You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@crunch.apache.org by bu...@apache.org on 2013/11/25 06:21:07 UTC
svn commit: r887976 - in /websites/staging/crunch/trunk/content: ./
intro.html
Author: buildbot
Date: Mon Nov 25 05:21:07 2013
New Revision: 887976
Log:
Staging update by buildbot for crunch
Modified:
websites/staging/crunch/trunk/content/ (props changed)
websites/staging/crunch/trunk/content/intro.html
Propchange: websites/staging/crunch/trunk/content/
------------------------------------------------------------------------------
--- cms:source-revision (original)
+++ cms:source-revision Mon Nov 25 05:21:07 2013
@@ -1 +1 @@
-1544354
+1545153
Modified: websites/staging/crunch/trunk/content/intro.html
==============================================================================
--- websites/staging/crunch/trunk/content/intro.html (original)
+++ websites/staging/crunch/trunk/content/intro.html Mon Nov 25 05:21:07 2013
@@ -250,7 +250,7 @@ return a type that has an associated obj
supports two serialization frameworks, called <em>type families</em>: one based on Hadoop's <code>Writable</code> interface, and another based on <code>Apache Avro</code>.
You can read more about how to work with Crunch's serialization libraries here. TODO</p>
<p>Because all of the core logic in our application is exposed via a single static method that operates on Crunch interfaces, we can use Crunch's
-in-memory API to test our business logic using a unit testing framework like JUnit. Let's look at an exampel unit test for the word count
+in-memory API to test our business logic using a unit testing framework like JUnit. Let's look at an example unit test for the word count
application:</p>
<div class="codehilite"><pre><span class="n">package</span> <span class="n">org</span><span class="p">.</span><span class="n">myorg</span><span class="p">;</span>
@@ -283,51 +283,55 @@ Collections Classes like <code>java.util
pipeline into the client and make decisions based on that data allows us to create sophisticated analytical
applications that can modify their downstream processing based on the results of upstream computations.</p>
<h3 id="data-model-and-operators">Data Model and Operators</h3>
-<p>The Java API is centered around three interfaces that represent distributed datasets: <code>PCollection<T></code>, <code>PTable<K, V></code>, and <code>PGroupedTable<K, V></code>.</p>
+<p>The Java API is centered around three interfaces that represent distributed datasets: <a href="apidocs/current/org/apache/crunch/PCollection.html">PCollection<T></a>,
+<a href="http://crunch.apache.org/apidocs/current/org/apache/crunch/PTable.html">PTable<K, V></a>, and <a href="apidocs/current/org/apache/crunch/PGroupedTable.html">PGroupedTable<K, V></a>.</p>
<p>A <code>PCollection<T></code> represents a distributed, unordered collection of elements of type T. For example, we represent a text file as a
-<code>PCollection<String></code> object. PCollection provides a method, <code>parallelDo</code>, that applies a <code>DoFn</code> to each element in a PCollection in parallel,
-and returns a new PCollection as its result. </p>
-<p>A <code>PTable<K, V></code> is a sub-interface of PCollection that represents a distributed, unordered multimap of its key type K to its value type V.
+<code>PCollection<String></code> object. <code>PCollection<T></code> provides a method, <code>parallelDo</code>, that applies a <a href="apidocs/current/org/apache/crunch/DoFn.html">DoFn<T, U></a>
+to each element in the <code>PCollection<T></code> in parallel, and returns an new <code>PCollection<U></code> as its result.</p>
+<p>A <code>PTable<K, V></code> is a sub-interface of <code>PCollection<Pair<K, V>></code> that represents a distributed, unordered multimap of its key type K to its value type V.
