org.apache.commons.lang3.concurrent.package-info Maven / Gradle / Ivy
Show all versions of commons-lang3 Show documentation
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/**
* Provides support classes for multi-threaded programming. This package is intended to be an extension to
* {@link java.util.concurrent}. These classes are thread-safe.
*
*
* A group of classes deals with the correct creation and initialization of objects that are accessed by multiple
* threads. All these classes implement the {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer} interface
* which provides just a single method:
*
*
* {@code
* public interface ConcurrentInitializer {
* T get() throws ConcurrentException;
* }
* }
*
*
*
* A {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer} produces an object. By calling the
* {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer#get() get()} method the object managed by the
* initializer can be obtained. There are different implementations of the interface available addressing various use
* cases:
*
*
*
* {@link org.apache.commons.lang3.concurrent.ConstantInitializer} is a very straightforward implementation of the
* {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer} interface: An instance is passed an object when it
* is constructed. In its {@code get()} method it simply returns this object. This is useful, for instance in unit tests
* or in cases when you want to pass a specific object to a component which expects a
* {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer}.
*
*
*
* The {@link org.apache.commons.lang3.concurrent.LazyInitializer} class can be used to defer the creation of an object
* until it is actually used. This makes sense, for instance, if the creation of the object is expensive and would slow
* down application startup or if the object is needed only for special executions.
* {@link org.apache.commons.lang3.concurrent.LazyInitializer} implements the double-check idiom for an instance
* field as discussed in Joshua Bloch's "Effective Java", 2nd edition, item 71. It uses volatile
* fields to reduce the amount of synchronization. Note that this idiom is appropriate for instance fields only. For
* static fields there are superior alternatives.
*
*
*
* We provide an example use case to demonstrate the usage of this class: A server application uses multiple worker
* threads to process client requests. If such a request causes a fatal error, an administrator is to be notified using
* a special messaging service. We assume that the creation of the messaging service is an expensive operation. So it
* should only be performed if an error actually occurs. Here is where
* {@link org.apache.commons.lang3.concurrent.LazyInitializer} comes into play. We create a specialized subclass for
* creating and initializing an instance of our messaging service.
* {@link org.apache.commons.lang3.concurrent.LazyInitializer} declares an abstract
* {@link org.apache.commons.lang3.concurrent.LazyInitializer#initialize() initialize()} method which we have to
* implement to create the messaging service object:
*
*
* {@code
* public class MessagingServiceInitializer extends LazyInitializer {
* protected MessagingService initialize() throws ConcurrentException {
* // Do all necessary steps to create and initialize the service object
* MessagingService service = ...
* return service;
* }
* }
* }
*
*
*
* Now each server thread is passed a reference to a shared instance of our new {@code MessagingServiceInitializer}
* class. The threads run in a loop processing client requests. If an error is detected, the messaging service is
* obtained from the initializer, and the administrator is notified:
*
*
* {@code
* public class ServerThread implements Runnable {
* // The initializer for obtaining the messaging service.
* private final ConcurrentInitializer initializer;
*
* public ServerThread(ConcurrentInitializer init) {
* initializer = init;
* }
*
* public void run() {
* while (true) {
* try {
* // wait for request
* // process request
* } catch (FatalServerException ex) {
* // get messaging service
* try {
* MessagingService svc = initializer.get();
* svc.notifyAdministrator(ex);
* } catch (ConcurrentException cex) {
* cex.printStackTrace();
* }
* }
* }
* }
* }
* }
*
*
*
* The {@link org.apache.commons.lang3.concurrent.AtomicInitializer} class is very similar to
* {@link org.apache.commons.lang3.concurrent.LazyInitializer}. It serves the same purpose: to defer the creation of an
* object until it is needed. The internal structure is also very similar. Again there is an abstract
* {@link org.apache.commons.lang3.concurrent.AtomicInitializer#initialize() initialize()} method which has to be
* implemented by concrete subclasses in order to create and initialize the managed object. Actually, in our example
* above we can turn the {@code MessagingServiceInitializer} into an atomic initializer by simply changing the
* extends declaration to refer to {@code AtomicInitializer} as super class.
