} options, where the keys correspond to #exometer_entry record attribute names. The following attributes make sense to preset:

{module, atom()}              % the callback module
{status, enabled | disabled}  % operational status of the entry
{cache, non_neg_integer()}    % cache lifetime (ms)
{options, [{atom(), any()}]}  % entry-specific options

Below is an example, from exometer/priv/app.config:

{exometer, [
    {defaults, [
        {['_'], function , [{module, exometer_function}]},
        {['_'], counter  , [{module, exometer}]},
        {['_'], histogram, [{module, exometer_histogram}]},
        {['_'], spiral   , [{module, exometer_spiral}]},
        {['_'], duration , [{module, exometer_folsom}]},
        {['_'], meter    , [{module, exometer_folsom}]},
        {['_'], gauge    , [{module, exometer_folsom}]}
    ]}
]}

In systems that use CuttleFish, the file exometer/priv/exometer.schema contains a schema for default settings. The setup corresponding to the above defaults would be as follows:

exometer.template.function.module  = exometer_function
exometer.template.counter.module   = exometer
exometer.template.histogram.module = exometer_histogram
exometer.template.spiral.module    = exometer_spiral
exometer.template.duration.module  = exometer_folsom
exometer.template.meter.module     = exometer_folsom
exometer.template.gauge.module     = exometer_folsom

Configuring statically defined entries

Using the exometer environment variable predefined, entries can be added at application startup. The variable should have one of the following values:

{script, File} - File will be processed using file:script/2. The return value (the result of the last expression in the script) should be a list of{Name, Type, Options} tuples.
{apply, M, F, A} - The result of apply(M, F, A) should be {ok, L} whereL is a list of {Name, Type, Options} tuples.
L, where L is a list of {Name, Type, Options} tuples.

Configuring static subscriptions

Static subscriptions, which are automatically setup at exometer startup without having to invoke exometer_report:subscribe(), are configured through the report sub section under exometer.

Below is an example, from exometer/priv/app.config:

{exometer, [
    {report, [
        {subscribers, [
            {exometer_report_collectd, [db, cache, hits], mean, 2000, true},
            {exometer_report_collectd, [db, cache, hits], max, 5000, false}
        ]}
    ]}
]}

The report section configures static subscriptions and reporter plugins. See Configuring reporter plugins for details on how to configure individual plugins.

The subscribers sub-section contains all static subscriptions to be setup att exometer applications start. Each tuple in the prop list should be of one of the following formats:

{Reporter, Metric, DataPoint, Interval} {Reporter, Metric, DataPoint, Interval, RetryFailedMetrics} {Reporter, Metric, DataPoint, Interval, RetryFailedMetrics, Extra} {apply, {M, F, A}}

In the case of {apply, M, F, A}, the result of apply(M, F, A) must be a list of subscribers tuples.

The meaning of the above tuple elements is:

Reporter :: module()

Specifies the reporter plugin module, such asexometer_report_collectd that is to receive updated metric's data points.
Metric :: [atoms()]

Specifies the path to a metric previously created with anexometer:new() call.
DataPoint :: atom() | [atom()]'

Specifies the data point within the given metric to send to the receiver. The data point must match one of the data points returned byexometer:info(Name, datapoints) for the given metrics name.
Interval :: integer()' (milliseconds)

Specifies the interval, in milliseconds, between each update of the given metric's data point. At the given interval, the data point will be samples, and the result will be sent to the receiver.
RetryFailedMetrics :: boolean()

Specifies if the metric should be continued to be reported even if it is not found during a reporting cycle. This would be the case if a metric is not created by the time it is reported for the first time. If the metric will be created at a later time, this value should be set to true. Set this value to false if all attempts to report the metric should stop if when is not found. The default value is true.
Extra :: any()

Provides a means to pass along extra information for a given subscription. An example is the syntax option for the SNMP reporter, in which case Extra needs to be a property list.

Configuring reporter plugins

The various reporter plugins to be loaded by exometer are configured in the report section under reporters

Each reporter has an entry named after its module, and the content of that entry is dependent on the reporter itself. The following chapters specifies the configuration parameters for the reporters shipped with exometer.