In addition to the parallelDo operation, PTable provides a <code>groupByKey</code> operation that aggregates all of the values in the PTable that
-have the same key into a single record. It is the groupByKey operation that triggers the sort phase of a MapReduce job.</p>
-<p>The result of a groupByKey operation is a <code>PGroupedTable<K, V></code> object, which is a distributed, sorted map of keys of type K to an Iterable
-collection of values of type V. In addition to parallelDo, the PGroupedTable provides a <code>combineValues</code> operation, which allows for
-a commutative and associative aggregation operator to be applied to the values of the PGroupedTable instance on both the map side and the
-reduce side of a MapReduce job.</p>
+have the same key into a single record. It is the groupByKey operation that triggers the sort phase of a MapReduce job. Developers can exercise
+fine-grained control over the number of reducers and the partitioning, grouping, and sorting strategies used during the shuffle by providing an instance
+of the <a href="apidocs/current/org/apache/crunch/GroupingOptions.html">GroupingOptions</a> class to the <code>groupByKey</code> function.</p>
+<p>The result of a groupByKey operation is a <code>PGroupedTable<K, V></code> object, which is a distributed, sorted map of keys of type K to an Iterable<V> that may
+be iterated over exactly once. In addition to <code>parallelDo</code> processing via DoFns, PGroupedTable provides a <code>combineValues</code> operation that allows a
+commutative and associative <a href="apidocs/current/org/apache/crunch/Aggregator.html">Aggregator<V></a> to be applied to the values of the PGroupedTable
+instance on both the map and reduce sides of the shuffle. A number of common <code>Aggregator<V></code> implementations are provided in the
+<a href="apidocs/current/org/apache/crunch/fn/Aggregators.html">Aggregators</a> class.</p>
<p>Finally, PCollection, PTable, and PGroupedTable all support a <code>union</code> operation, which takes a series of distinct PCollections that all have
the same data type and treats them as a single virtual PCollection.</p>
-<p>All of the other MapReduce patterns supported by the Crunch APIs (aggregations, joins, sorts, secondary sorts, and cogrouping) are all implemented
-in terms of these four primitives. The patterns themselves are defined in the <code>org.apache.crunch.lib</code> package and its children, and a few of
-the most common patterns have convenience functions defined on the PCollection and PTable interfaces. We will do a more detailed review of these
-patterns later in this document, but here are a few examples to get you started: TODO</p>
+<p>All of the other data transformation operations supported by the Crunch APIs (aggregations, joins, sorts, secondary sorts, and cogrouping) are implemented
+in terms of these four primitives. The patterns themselves are defined in the <a href="apidocs/current/org/apache/crunch/lib/package-summary.html">org.apache.crunch.lib</a>
+package and its children, and a few of of the most common patterns have convenience functions defined on the PCollection and PTable interfaces.</p>
<h3 id="writing-dofns">Writing DoFns</h3>
<p>DoFns represent the logical computations of your Crunch pipelines. They are designed to be easy to write, easy to test, and easy to deploy
within the context of a MapReduce job. Much of your work with the Crunch APIs will be writing DoFns, and so having a good understanding of
how to use them effectively is critical to crafting elegant and efficient pipelines.</p>
<h4 id="dofn-extends-serializable">DoFn extends Serializable</h4>
<p>The most important thing to remember about DoFns is that they all implement the <code>java.io.Serializable</code> interface, which means that all of the
-state information associated with a DoFn must also be serializable. There is an excellent overview of Java serializability here that is worth
-reviewing if you aren't familiar with Java's serializability model. TODO</p>
-<p>If your DoFn needs to work with a class that does not implement Serializable and cannot be modified (e.g., because it is defined in a third-party
+state information associated with a DoFn must also be serializable. There is an <a href="http://docs.oracle.com/javase/tutorial/jndi/objects/serial.html">excellent overview of Java serializability</a> that is worth reviewing if you aren't familiar with it already.</p>
+<p>If your DoFn needs to work with a class that does not implement Serializable and cannot be modified (for example, because it is defined in a third-party
library), you should use the <code>transient</code> keyword on that member variable so that serializing the DoFn won't fail if that object happens to be
defined. You can create an instance of the object during runtime using the <code>initialize</code> method described in the following section.</p>
<h4 id="runtime-processing-steps">Runtime Processing Steps</h4>
<p>After the Crunch runtime loads the serialized DoFns into its map and reduce tasks, the DoFns are executed on the input data via the following
sequence:</p>
-<h1 id="first-the-dofn-is-given-access-to-the-taskinputoutputcontext-implementation-for-the-current-task-this-allows-the-dofn-to-access-any">First, the DoFn is given access to the <code>TaskInputOutputContext</code> implementation for the current task. This allows the DoFn to access any</h1>
-<p>necessary configuration and runtime information needed before or during processing.</p>
-<h1 id="next-the-dofns-initialize-method-is-called-the-initialize-method-is-similar-to-the-setup-method-used-in-the-mapper-and-reducer-classes">Next, the DoFn's <code>initialize</code> method is called. The initialize method is similar to the <code>setup</code> method used in the Mapper and Reducer classes;</h1>
-<p>it is called before processing begins in order to enable any necessary initialization or configuration of the DoFn to be performed. For example,
-if we were making use of a non-serializable third-party library, we would create an instance of it here.</p>
-<h1 id="at-this-point-data-processing-begins-the-map-or-reduce-task-will-begin-passing-records-in-to-the-dofns-process-method-and-capturing-the">At this point, data processing begins. The map or reduce task will begin passing records in to the DoFn's <code>process</code> method, and capturing the</h1>
-<p>output of the process method into an <code>Emitter<T></code> that can either pass the data along to another DoFn for processing or serialize it as the output
-of the current processing stage.</p>
-<h1 id="finally-after-all-of-the-records-have-been-processed-the-void-cleanupemittert-emitter-method-is-called-on-each-dofn-the-cleanup-method">Finally, after all of the records have been processed, the <code>void cleanup(Emitter<T> emitter)</code> method is called on each DoFn. The cleanup method</h1>
-<p>has a dual purpose: it can be used to emit any state information that the DoFn wants to pass along to the next stage (for example, cleanup could
+<ol>
+<li>First, the DoFn is given access to the <code>TaskInputOutputContext</code> implementation for the current task. This allows the DoFn to access any
+necessary configuration and runtime information needed before or during processing.</li>
+<li>Next, the DoFn's <code>initialize</code> method is called. The initialize method is similar to the <code>setup</code> method used in the Mapper and Reducer classes;
+it is called before processing begins in order to enable any necessary initialization or configuration of the DoFn to be performed. For example,
+if we were making use of a non-serializable third-party library, we would create an instance of it here.</li>
+<li>At this point, data processing begins. The map or reduce task will begin passing records in to the DoFn's <code>process</code> method, and capturing the
+output of the process method into an <code>Emitter<T></code> that can either pass the data along to another DoFn for processing or serialize it as the output
+of the current processing stage.</li>
+<li>Finally, after all of the records have been processed, the <code>void cleanup(Emitter<T> emitter)</code> method is called on each DoFn. The cleanup method
+has a dual purpose: it can be used to emit any state information that the DoFn wants to pass along to the next stage (for example, cleanup could
be used to emit the sum of a list of numbers that was passed in to the DoFn's process method), as well as to release any resources or perform any
-other cleanup task that is appropriate once the job has finished executing.</p>
+other cleanup task that is appropriate once the job has finished executing.</li>
+</ol>
<h4 id="accessing-runtime-mapreduce-apis">Accessing Runtime MapReduce APIs</h4>
-<p>DoFns provide direct access to the <code>TaskInputOutputContext</code> object that is used within a given Map or Reduce task via the protected <code>getContext</code>
+<p>DoFns provide direct access to the <code>TaskInputOutputContext</code> object that is used within a given Map or Reduce task via the <code>getContext</code>
method. There are also a number of helper methods for working with the objects associated with the TaskInputOutputContext, including:</p>
<ul>
<li><code>getConfiguration()</code> for accessing the <code>Configuration</code> object that contains much of the detail about system and user-specific parameters for a
@@ -337,57 +341,86 @@ framework won't kill it,</li>
<li><code>setStatus(String status)</code> and <code>getStatus</code> for setting task status information, and</li>
<li><code>getTaskAttemptID()</code> for accessing the current <code>TaskAttemptID</code> information.</li>
</ul>
-<p>Crunch provides a number of helper methods, all named <code>increment</code> and having various signatures, for working with Hadoop Counters.
