A pool of gen servers.
The goal of worker pool is pretty straightforward: To provide a transparent way to manage a pool of workers and do the best effort in balancing the load among them distributing the tasks requested to the pool.
The documentation can be generated from code using edoc with make doc
or using erldocs with make erldocs
. It is also available online here
All user functions are exposed through the wpool module.
Worker Pool is an erlang application that can be started using the functions in the application
module. For convinience, wpool:start/0
and wpool:stop/0
are also provided.
To start a new worker pool, you can either use wpool:start_pool
(if you want to supervise it yourself) or wpool:start_sup_pool
(if you want the pool to live under wpool's supervision tree). You can provide several options on any of those calls:
- overrun_warning: The number of milliseconds after which a task is considered overrun (i.e. delayed) so a warning is emitted using overrun_handler. The default value for this setting is
infinity
(i.e. no warnings are emitted) - overrun_handler: The module and function to call when a task is overrun. The default value for this setting is
{error_logger, warning_report}
. - workers: The number of workers in the pool. The default value for this setting is
100
- worker: The
gen_server
module that each worker will run and theInitArgs
to use on the correspondingstart_link
call used to initiate it. The default value for this setting is{wpool_worker, undefined}
. That means that if you don't provide a worker implementation, the pool will be generated with this default one.wpool_worker
is a module that implements a very simple rfc-like interface.
Since the workers are gen_server
s, messages can be call
ed or cast
ed to them. To do that you can use wpool:call
and wpool:cast
as you would use the equivalent functions on gen_server
.
Beyond the regular parameters for gen_server:call
and gen_server:cast
, wpool also provides an extra optional parameter: Strategy.
The strategy used to pick up the worker to perform the task. If not provided, the result of wpool:default_strategy/0
is used. The available strategies are defined in the wpool:strategy/0
type and also described below:
Picks the worker with the smaller queue of messages. Loosely based on this article. This strategy is usually useful when your workers always perform the same task, or tasks with expectedly similar runtimes.
Just picks a random worker. This strategy is the fastest one when to select a worker. It's ideal if your workers will perform many short tasks.
Picks the next worker in a round-robin fashion. That ensures evenly distribution of tasks.
Instead of just picking one of the workers in the queue and sending the request to it, this strategy queues the request and waits until a worker is available to perform it. That may render the worker selection part of the process much slower (thus generating the need for an aditional paramter: Worker_Timeout that controls how many milliseconds is the client willing to spend in that, regardless of the global Timeout for the call). This strategy ensures that, if a worker crashes, no messages are lost in its message queue. It also ensures that, if a task takes too long, that doesn't block other tasks since, as soon as other worker is free it can pick up the next task in the list.
Wpool provides a way to get live statistics about a pool. To do that, you can use wpool:stats/1
.
To stop a pool, just use wpool:stop/1
.
To see how wpool
is used you can check the test folder where you'll find many different scenarios excercised in the different suites.
If you want to see worker_pool in a real life project, I recommend you to check sumo_db, another open-source library from Inaka that uses wpool intensively.
wpool comes with a very basic benchmarker that let's you compare different strategies against the default wpool_worker
. If you want to do the same in your project, you can use wpool_bench
as a template and replace the worker and the tasks by your own ones.
For questions or general comments regarding the use of this library, please use our public hipchat room.
If you find any bugs or have a problem while using this library, please open an issue in this repo (or a pull request :)).
And you can check all of our open-source projects at inaka.github.io