Configuring collectd reporter

Below is an example of the collectd reporter application environment, with its correct location in the hierarchy:

{exometer, [
    {report, [
        {reporters, [ 
            {exometer_report_collectd, [ 
                {reconnect_interval, 10},
                {refresh_interval, 20}, 
                {read_timeout, 5000}, 
                {connect_timeout, 8000}, 
                {hostname, "testhost"}, 
                {path, "/var/run/collectd-unixsock"},
                {plugin_name, "testname"},
                {plugin_instance, "testnode"},
                {type_map, 
                    [{[db, cache, hits, max], "gauge"}]
                }
            ]}
        ]}
    ]}
]}

The following attributes are available for configuration:

reconnect_interval (seconds - default: 30)

Specifies the duration between each reconnect attempt to a collectd server that is not available. Should the server either be unavailable at exometer startup, or become unavailable during exometer's operation, exometer will attempt to reconnect at the given number of seconds.
refresh_interval (seconds - default: 10)

Specifies how often a value, which has not been updated by exometer, is to be resent with its current value to collectd. If collectd does not see an identifier updated at a given number of seconds (see Timeout in collectd.conf(5)), it will be removed from the list of maintained identifiers.
read_timeout (milliseconds - default: 5000)

Specifies how long the collectd reporter plugin shall wait for an acknowledgement from collectd after sending it an updated value. If an acknowledgment is not received within the given interval, the plugin will disconnect from the collectd server and reconnect to it after the given reconnect interval (see item one above).
connect_timeout (milliseconds - default: 5000)

Specifies how long the collectd reporter plugin shall wait for a unix socket connection to complete before timing out. A timed out connection attempt will be retried after the reconnect interval has passed see item 1 above).
path (file path - default: "/var/run/collectd-unixsock")

Specifies the path to the named unix socket that collectd is listening on. When exometer starts and loads the collectd reporter plugin, the plugin will connect to the given socket.
plugin_name (string - default: "exometer")

Specifies the plugin name to use when constructing an collectd identifier. Please see Configuring collectd reporter for details.
plugin_instance (string - default: left hand side of node())

Specifies the plugin instance id to use when constructing an collectd identifier. Please see Configuring collectd reporter for details.
plugin_instance (string - default: left hand side of node())

Specifies the plugin instance id to use when constructing an collectd identifier. Please see Configuring collectd reporter for details.
hostname (string - default: net_adm:localhost())

Specifies the host name to use when constructing an collectd identifier. Please see Configuring collectd reporter for details.
type_map (prop list - default: n/a)

Specifies the mapping between metrics/datapoints and the collectd type to use when sending an updated metric value. See below.

Type maps must be provided since there is no natural connection between the type of a metric/datapoint pair and an identifier in collectd. The type_map consists of a prop list that converts a path to a metric/datapoint to a string that can be used as a type when reporting to collectd.

The key part of each element in the list consists of a list of atoms that matches the name of the metrics, with the name of the data point added as a final element. If the metric is identified as [ webserver, https, get_count ], and the data point is called total, the key in the type_map would be [ webserver, https, get_count, total ], The value part of a property is the type string to use when reporting to collectd. Please see types.db(5) for a list of available collectd types. A complete entry in the type_map list would be: { [ webserver, https, get_count, total ], "counter" }.

Configuring graphite reporter

Below is an example of the a graphite reporter application environment, with its correct location in the hierarchy:

{exometer, [
    {report, [
        {reporters, [ 
            {exometer_report_graphite, [ 
                {connect_timeout, 5000},
                {prefix, "web_stats"}, 
                {host, "carbon.hostedgraphite.com"}, 
                {port, 2003}, 
                {api_key, "267d121c-8387-459a-9326-000000000000"}
            ]}
        ]}
    ]}
]}

The following attributes are available for configuration:

connect_timeout (milliseconds - default: 5000)

Specifies how long the graphie reporter plugin shall wait for a tcp connection to complete before timing out. A timed out connection will not be reconnected to automatically. (To be fixed.)
prefix (string - default: "")

Specifies an optional prefix to prepend all metric names with before they are sent to the graphite server.
host (string - default: "carbon.hostedgraphite.com")

Specifies the name (or IP address) of the graphite server to report to.
port (integer - default: 2003)

Specifies the TCP port on the given graphite server to connect to.
api_key (string - default: n/a)

Specifies the api key to use when reporting to a hosted graphite server.

If prefix is not specified, but api_key is, each metrics will be reported as ApiKey.Metric.

If prefix is specified, but api_key is not, each metrics will be reported as Prefix.Metric.

if neither prefix or api_key is specified, each metric will be reported simply as Metric.

Configuring snmp reporter

Below is an example of the a snmp reporter application environment, with its correct location in the hierarchy:

{exometer, [
    {report, [
        {reporters, [ 
            {exometer_report_snmp, [ 
                {mib_template, "priv/MYORG-EXOMETER-METRICS.mib"},
                {mib_dir, "/tmp/exometer"}
            ]}
        ]}
    ]}
]}

The following attributes are available for configuration:

mib_template (string - default: "mibs/EXOMETER-METRICS-MIB.mib")

Specifies where to find the MIB template used for dynamically assembline an internal MIB. Take a look at the MIB template shipped with Exometer for reference in case you want to define your own template.
mib_dir (string - default: "tmp/exometer_report_snmp")

Specifies temporary direction which will be used by Exometer to store dymanically created MIB files.