-There was a change in the Counters API from Hadoop 1.0 to Hadoop 2.0, and thus we do not recommend that you work with the <code>Counter</code> classes
-directly in your Crunch pipelines (the two <code>getCounter</code> methods that were defined in DoFn are both deprecated) so that you will not be
-required to recompile your job jars when you move from a Hadoop 1.x cluster to a Hadoop 2.x cluster.</p>
+<p>Crunch provides a number of helper methods for working with <a href="http://codingwiththomas.blogspot.com/2011/04/controlling-hadoop-job-recursion.html">Hadoop Counters</a>, all named <code>increment</code>. Counters are an incredibly useful way of keeping track of the state of long running data pipelines and detecting any exceptional conditions that
+occur during processing, and they are supported in both the MapReduce-based and in-memory Crunch pipeline contexts. You can retrive the value of the Counters
+in your client code at the end of a MapReduce pipeline by getting them from the <a href="apidocs/current/org/apache/crunch/PipelineResult.StageResult.html">StageResult</a>
+objects returned by Crunch at the end of a run.</p>
+<p>(Note that there was a change in the Counters API from Hadoop 1.0 to Hadoop 2.0, and thus we do not recommend that you work with the
+Counter classes directly in yoru Crunch pipelines (the two <code>getCounter</code> methods that were defined in DoFn are both deprecated) so that you will not be
+required to recompile your job jars when you move from a Hadoop 1.0 cluster to a Hadoop 2.0 cluster.)</p>
<h4 id="configuring-the-crunch-planner-and-mapreduce-jobs-with-dofns">Configuring the Crunch Planner and MapReduce Jobs with DoFns</h4>
<p>Although most of the DoFn methods are focused on runtime execution, there are a handful of methods that are used during the planning phase
before a pipeline is converted into MapReduce jobs. The first of these functions is <code>float scaleFactor()</code>, which should return a floating point
value greater than 0.0f. You can override the scaleFactor method in your custom DoFns in order to provide a hint to the Crunch planner about
-how much larger (or smaller) an input data set will become after passing through the process method. If the groupByKey method is called without
+how much larger (or smaller) an input data set will become after passing through the process method. If the <code>groupByKey</code> method is called without
an explicit number of reducers provided, the planner will try to guess how many reduce tasks should be used for the job based on the size of
-the input data, which is determined in part by using the scaleFactor results.</p>
+the input data, which is determined in part by using the result of calling the <code>scaleFactor</code> method on the DoFns in the processing path.</p>
<p>Sometimes, you may know that one of your DoFns has some unusual parameter settings that need to be specified on any job that includes that
DoFn as part of its processing. A DoFn can modify the Hadoop Configuration object that is associated with the MapReduce job it is assigned to
on the client before processing begins by overriding the <code>void configure(Configuration conf)</code> method. For example, you might know that the DoFn
will require extra memory settings to run, and so you could make sure that the value of the <code>mapred.child.java.opts</code> argument had a large enough
memory setting for the DoFn's needs before the job was launched on the cluster.</p>
-<h4 id="dofn-extensions-and-helper-classes">DoFn Extensions and Helper Classes</h4>
+<h4 id="common-dofn-patterns">Common DoFn Patterns</h4>
<p>The Crunch APIs contain a number of useful subclasses of DoFn that handle common data processing scenarios and are easier
-to write and test. The top-level <code>org.apache.crunch</code> package contains three of the most important specializations, which we will
-discuss now. Each of these specialized DoFn implementations has associated methods on the PCollection, PTable, and PGroupedTable
-interfaces to support these common data processing tasks.</p>
-<p>The simplest extension is the <code>FilterFn<T></code> class, which defines a single abstract method, <code>boolean accept(T input)</code>. The FilterFn can be applied
-to a <code>PCollection<T></code> by calling the <code>filter(FilterFn<T> fn)</code> method, and will return a new <code>PCollection<T></code> that only contains the elements
-of the input PCollection for which the accept method returned true. Note that the filter function does not include a PType argument in its
+to write and test. The top-level <a href="apidocs/current/org/apache/crunch/package-summary.html">org.apache.crunch</a> package contains three
+of the most important specializations, which we will discuss now. Each of these specialized DoFn implementations has associated methods
+on the PCollection, PTable, and PGroupedTable interfaces to support common data processing steps.</p>
+<p>The simplest extension is the <a href="apidocs/current/org/apache/crunch/FilterFn.html">FilterFn<T></a> class, which defines a single abstract method, <code>boolean accept(T input)</code>.