*
*
*
* With {@link org.apache.commons.lang3.concurrent.AtomicSafeInitializer} there is yet another variant implementing the
* lazy initializing pattern. Its implementation is close to
* {@link org.apache.commons.lang3.concurrent.AtomicInitializer}; it also uses atomic variables internally and therefore
* does not need synchronization. The name "Safe" is derived from the fact that it implements an additional
* check which guarantees that the {@link org.apache.commons.lang3.concurrent.AtomicSafeInitializer#initialize()
* initialize()} method is called only once. So it behaves exactly in the same way as
* {@link org.apache.commons.lang3.concurrent.LazyInitializer}.
*
*
*
* Now, which one of the lazy initializer implementations should you use? First of all we have to state that is
* problematic to give general recommendations regarding the performance of these classes. The initializers make use of
* low-level functionality whose efficiency depends on multiple factors including the target platform and the number of
* concurrent threads. So developers should make their own benchmarks in scenarios close to their specific use cases.
* The following statements are rules of thumb which have to be verified in practice.
*
*
*
* {@link org.apache.commons.lang3.concurrent.AtomicInitializer} is probably the most efficient implementation due to
* its lack of synchronization and further checks. Its main drawback is that the {@code initialize()} method can be
* called multiple times. In cases where this is not an issue
* {@link org.apache.commons.lang3.concurrent.AtomicInitializer} is a good choice.
* {@link org.apache.commons.lang3.concurrent.AtomicSafeInitializer} and
* {@link org.apache.commons.lang3.concurrent.LazyInitializer} both guarantee that the initialization method is called
* only once. Because {@link org.apache.commons.lang3.concurrent.AtomicSafeInitializer} does not use synchronization it
* is probably slightly more efficient than {@link org.apache.commons.lang3.concurrent.LazyInitializer}, but the
* concrete numbers might depend on the level of concurrency.
*
*
*
* Another implementation of the {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer} interface is
* {@link org.apache.commons.lang3.concurrent.BackgroundInitializer}. It is again an abstract base class with an
* {@link org.apache.commons.lang3.concurrent.BackgroundInitializer#initialize() initialize()} method that has to be
* defined by concrete subclasses. The idea of {@link org.apache.commons.lang3.concurrent.BackgroundInitializer} is that
* it calls the {@code initialize()} method in a separate worker thread. An application creates a background initializer
* and starts it. Then it can continue with its work while the initializer runs in parallel. When the application needs
* the results of the initializer it calls its {@code get()} method. {@code get()} blocks until the initialization is
* complete. This is useful for instance at application startup. Here initialization steps (e.g. reading configuration
* files, opening a database connection, etc.) can be run in background threads while the application shows a splash
* screen and constructs its UI.
*
*
*
* As a concrete example consider an application that has to read the content of a URL - maybe a page with news - which
* is to be displayed to the user after login. Because loading the data over the network can take some time a
* specialized implementation of {@link org.apache.commons.lang3.concurrent.BackgroundInitializer} can be created for
* this purpose:
*
*
* {@code
* public class URLLoader extends BackgroundInitializer {
* // The URL to be loaded.
* private final URL url;
*
* public URLLoader(URL u) {
* url = u;
* }
*
* protected String initialize() throws ConcurrentException {
* try {
* InputStream in = url.openStream();
* // read content into string
* ...
* return content;
* } catch (IOException ioex) {
* throw new ConcurrentException(ioex);
* }
* }
* }
* }
*
*
*
* An application creates an instance of {@code URLLoader} and starts it. Then it can do other things. When it needs the
* content of the URL it calls the initializer's {@code get()} method:
*
*
*
*
* URL url = new URL("http://www.application-home-page.com/");
* URLLoader loader = new URLLoader(url);
* loader.start(); // this starts the background initialization
*
* // do other stuff
* ...
* // now obtain the content of the URL
* String content;
* try {
* content = loader.get(); // this may block
* } catch (ConcurrentException cex) {
* content = "Error when loading URL " + url;
* }
* // display content
*
*
*
*
* Related to {@link org.apache.commons.lang3.concurrent.BackgroundInitializer} is the
* {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer} class. As the name implies, this class can
* handle multiple initializations in parallel. The basic usage scenario is that a
* {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer} instance is created. Then an arbitrary number
* of {@link org.apache.commons.lang3.concurrent.BackgroundInitializer} objects is added using the
* {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer#addInitializer(String, BackgroundInitializer)}
* method. When adding an initializer a string has to be provided which is later used to obtain the result for this
* initializer. When all initializers have been added the
* {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer#start()} method is called. This starts
* processing of all initializers. Later the {@code get()} method can be called. It waits until all initializers have
* finished their initialization. {@code get()} returns an object of type
* {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer.MultiBackgroundInitializerResults}. This object
* provides information about all initializations that have been performed. It can be checked whether a specific
* initializer was successful or threw an exception. Of course, all initialization results can be queried.
*
*
*
* With {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer} we can extend our example to perform
* multiple initialization steps. Suppose that in addition to loading a web site we also want to create a JPA entity
* manager factory and read a configuration file. We assume that corresponding
* {@link org.apache.commons.lang3.concurrent.BackgroundInitializer} implementations exist. The following example
* fragment shows the usage of {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer} for this purpose:
*
*
*
* {@code
* MultiBackgroundInitializer initializer = new MultiBackgroundInitializer();
* initializer.addInitializer("url", new URLLoader(url));
* initializer.addInitializer("jpa", new JPAEMFInitializer());
* initializer.addInitializer("config", new ConfigurationInitializer());
* initializer.start(); // start background processing
*
* // do other interesting things in parallel
* ...
* // evaluate the results of background initialization
* MultiBackgroundInitializer.MultiBackgroundInitializerResults results =
* initializer.get();
* String urlContent = (String) results.getResultObject("url");
* EntityManagerFactory emf =
* (EntityManagerFactory) results.getResultObject("jpa");
* ...
* }
*
*
*
* The child initializers are added to the multi initializer and are assigned a unique name. The object returned by the
* {@code get()} method is then queried for the single results using these unique names.
*
*
*
* If background initializers - including {@link org.apache.commons.lang3.concurrent.MultiBackgroundInitializer} - are
* created using the standard constructor, they create their own {@link java.util.concurrent.ExecutorService} which is
* used behind the scenes to execute the worker tasks. It is also possible to pass in an
* {@link java.util.concurrent.ExecutorService} when the initializer is constructed. That way client code can configure
* the {@link java.util.concurrent.ExecutorService} according to its specific needs; for instance, the number of threads
* available could be limited.
*
*
* Utility Classes
*
*
* Another group of classes in the new {@code concurrent} package offers some generic functionality related to
* concurrency. There is the {@link org.apache.commons.lang3.concurrent.ConcurrentUtils} class with a bunch of static
* utility methods. One focus of this class is dealing with exceptions thrown by JDK classes. Many JDK classes of the
* executor framework throw exceptions of type {@link java.util.concurrent.ExecutionException} if something goes wrong.
* The root cause of these exceptions can also be a runtime exception or even an error. In typical Java programming you
* often do not want to deal with runtime exceptions directly; rather you let them fall through the hierarchy of method
* invocations until they reach a central exception handler. Checked exceptions in contrast are usually handled close to
* their occurrence. With {@link java.util.concurrent.ExecutionException} this principle is violated. Because it is a
* checked exception, an application is forced to handle it even if the cause is a runtime exception. So you typically
* have to inspect the cause of the {@link java.util.concurrent.ExecutionException} and test whether it is a checked
* exception which has to be handled. If this is not the case, the causing exception can be rethrown.
*
*
*
* The {@link org.apache.commons.lang3.concurrent.ConcurrentUtils#extractCause(java.util.concurrent.ExecutionException)}
* method does this work for you. It is passed an {@link java.util.concurrent.ExecutionException} and tests its root
* cause. If this is an error or a runtime exception, it is directly rethrown. Otherwise, an instance of
* {@link org.apache.commons.lang3.concurrent.ConcurrentException} is created and initialized with the root cause
* ({@link org.apache.commons.lang3.concurrent.ConcurrentException} is a new exception class in the
* {@code o.a.c.l.concurrent} package). So if you get such a
* {@link org.apache.commons.lang3.concurrent.ConcurrentException}, you can be sure that the original cause for the
* {@link java.util.concurrent.ExecutionException} was a checked exception. For users who prefer runtime exceptions in
* general there is also an
* {@link org.apache.commons.lang3.concurrent.ConcurrentUtils#extractCauseUnchecked(java.util.concurrent.ExecutionException)}
* method which behaves like {@code extractCause()}, but returns the unchecked exception
* {@link org.apache.commons.lang3.concurrent.ConcurrentRuntimeException} instead.
