Skip to content
/ statman Public

Efficiently collect massive volumes of metrics inside the Erlang VM

License

Notifications You must be signed in to change notification settings

knutin/statman

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

statman - Statistics man to the rescue!

Statman makes it possible to instrument and collect statistics from your high-traffic production Erlang systems with very low overhead. The collected data points are aggregated in the VM and can be sent to services like Graphite, Munin, New Relic, etc.

Statman uses in-memory ETS tables for low overhead logging and to avoid single process bottlenecks. See "How does it work" below.

Integration options:

  • statman_elli: real-time (mobile friendly) web dashboard. Exposes a small web app and a HTTP API where external tools like Munin(plugin included), Librato, etc, can pull aggregated stats.

  • newrelic-erlang: Track web transactions happening in any Erlang webserver in New Relic, a hosted application monitoring service.

  • statman_graphite: Push data to a Graphite instance, also works with hostedgraphite.com.

  • hatman: Push data to stathat

Usage

Add statman_server to one of your supervisors with the following child specification. You can adjust the poll interval to your liking, it determines how frequently metrics will be pushed to the subscribers:

    {statman_server, {statman_server, start_link, [1000]},
     permanent, 5000, worker, []}.

Statman offers three data types. Here's how to use them:

%% Counters measure the frequency of an event
statman_counter:incr(my_queue_in).

%% A gauge is a point in time snapshot of a value
statman_gauge:set(queue_size, N).

%% Histograms show you the distribution of values
Result = statman:run({foo, bar}, fun () -> do_something() end)

Updates to counters, gauges and histograms involves one atomic write in ETS.

Decorators

You can instrument a function using one of the supplied decorators:

-decorate({statman_decorators, call_rate}).
my_function(A, B) ->
    A + B.

-decorate({statman_decorators, runtime, [{key, {statman, key}}]}).
other_function(foo) ->
    bar.

statman_poller

It's quite common to want to poll something at an interval, like memory usage, reduction counts, etc. To this end, Statman includes statman_poller which can run functions at intervals on your behalf. Add the supervisor to your supervision tree with the following child specification:

    {statman_poller_sup, {statman_poller_sup, start_link, []},
        permanent, 5000, worker, []}]}}.

In your app startup, you can then create pollers, which will be restarted if they crash and shut down together with your application:

queue_sizes() ->
    [{my_queue_size, my_queue:get_size()},
     {other_queue, foo:queue_size()}].

app_setup() ->
    ok = statman_poller:add_gauge(fun ?MODULE:queue_sizes/0, 1000).

A polling function can also be "stateful". Allowing you to measure the rate of change in an absolute number. If the function has arity 1, it will be passed the state and expected to return a new state:

widget_rate(undefined) ->
    TotalWidgets = count_total_widgets(),
    {TotalWidgets, []};
widget_rate(PrevTotalWidgets) ->
    TotalWidgets = count_total_widgets(),
    {TotalWidgets, [{created_widgets, TotalWidgets - PrevTotalWidgets}]}.

app_setup() ->
    ok = statman_poller:add_counter(fun ?MODULE:widget_rate/1, 1000).

It's important to pass a function reference rather than the function itself, to make code upgrades smoother.

How does it work

Using ets:update_counter/3 we get very efficient atomic increments / decrements of counters. With this primitive, counters, gauges and histograms become very efficient.

A histogram is really a frequency table of values. By keeping a count (weight) of how many times we have seen the different values, we have enough information to calculate the mean, min, max, standard deviation and percentiles.

Now, from this we can build something really cool:

  • The space required is proportionate to how many different values we have seen, not by the total number of observations. Binning values requires even less space.
  • Basic aggregation is done very early in the process. Binning also helps with this.
  • The frequency tables can easily be merged together, either to create an aggregate of multiple nodes to create a cluster view or aggregate over time to create for example 5 minute summaries.

Clusters

In a single node application, you can collect, aggregate and push out metrics from that single node. In bigger applications it might be helpful to collect metrics inside of each node, but aggregate together and view metrics for the whole cluster in one place. Having a "ops dashboard" showing message queues in key processes, node throughput, cluster throughput, request latency per node, request latency as a whole, etc, is extremely useful.

Setup

Statman has two parts, statman_server and statman_aggregator. The server owns the ETS-tables and periodically forwards the changes to any interested aggregator. The aggregator keeps a moving window of metrics coming from one ore more servers. You can ask the aggregator for the stats collected in the last N seconds.

You need to run one server under a supervisor in each node. If you have a cluster of nodes, you can run the aggregator on just one of them, collecting stats for the whole cluster.

About

Efficiently collect massive volumes of metrics inside the Erlang VM

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages