§Understanding Play thread pools
Play framework is, from the bottom up, an asynchronous web framework. Streams are handled asynchronously using iteratees. Thread pools in Play are tuned to use fewer threads than in traditional web frameworks, since IO in play-core never blocks.
Because of this, if you plan to write blocking IO code, or code that could potentially do a lot of CPU intensive work, you need to know exactly which thread pool is bearing that workload, and you need to tune it accordingly. Doing blocking IO without taking this into account is likely to result in very poor performance from Play framework, for example, you may see only a few requests per second being handled, while CPU usage sits at 5%. In comparison, benchmarks on typical development hardware (eg, a MacBook Pro) have shown Play to be able to handle workloads in the hundreds or even thousands of requests per second without a sweat when tuned correctly.
§Knowing when you are blocking
The most common place where a typical Play application will block is when it’s talking to a database. Unfortunately, none of the major databases provide asynchronous database drivers for the JVM, so for most databases, your only option is to using blocking IO. A notable exception to this is ReactiveMongo, a driver for MongoDB that uses Play’s Iteratee library to talk to MongoDB.
Other cases when your code may block include:
- Using REST/WebService APIs through a 3rd party client library (ie, not using Play’s asynchronous WS API)
- Some messaging technologies only provide synchronous APIs to send messages
- When you open files or sockets directly yourself
- CPU intensive operations that block by virtue of the fact that they take a long time to execute
In general, if the API you are using returns Future
s, it is non-blocking, otherwise it is blocking.
Note that you may be tempted to therefore wrap your blocking code in Futures. This does not make it non-blocking, it just means the blocking will happen in a different thread. You still need to make sure that the thread pool that you are using has enough threads to handle the blocking.
In contrast, the following types of IO do not block:
- The Play WS API
- Asynchronous database drivers such as ReactiveMongo
- Sending/receiving messages to/from Akka actors
§Play’s thread pools
Play uses a number of different thread pools for different purposes:
- Netty boss/worker thread pools - These are used internally by Netty for handling Netty IO. An application’s code should never be executed by a thread in these thread pools.
- Play default thread pool - This is the thread pool in which all of your application code in Play Framework is executed. It is an Akka dispatcher, and is used by the application
ActorSystem
. It can be configured by configuring Akka, described below.
Note that in Play 2.4 several thread pools were combined together into the Play default thread pool.
§Using the default thread pool
All actions in Play Framework use the default thread pool. When doing certain asynchronous operations, for example, calling map
or flatMap
on a future, you may need to provide an implicit execution context to execute the given functions in. An execution context is basically another name for a ThreadPool
.
In most situations, the appropriate execution context to use will be the Play default thread pool. This can be used by importing it into your Scala source file:
import play.api.libs.concurrent.Execution.Implicits._
def someAsyncAction = Action.async {
import play.api.Play.current
WS.url("http://www.playframework.com").get().map { response =>
// This code block is executed in the imported default execution context
// which happens to be the same thread pool in which the outer block of
// code in this action will be executed.
Results.Ok("The response code was " + response.status)
}
}
§Configuring the Play default thread pool
The default thread pool can be configured using standard Akka configuration in application.conf
under the akka
namespace. Here is default configuration for Play’s thread pool:
akka {
fork-join-executor {
# Settings this to 1 instead of 3 seems to improve performance.
parallelism-factor = 1.0
parallelism-max = 24
# Setting this to LIFO changes the fork-join-executor
# to use a stack discipline for task scheduling. This usually
# improves throughput at the cost of possibly increasing
# latency and risking task starvation (which should be rare).
task-peeking-mode = LIFO
}
}
This configuration instructs Akka to create 1 thread per available processor, with a maximum of 24 threads in the pool.
You can also try the default Akka configuration:
akka {
fork-join-executor {
# The parallelism factor is used to determine thread pool size using the
# following formula: ceil(available processors * factor). Resulting size
# is then bounded by the parallelism-min and parallelism-max values.
parallelism-factor = 3.0
# Min number of threads to cap factor-based parallelism number to
parallelism-min = 8
# Max number of threads to cap factor-based parallelism number to
parallelism-max = 64
}
}
The full configuration options available to you can be found here.
§Using other thread pools
In certain circumstances, you may wish to dispatch work to other thread pools. This may include CPU heavy work, or IO work, such as database access. To do this, you should first create a ThreadPool
, this can be done easily in Scala:
object Contexts {
implicit val myExecutionContext: ExecutionContext = Akka.system.dispatchers.lookup("my-context")
}
In this case, we are using Akka to create the ExecutionContext
, but you could also easily create your own ExecutionContext
s using Java executors, or the Scala fork join thread pool, for example. To configure this Akka execution context, you can add the following configuration to your application.conf
:
my-context {
fork-join-executor {
parallelism-factor = 20.0
parallelism-max = 200
}
}
To use this execution context in Scala, you would simply use the scala Future
companion object function:
Future {
// Some blocking or expensive code here
}(Contexts.myExecutionContext)
or you could just use it implicitly:
import Contexts.myExecutionContext
Future {
// Some blocking or expensive code here
}
§Class loaders and thread locals
Class loaders and thread locals need special handling in a multithreaded environment such as a Play program.
§Application class loader
In a Play application the thread context class loader may not always be able to load application classes. You should explicitly use the application class loader to load classes.
- Java
-
Class myClass = Play.application().classloader().loadClass(myClassName);
- Scala
-
val myClass = Play.current.classloader.loadClass(myClassName)
Being explicit about loading classes is most important when running Play in development mode (using run
) rather than production mode. That’s because Play’s development mode uses multiple class loaders so that it can support automatic application reloading. Some of Play’s threads might be bound to a class loader that only knows about a subset of your application’s classes.
In some cases you may not be able to explicitly use the application classloader. This is sometimes the case when using third party libraries. In this case you may need to set the thread context class loader explicitly before you call the third party code. If you do, remember to restore the context class loader back to its previous value once you’ve finished calling the third party code.
§Java thread locals
Java code in Play uses a ThreadLocal
to find out about contextual information such as the current HTTP request. Scala code doesn’t need to use ThreadLocal
s because it can use implicit parameters to pass context instead. ThreadLocal
s are used in Java so that Java code can access contextual information without needing to pass context parameters everywhere.
Java ThreadLocal
s, along with the correct context ClassLoader
, are propagated automatically by ExecutionContextExecutor
objects provided through the HttpExecution
class. (An ExecutionContextExecutor
is both a Scala ExecutionContext
and a Java Executor
.) These special ExecutionContextExecutor
objects are automatically created and used by Java actions and Java Promise
methods. The default objects wrap the default user thread pool. If you want to do your own threading then you should use the HttpExecution
class’ helper methods to get an ExecutionContextExecutor
object yourself.
In the example below, a user thread pool is wrapped to create a new ExecutionContext
that propagates thread locals correctly.
import play.libs.HttpExecution;
import scala.concurrent.ExecutionContext;
public Promise<Result> index2() {
// Wrap an existing thread pool, using the context from the current thread
ExecutionContext myEc = HttpExecution.fromThread(myThreadPool);
return Promise.promise(() -> intensiveComputation(), myEc)
.map((Integer i) -> ok("Got result: " + i), myEc);
}
§Best practices
How you should best divide work in your application between different thread pools greatly depends on the types of work that your application is doing, and the control you want to have over how much of which work can be done in parallel. There is no one size fits all solution to the problem, and the best decision for you will come from understanding the blocking-IO requirements of your application and the implications they have on your thread pools. It may help to do load testing on your application to tune and verify your configuration.
Given the fact that JDBC is blocking thread pools can be sized to the # of connections available to a db pool assuming that the thread pool is used exclusively for database access. Any lesser amount of threads will not consume the number of connections available. Any more threads than the number of connections available could be wasteful given contention for the connections.
Below we outline a few common profiles that people may want to use in Play Framework:
§Pure asynchronous
In this case, you are doing no blocking IO in your application. Since you are never blocking, the default configuration of one thread per processor suits your use case perfectly, so no extra configuration needs to be done. The Play default execution context can be used in all cases.
§Highly synchronous
This profile matches that of a traditional synchronous IO based web framework, such as a Java servlet container. It uses large thread pools to handle blocking IO. It is useful for applications where most actions are doing database synchronous IO calls, such as accessing a database, and you don’t want or need control over concurrency for different types of work. This profile is the simplest for handling blocking IO.
In this profile, you would simply use the default execution context everywhere, but configure it to have a very large number of threads in its pool, like so:
akka {
akka.loggers = ["akka.event.slf4j.Slf4jLogger"]
loglevel = WARNING
actor {
default-dispatcher = {
fork-join-executor {
parallelism-min = 300
parallelism-max = 300
}
}
}
}
This profile is recommended for Java applications that do synchronous IO, since it is harder in Java to dispatch work to other threads.
Note that we use the same value for parallelism-min
and parallelism-max
. The reason is that the number of threads is defined by the following formulas :
base-nb-threads = nb-processors * parallelism-factor
parallelism-min <= actual-nb-threads <= parallelism-max
So if you don’t have enough available processors, you will never be able to reach the parallelism-max
setting.
§Many specific thread pools
This profile is for when you want to do a lot of synchronous IO, but you also want to control exactly how much of which types of operations your application does at once. In this profile, you would only do non blocking operations in the default execution context, and then dispatch blocking operations to different execution contexts for those specific operations.
In this case, you might create a number of different execution contexts for different types of operations, like this:
object Contexts {
implicit val simpleDbLookups: ExecutionContext = Akka.system.dispatchers.lookup("contexts.simple-db-lookups")
implicit val expensiveDbLookups: ExecutionContext = Akka.system.dispatchers.lookup("contexts.expensive-db-lookups")
implicit val dbWriteOperations: ExecutionContext = Akka.system.dispatchers.lookup("contexts.db-write-operations")
implicit val expensiveCpuOperations: ExecutionContext = Akka.system.dispatchers.lookup("contexts.expensive-cpu-operations")
}
These might then be configured like so:
contexts {
simple-db-lookups {
fork-join-executor {
parallelism-factor = 10.0
}
}
expensive-db-lookups {
fork-join-executor {
parallelism-max = 4
}
}
db-write-operations {
fork-join-executor {
parallelism-factor = 2.0
}
}
expensive-cpu-operations {
fork-join-executor {
parallelism-max = 2
}
}
}
Then in your code, you would create Future
s and pass the relevant ExecutionContext
for the type of work that Future
was doing.
Note: The configuration namespace can be chosen freely, as long as it matches the dispatcher ID passed to
Akka.system.dispatchers.lookup
.
§Few specific thread pools
This is a combination between the many specific thread pools and the highly synchronized profile. You would do most simple IO in the default execution context and set the number of threads there to be reasonably high (say 100), but then dispatch certain expensive operations to specific contexts, where you can limit the number of them that are done at one time.
Next: Configuring logging