*
*
*
* In addition to the {@code extractCause()} methods there are corresponding
* {@link org.apache.commons.lang3.concurrent.ConcurrentUtils#handleCause(java.util.concurrent.ExecutionException)} and
* {@link org.apache.commons.lang3.concurrent.ConcurrentUtils#handleCauseUnchecked(java.util.concurrent.ExecutionException)}
* methods. These methods extract the cause of the passed in {@link java.util.concurrent.ExecutionException} and throw
* the resulting {@link org.apache.commons.lang3.concurrent.ConcurrentException} or
* {@link org.apache.commons.lang3.concurrent.ConcurrentRuntimeException}. This makes it easy to transform an
* {@link java.util.concurrent.ExecutionException} into a
* {@link org.apache.commons.lang3.concurrent.ConcurrentException} ignoring unchecked exceptions:
*
*
* {@code
* Future
*
*
* There is also some support for the concurrent initializers introduced in the last sub section. The
* {@code initialize()} method is passed a {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer} object and
* returns the object created by this initializer. It is null-safe. The {@code initializeUnchecked()} method works
* analogously, but a {@link org.apache.commons.lang3.concurrent.ConcurrentException} throws by the initializer is
* rethrown as a {@link org.apache.commons.lang3.concurrent.ConcurrentRuntimeException}. This is especially useful if
* the specific {@link org.apache.commons.lang3.concurrent.ConcurrentInitializer} does not throw checked exceptions.
* Using this method the code for requesting the object of an initializer becomes less verbose. The direct invocation
* looks as follows:
*
*
* {@code
* ConcurrentInitializer initializer = ...;
* try {
* MyClass obj = initializer.get();
* // do something with obj
* } catch (ConcurrentException cex) {
* // exception handling
* }
* }
*
*
*
* Using the {@link org.apache.commons.lang3.concurrent.ConcurrentUtils#initializeUnchecked(ConcurrentInitializer)}
* method, this becomes:
*
*
*
* {@code
* ConcurrentInitializer initializer = ...;
* MyClass obj = ConcurrentUtils.initializeUnchecked(initializer);
* // do something with obj
* }
*
*
*
* Another utility class deals with the creation of threads. When using the Executor framework new in JDK 1.5
* the developer usually does not have to care about creating threads; the executors create the threads they need on
* demand. However, sometimes it is desired to set some properties of the newly created worker threads. This is possible
* through the {@link java.util.concurrent.ThreadFactory} interface; an implementation of this interface has to be
* created and passed to an executor on creation time. Currently, the JDK does not provide an implementation of
* {@link java.util.concurrent.ThreadFactory}, so one has to start from scratch.
*
*
*
* With {@link org.apache.commons.lang3.concurrent.BasicThreadFactory} Commons Lang has an implementation of
* {@link java.util.concurrent.ThreadFactory} that works out of the box for many common use cases. For instance, it is
* possible to set a naming pattern for the new threads, set the daemon flag and a priority, or install a handler for
* uncaught exceptions. Instances of {@link org.apache.commons.lang3.concurrent.BasicThreadFactory} are created and
* configured using the nested {@link org.apache.commons.lang3.concurrent.BasicThreadFactory.Builder} class. The
* following example shows a typical usage scenario:
*
*
*
* {@code
* BasicThreadFactory factory = new BasicThreadFactory.Builder()
* .namingPattern("worker-thread-%d")
* .daemon(true)
* .uncaughtExceptionHandler(myHandler)
* .build();
* ExecutorService exec = Executors.newSingleThreadExecutor(factory);
* }
*
*
*
* The nested {@link org.apache.commons.lang3.concurrent.BasicThreadFactory.Builder} class defines some methods for
* configuring the new {@link org.apache.commons.lang3.concurrent.BasicThreadFactory} instance. Objects of this class
* are immutable, so these attributes cannot be changed later. The naming pattern is a string which can be passed to
* {@link String#format(java.util.Locale, String, Object...)}. The placeholder %d is replaced by an
* increasing counter value. An instance can wrap another {@link java.util.concurrent.ThreadFactory} implementation;
* this is achieved by calling the builder's
* {@link org.apache.commons.lang3.concurrent.BasicThreadFactory.Builder#wrappedFactory(java.util.concurrent.ThreadFactory)
* wrappedFactory(ThreadFactory)} method. This factory is then used for creating new threads; after that the specific
* attributes are applied to the new thread. If no wrapped factory is set, the default factory provided by the JDK is
* used.
*
*
* Synchronization objects
*
*
* The {@code concurrent} package also provides some support for specific synchronization problems with threads.