+The FilterFn can be applied to a <code>PCollection<T></code> by calling the <code>filter(FilterFn<T> fn)</code> method, and will return a new <code>PCollection<T></code> that only contains
+the elements of the input PCollection for which the accept method returned true. Note that the filter function does not include a PType argument in its
signature, because there is no change in the data type of the PCollection when the FilterFn is applied. It is possible to compose new FilterFn
-instances by combining multiple FilterFns together using the <code>and</code>, <code>or</code>, and <code>not</code> factory methods defined in the FilterFns helper class.</p>
-<p>The second extension is the <code>MapFn<S, T></code> class, which defines a single abstract method, <code>T map(S input)</code>. For simple transform tasks in which
-every input record will have exactly one output, it's easy to test a MapFn by verifying that a given input returns a given output. MapFns are
-also used by Crunch's data serialization libraries to map between serialized data types (such as Writables or Avro records) and POJOs.</p>
+instances by combining multiple FilterFns together using the <code>and</code>, <code>or</code>, and <code>not</code> factory methods defined in the
+<a href="apidocs/current/org/apache/crunch/fn/FilterFns.html">FilterFns</a> helper class.</p>
+<p>The second extension is the <a href="apidocs/current/org/apache/crunch/MapFn.html">MapFn<S, T></a> class, which defines a single abstract method, <code>T map(S input)</code>.
+For simple transform tasks in which every input record will have exactly one output, it's easy to test a MapFn by verifying that a given input returns a
+every input record will have exactly one output, it's easy to test a MapFn by verifying that a given input returns a given output.</p>
<p>MapFns are also used in specialized methods on the PCollection and PTable interfaces. <code>PCollection<V></code> defines the method
<code>PTable<K,V> by(MapFn<V, K> mapFn, PType<K> keyType)</code> that can be used to create a PTable from a PCollection by writing a
function that extracts the key (of type K) from the value (of type V) contained in the PCollection. The by function only requires that the PType of
the key be given and constructs a <code>PTableType<K, V></code> from the given key type and the PCollection's existing value type. <code>PTable<K, V></code>, in turn,
has methods <code>PTable<K1, V> mapKeys(MapFn<K, K1> mapFn)</code> and <code>PTable<K, V2> mapValues(MapFn<V, V2>)</code> that handle the common case of converting
just one of the paired values in a PTable instance from one type to another while leaving the other type the same.</p>
-<p>The final top-level extension to DoFn is the <code>CombineFn<K, V></code> class, which is used in conjunction with the <code>combineValues</code> method defined on the
-PGroupedTable interface. CombineFns are used to represent the associative operations that can be applied using the MapReduce Combiner concept in
-order to reduce the amount of data that is shipped over the network during the shuffle. The CombineFn extension is different from the FilterFn and
-MapFn classes in that it does not define an abstract method for handling data besides the default <code>process</code> method that any other DoFn would use;
-rather, extending the CombineFn class signals to the Crunch planner that the logic contained in this class satisfies the conditions required for use
-with the MapReduce combiner. Crunch supports many types of these associative patterns, such as sums, counts, and set unions, via the <code>Aggregator<V></code> interface,
-which is defined right alongside the CombineFn class in the top-level <code>org.apache.crunch</code> package. There are a number of implementations of the Aggregator
-interface defined via static factory methods in the <code>org.apache.crunch.fn.Aggregators</code> class.</p>
+<p>The final top-level extension to DoFn is the <a href="apidocs/current/org/apache/crunch/CombineFn.html">CombineFn<K, V></a> class, which is used in conjunction with
+the <code>combineValues</code> method defined on the PGroupedTable interface. CombineFns are used to represent the associative operations that can be applied using
+the MapReduce Combiner concept in order to reduce the amount data that is shipped over the network during a shuffle.</p>
+<p>The CombineFn extension is different from the FilterFn and MapFn classes in that it does not define an abstract method for handling data
+beyond the default <code>process</code> method that any other DoFn would use; rather, extending the CombineFn class signals to the Crunch planner that the logic
+contained in this class satisfies the conditions required for use with the MapReduce combiner.</p>
+<p>Crunch supports many types of these associative patterns, such as sums, counts, and set unions, via the <a href="apidocs/current/org/apache/crunch/Aggregator.html">Aggregator<V></a>
