Twitter Ostrich

Ostrich is a library for scala servers that makes it easy to:

• load & reload per-environment configuration
• collect runtime statistics (counters, gauges, metrics, and labels)
• report those statistics through a simple web interface (optionally with graphs) or into log files
• interact with the server over HTTP to check build versions or shut it down

The idea is that it should be simple and straightforward, allowing you to plug it in and get started quickly.

This library is released under the Apache Software License, version 2, which should be included with the source in a file named LICENSE.

Building

Use sbt (simple-build-tool) to build:

$ sbt clean update package-dist 

The finished jar will be in dist/.

Counters, Gauges, Metrics, and Labels

There are four kinds of statistics that ostrich captures:

• counters

  A counter is a value that never decreases. Examples might be "widgets_sold" or "births". You just increment the counter each time a countable event happens, and graphing utilities usually graph the deltas over time. To increment a counter, use:

   Stats.incr("births") 

  or

   Stats.incr("widgets_sold", 5) 

• gauges

  A gauge is a value that has a discrete value at any given moment, like "heap_used" or "current_temperature". It's usually a measurement that you only need to take when someone asks. To define a gauge, stick this code somewhere in the server initialization:

   Stats.addGauge("current_temperature") { myThermometer.temperature } 

  A gauge method must always return a double.

• metrics

  A metric is tracked via distribution, and is usually used for timings, like so:

   Stats.time("translation") { document.translate("de", "en") } 

  But you can also add metrics directly:

   Stats.addMetric("query_results", results.size) 

  Metrics are collected by tracking the count, min, max, mean (average), and a simple bucket-based histogram of the distribution. This distribution can be used to determine median, 90th percentile, etc.

• labels

  A label is just a key/value pair of strings, usually used to report a subsystem's state, like "boiler=offline". They're set with:

   Stats.setLabel("boiler", "online") 

  They have no real statistical value, but can be used to raise flags in logging and monitoring.

RuntimeEnvironment

If you build with standard-project http://github.com/twitter/standard-project, RuntimeEnvironment can pull build and environment info out of the build.properties file that's tucked into your jar. Typical use is to pass your server object (or any object from your jar) and any command-line arguments you haven't already parsed:

val runtime = RuntimeEnvironment(this, args) 

The command-line argument parsing is optional, and supports only:

• --version to print out the jar's build info (name, version, build)

• -f <filename> to specify a config file manually

• --validate to validate that your config file can be compiled

Your server object is used as the home jar of the build.properties file. Then the classpath is scanned to find that jar's home and the config files that are located nearby.

Quick Start

A good example server is created by the scala-bootstrapper project here: http://github.com/twitter/scala-bootstrapper

Define a server config class:

class MyServerConfig extends ServerConfig[MyServer] { var serverPort: Int = 9999 def apply(runtime: RuntimeEnvironment) = { new MyServer(serverPort) } } 

A ServerConfig class contains things you want to configure on your server, as vars, and an apply method that turns a RuntimeEnvironment into your server. ServerConfig is actually a helper for Config that adds logging configuration, sets up the optional admin HTTP server if it was configured, and registers your service with the ServiceTracker so that it will be shutdown when the admin port receives a shutdown command.

Next, make a simple config file for development:

import com.twitter.conversions.time._ import com.twitter.logging.config._ import com.twitter.ostrich.admin.config._ import com.example.config._ new MyServerConfig { serverPort = 9999 admin.httpPort = 9900 loggers = new LoggerConfig { level = Level.INFO handlers = new ConsoleHandlerConfig() } } 

The config file will be evaluated at runtime by this code in your Main class:

object Main { val log = Logger.get(getClass.getName) def main(args: Array[String]) { val runtime = RuntimeEnvironment(this, args) val server = runtime.loadRuntimeConfig[MyServer]() log.info("Starting my server!") try { server.start() } catch { case e: Exception => e.printStackTrace() log.error(e, "Unexpected exception: %s", e.getMessage) System.exit(0) } } } 

Your MyServer class should implement the Service interface so it can be started and shutdown. The runtime environment will find your config file and evaluate it, returning the MyServer object to you so you can start it. And you're set!

Stats API

The base trait of the stats API is StatsProvider, which defines methods for setting and getting each type of collected stat. The concrete implementation is StatsCollection, which stores them all in java concurrent hash maps.