*
*
*
* {@link org.apache.commons.lang3.concurrent.TimedSemaphore} allows restricted access to a resource in a given time
* frame. Similar to a semaphore, a number of permits can be acquired. What is new is the fact that the permits
* available are related to a given time unit. For instance, the timed semaphore can be configured to allow 10 permits
* in a second. Now multiple threads access the semaphore and call its
* {@link org.apache.commons.lang3.concurrent.TimedSemaphore#acquire()} method. The semaphore keeps track about the
* number of granted permits in the current time frame. Only 10 calls are allowed; if there are further callers, they
* are blocked until the time frame (one second in this example) is over. Then all blocking threads are released, and
* the counter of available permits is reset to 0. So the game can start anew.
*
*
*
* What are use cases for {@link org.apache.commons.lang3.concurrent.TimedSemaphore}? One example is to artificially
* limit the load produced by multiple threads. Consider a batch application accessing a database to extract statistical
* data. The application runs multiple threads which issue database queries in parallel and perform some calculation on
* the results. If the database to be processed is huge and is also used by a production system, multiple factors have
* to be balanced: On one hand, the time required for the statistical evaluation should not take too long. Therefore you
* will probably use a larger number of threads because most of its life time a thread will just wait for the database
* to return query results. On the other hand, the load on the database generated by all these threads should be limited
* so that the responsiveness of the production system is not affected. With a
* {@link org.apache.commons.lang3.concurrent.TimedSemaphore} object this can be achieved. The semaphore can be
* configured to allow e.g. 100 queries per second. After these queries have been sent to the database the threads have
* to wait until the second is over - then they can query again. By fine-tuning the limit enforced by the semaphore a
* good balance between performance and database load can be established. It is even possible to chang? the number of
* available permits at runtime. So this number can be reduced during the typical working hours and increased at night.
*
*
*
* The following code examples demonstrate parts of the implementation of such a scenario. First the batch application
* has to create an instance of {@link org.apache.commons.lang3.concurrent.TimedSemaphore} and to initialize its
* properties with default values:
*
*
* {@code TimedSemaphore semaphore = new TimedSemaphore(1, TimeUnit.SECONDS, 100);}
*
*
* Here we specify that the semaphore should allow 100 permits in one second. This is effectively the limit of database
* queries per second in our example use case. Next the server threads issuing database queries and performing
* statistical operations can be initialized. They are passed a reference to the semaphore at creation time. Before they
* execute a query they have to acquire a permit.
*
*
*
*
* public class StatisticsTask implements Runnable {
* // The semaphore for limiting database load.
* private final TimedSemaphore semaphore;
*
* public StatisticsTask(TimedSemaphore sem, Connection con) {
* semaphore = sem;
* ...
* }
*
* //The main processing method. Executes queries and evaluates their results.
* public void run() {
* try {
* while (!isDone()) {
* semaphore.acquire(); // enforce the load limit
* executeAndEvaluateQuery();
* }
* } catch (InterruptedException iex) {
* // fall through
* }
* }
* }
*
*
*
*
* The important line here is the call to {@code semaphore.acquire()}. If the number of permits in the current time
* frame has not yet been reached, the call returns immediately. Otherwise, it blocks until the end of the time frame.
* The last piece missing is a scheduler service which adapts the number of permits allowed by the semaphore according
* to the time of day. We assume that this service is pretty simple and knows only two different time slots: working
* shift and night shift. The service is triggered periodically. It then determines the current time slot and configures
* the timed semaphore accordingly.
*
*
*
*
* public class SchedulerService {
* // The semaphore for limiting database load.
* private final TimedSemaphore semaphore;
* ...
*
* // Configures the timed semaphore based on the current time of day. This method is called periodically.
* public void configureTimedSemaphore() {
* int limit;
* if (isWorkshift()) {
* limit = 50; // low database load
* } else {
* limit = 250; // high database load
* }
*
* semaphore.setLimit(limit);
* }
* }
*
*
*
*
* With the {@link org.apache.commons.lang3.concurrent.TimedSemaphore#setLimit(int)} method the number of permits
* allowed for a time frame can be changed. There are some other methods for querying the internal state of a timed
* semaphore. Also some statistical data is available, e.g. the average number of {@code acquire()} calls per time
* frame. When a timed semaphore is no more needed, its {@code shutdown()} method has to be called.
*
*/
package org.apache.commons.lang3.concurrent;