+interface, which is defined right alongside the CombineFn class in the top-level <code>org.apache.crunch</code> package. There are a number of implementations of the Aggregator
+interface defined via static factory methods in the <a href="apidocs/current/org/apache/crunch/fn/Aggregators.html">Aggregators</a> class.</p>
<h3 id="serializing-data-with-ptypes">Serializing Data with PTypes</h3>
<p>Why PTypes Are Necessary, the two type families, the core methods and tuples.</p>
<h4 id="extending-ptypes">Extending PTypes</h4>
-<h3 id="reading-data-sources">Reading Data: Sources</h3>
-<h3 id="writing-data-targets">Writing Data: Targets</h3>
+<p>The simplest way to create a new <code>PType<T></code> for a data object is to create a <em>derived</em> PType from one of the built-in PTypes for the Avro
+and Writable type families. If we have a base <code>PType<S></code>, we can create a derived <code>PType<T></code> by implementing an input <code>MapFn<S, T></code> and an
+output <code>MapFn<T, S></code> and then calling <code>PTypeFamily.derived(Class<T>, MapFn<S, T> in, MapFn<T, S> out, PType<S> base)</code>, which will return
+a new <code>PType<T></code>. There are examples of derived PTypes in the <a href="apidocs/current/org/apache/crunch/types/PTypes.html">PTypes</a> class, including
+serialization support for protocol buffers, Thrift records, Java Enums, BigInteger, and UUIDs.</p>
+<h3 id="reading-and-writing-data-sources-targets-and-sourcetargets">Reading and Writing Data: Sources, Targets, and SourceTargets</h3>
+<p>MapReduce developers are familiar with the <code>InputFormat<K, V></code> and <code>OutputFormat<K, V></code> classes for reading and writing data during
+MapReduce processing. Crunch has the analogous concepts of a <code>Source<T></code> for reading data and a <code>Target</code> for writing data. For data
+sources that may be treated as both the output of one pipeline phase and the input to another, Crunch has a <code>SourceTarget<T></code> interface
+that combines the functionality of both <code>Source<T></code> and <code>Target</code>.</p>
+<p>Sources and Targets provide several useful extensions to the functionality provided by InputFormat and OutputFormat. First, a Source can
+encapsulate an InputFormat as well as any special Configuration settings that are needed by that InputFormat. For example, the
+<code>AvroInputFormat</code> needs to know the Avro schema of the input Avro file and expects to find that schema associated with the "avro.schema" key
+in the <code>Configuration</code> object for a pipeline. But if you need to read multiple Avro files, each with its own schema, during a single MapReduce
+job, you need a way of ensuring that the different schemas for each file do not all overwrite the "avro.schema" key in the shared
+<code>Configuration</code> object. Crunch's <code>Source<T></code> allows you to specify a set of key-value entries that need to be set in the <code>Configuration</code>
+before a particular input is read in a way that prevents them from conflicting with each other, while the Target interface provides the same
+functionality for OutputFormats.</p>
+<p>The <code>Source<T></code> interface has two useful extensions. The first is <code>TableSource<K, V></code> which extends <code>Source<Pair<K, V>></code> and can be
+used to read in a <code>PTable<K, V></code> instance instead of a <code>PCollection<Pair<K, V>></code> instance. The second extension is <code>ReadableSource<T></code>, which
+declares a <code>Iterable<T> read(Configuration conf)</code> method that allows the contents of the Source to be read directly, either into the client
+or into a DoFn implementation that can use the data read from the source to perform additional transforms on the main input data that is
+processed using the DoFn's <code>process</code> method (this is how Crunch supports mapside-join operations.)</p>
+<p>Support for the most common Source, Target, and SourceTarget implementations are provided by the factory functions declared in the
+<a href="apidocs/current/org/apache/crunch/io/From.html">From</a> (Sources), <a href="apidocs/current/org/apache/crunch/io/To.html">To</a> (Targets), and
+<a href="apidocs/current/org/apache/crunch/io/At.html">At</a> (SourceTargets) classes in the <a href="apidocs/current/org/apache/crunch/io/package-summary.html">org.apache.crunch.io</a>
+package.</p>
<h3 id="pipeline-building-and-execution">Pipeline Building and Execution</h3>
<h4 id="creating-a-new-crunch-pipeline">Creating A New Crunch Pipeline</h4>
-<p>Section here on Configuration of pipelines.</p>
<h4 id="managing-pipeline-execution-and-cleanup">Managing Pipeline Execution and Cleanup</h4>
<h2 id="more-information">More Information</h2>
<p><a href="pipelines.html">Writing Your Own Pipelines</a></p>