To log or report stats, attach a StatsReporter to a StatsCollection. A StatsReporter keeps its own state, and resets that state each time it reports. You can attach multiple StatsReporters to track independent state without affecting the StatsCollection.

The simplest (and most common) pattern is to use the global singleton named Stats, like so:

import com.twitter.ostrich.stats.Stats Stats.incr("cache_misses") Stats.time("memcache_timing") { memcache.set(key, value) } 

Stat names can be any string, though conventionally they contain only letters, digits, underline (_), and dash (-), to make it easier for reporting.

You can immediately see any reported stats on the admin web server, if you've activated it, through the "stats" command:

curl localhost:PPPP/stats.txt 

(where PPPP is your configured admin port)

ServiceTracker

The global "shutdown" and "quiesce" commands work by talking to a global ServiceTracker object. This is just a set of running Service objects.

Each Service knows how to start and shutdown, so registering a service with the global ServiceTracker will cause it to be shutdown when the server as a whole is shutdown:

ServiceTracker.register(this) 

Some helper classes like BackgroundProcess and PeriodicBackgroundProcess implement Service, so they can be used to build simple background tasks that will be automatically shutdown when the server exits.

Admin web service

The easiest way to start the admin service is to construct an AdminServiceConfig with desired configuration, and call apply on it.

To reduce boilerplate in the common case of configuring a server with an admin port and logging, a helper trait called ServerConfig is defined with both:

var loggers: List[LoggerConfig] = Nil var admin = new AdminServiceConfig() 

The apply method on ServerConfig will create and start the admin service if a port is defined, and setup any configured logging.

You can also build an admin service directly from its config:

val adminConfig = new AdminServiceConfig { httpPort = 8888 statsNodes = new StatsConfig { reporters = new TimeSeriesCollectorConfig } } val runtime = RuntimeEnvironment(this, Nil) val admin = adminConfig()(runtime) 

If httpPort isn't set, the admin service won't start, and admin will be None. Otherwise it will be an Option[AdminHttpService].

statsNodes can attach a list of reporters to named stats collections. In the above example, a time-series collector is added to the global Stats object. This is used to provide the web graphs described below under "Web graphs".

Web/socket commands

Commands over the admin interface take the form of an HTTP "get" request:

GET /<command>[/<parameters...>][.<type>] 

which can be performed using 'curl' or 'wget':

$ curl http://localhost:PPPP/shutdown 

The result body may be json or plain-text, depending on . The default is json, but you can ask for text like so:

$ curl http://localhost:PPPP/stats.txt 

For simple commands like shutdown, the response body may simply be the JSON encoding of the string "ok". For others like stats, it may be a nested structure.

The commands are:

• ping

  Verify that the admin interface is working; server should say "pong" back.

• reload

  Reload the server config file for any services that support it (most do not).

• shutdown

  Immediately shutdown the server.

• quiesce

  Close any listening sockets, stop accepting new connections, and shutdown the server as soon as the last client connection is done.

• stats

  Dump server statistics as 4 groups: counters, gauges, metrics, and labels.

  • If the period query parameter is specified (e.g. /stats.json?period=10), a StatsListener is acquired for that time period, and all requests with this period value will receive the same stats values throughout that period.
  • Otherwise, if the namespace argument is provided (e.g. /stats.json?namespace=ganglia), a StatsListener is acquired for that namespace, and each request with this namespace value will reset the stats listener, effectively returning the delta since the prior request with that namespace. (See src/scripts/json_stats_fetcher.rb for an example.)
  • If neither period nor namespace parameters are specified, the main stats object will be fetched, returning non-differerential counters and metrics over the life-time of the process.

• server_info

  Dump server info (server name, version, build, and git revision).

• threads

  Dump stack traces and stats about each currently running thread.

• gc

  Force a garbage collection cycle.

Web graphs

If TimeSeriesCollector is attached to a stats collection, the web interface will include a small graph server that can be used to look at the last hour of data on collected stats.

The url

http://localhost:PPPP/graph/ 

(where PPPP is your admin httpPort) will give a list of currently-collected stats, and links to the current hourly graph for each stat. The graphs are generated in javascript using flot.

Profiling

If you're using heapster, you can generate a profile suitable for reading with google perftools

Example use:

curl -s 'localhost:9990/pprof/heap?pause=10' >| /tmp/prof 

This will result in a file that you can be read with pprof