Cascading 2.1 User Guide

Cascading is a data processing API and processing query planner used for defining, sharing, and executing data-processing workflows on a single computing node or distributed computing cluster. On a single node, Cascading's "local mode" can be used to efficiently test code and process local files before being deployed on a cluster. On a distributed computing cluster using Apache Hadoop platform, Cascading adds an abstraction layer over the Hadoop API, greatly simplifying Hadoop application development, job creation, and job scheduling.

From the Hadoop website, it “is a software platform that lets one easily write and run applications that process vast amounts of data”. Hadoop does this by providing a storage layer that holds vast amounts of data, and an execution layer that runs an application in parallel across the cluster, using coordinated subsets of the stored data.

The storage layer, called the Hadoop File System (HDFS), looks like a single storage volume that has been optimized for many concurrent serialized reads of large data files - where "large" might be measured in gigabytes or petabytes. However, it does have limitations. For example, random access to the data is not really possible in an efficient manner. And Hadoop only supports a single writer for output. But this limit helps make Hadoop very performant and reliable, in part because it allows for the data to be replicated across the cluster, reducing the chance of data loss.

The execution layer, called MapReduce, relies on a divide-and-conquer strategy to manage massive data sets and computing processes. Explaining MapReduce is beyond the scope of this document, but its complexity, and the difficulty of creating real-world applications against it, are the chief driving force behind the creation of Cascading.

Hadoop, according to its documentation, can be configured to run in three modes: standalone mode (i.e., on the local computer, useful for testing and debugging), pseudo-distributed mode (i.e., on an emulated "cluster" of one computer, useful for testing), and fully-distributed mode (on a full cluster, for production purposes). The pseudo-distributed mode does not add value for most purposes, and will not be discussed further. Cascading itself can run locally or on the Hadoop platform, where Hadoop itself may be in standalone or distributed mode. The primary difference between these two platforms is that, when Cascading is running locally, it makes no use of Hadoop APIs and performs all of its work in memory, allowing it to be very fast - but consequently not as robust or scalable as when it is running on the Hadoop platform.

Apache Hadoop is an Open Source Apache project and is freely available. It can be downloaded from the Hadoop website: http://hadoop.apache.org/core/.

The Cascading processing model is based on a metaphor of pipes (data streams) and filters (data operations). Thus the Cascading API allows the developer to assemble pipe assemblies that split, merge, group, or join streams of data while applying operations to each data record or groups of records.

In Cascading, we call a data record a tuple, a simple chain of pipes without forks or merges a branch, an interconnected set of pipe branches a pipe assembly, and a series of tuples passing through a pipe branch or assembly a tuple stream.

Pipe assemblies are specified independently of the data source they are to process. So before a pipe assembly can be executed, it must be bound to taps, i.e., data sources and sinks. The result of binding one or more pipe assemblies to taps is a flow, which is executed on a computer or cluster using the Hadoop framework.

Multiple flows can be grouped together and executed as a single process. In this context, if one flow depends on the output of another, it is not executed until all of its data dependencies are satisfied. Such a collection of flows is called a cascade.

Pipe assemblies define what work should be done against tuple streams, which are read from tap sources and written to tap sinks. The work performed on the data stream may include actions such as filtering, transforming, organizing, and calculating. Pipe assemblies may use multiple sources and multiple sinks, and may define splits, merges, and joins to manipulate the tuple streams.

Pipe Assembly Workflow

Pipe assemblies are created by chaining cascading.pipe.Pipe classes and subclasses together. Chaining is accomplished by passing the previous Pipe instances to the constructor of the next Pipe instance.

The following example demonstrates this type of chaining. It creates two pipes - a "left-hand side" (lhs) and a "right-hand side" (rhs) - and performs some processing on them both, using the Each pipe. Then it joins the two pipes into one, using the CoGroup pipe, and performs several operations on the joined pipe using Every and GroupBy. The specific operations performed are not important in the example; the point is to show the general flow of the data streams. The diagram after the example gives a visual representation of the workflow.

Example 3.1. Chaining Pipes

// the "left hand side" assembly head
Pipe lhs = new Pipe( "lhs" );

lhs = new Each( lhs, new SomeFunction() );
lhs = new Each( lhs, new SomeFilter() );

// the "right hand side" assembly head
Pipe rhs = new Pipe( "rhs" );

rhs = new Each( rhs, new SomeFunction() );

// joins the lhs and rhs
Pipe join = new CoGroup( lhs, rhs );

join = new Every( join, new SomeAggregator() );

join = new GroupBy( join );

join = new Every( join, new SomeAggregator() );

// the tail of the assembly
join = new Each( join, new SomeFunction() );

The following diagram is a visual representation of the example above.

The code for the sample pipe assembly above, Chaining Pipes, consists almost entirely of a series of Pipe constructors. This section describes the various Pipe classes in detail. The base class cascading.pipe.Pipe and its subclasses are shown in the diagram below.

The Each and Every Pipes

The Each and Every pipes perform operations on tuple data - for instance, perform a search-and-replace on tuple contents, filter out some of the tuples based on their contents, or count the number of tuples in a stream that share a common field value.

Here is the syntax for these pipes:

new Each( previousPipe, argumentSelector, operation, outputSelector )

new Every( previousPipe, argumentSelector, operation, outputSelector )

Both types take four arguments:

• a Pipe instance

• an argument selector

• an Operation instance

• an output selector on the constructor (selectors here are Fields instances)

The key difference between Each and Every is that the Each operates on individual tuples, and Every operates on groups of tuples emitted by GroupBy or CoGroup. This affects the kind of operations that these two pipes can perform, and the kind of output they produce as a result.

The Each pipe applies operations that are subclasses of Functions and Filters (described in the Javadoc). For example, using Each you can parse lines from a logfile into their constituent fields, filter out all lines except the HTTP GET requests, and replace the timestring fields with date fields.

Similarly, since the Every pipe works on tuple groups (the output of a GroupBy or CoGroup pipe), it applies operations that are subclasses of Aggregators and Buffers. For example, you could use GroupBy to group the output of the above Each pipe by date, then use an Every pipe to count the GET requests per date. The pipe would then emit the operation results as the date and count for each group.

In the syntax shown at the start of this section, the argument selector specifies fields from the input tuple to use as input values. If the argument selector is not specified, the whole input tuple (Fields.ALL) is passed to the operation as a set of argument values.

Most Operation subclasses declare result fields (shown as "declared fields" in the diagram). The output selector specifies the fields of the output Tuple from the fields of the input Tuple and the operation result. This new output Tuple becomes the input Tuple to the next pipe in the pipe assembly. If the output selector is Fields.ALL, the output is the input Tuple plus the operation result, merged into a single Tuple.

Note that it's possible for a Function or Aggregator to return more than one output Tuple per input Tuple. In this case, the input tuple is duplicated as many times as necessary to create the necessary output tuples. This is similar to the reiteration of values that happens during a join. If a function is designed to always emit three result tuples for every input tuple, each of the three outgoing tuples will consist of the selected input tuple values plus one of the three sets of function result values.

If the result selector is not specified for an Each pipe performing a Functions operation, the operation results are returned by default (Fields.RESULTS), discarding the input tuple values in the tuple stream. (This is not true of Filters , which either discard the input tuple or return it intact, and thus do not use an output selector.)

For the Every pipe, the Aggregator results are appended to the input Tuple (Fields.ALL) by default.

Note that the Every pipe associates Aggregator results with the current group Tuple (the unique keys currently being grouped on). For example, if you are grouping on the field "department" and counting the number of "names" grouped by that department, the resulting output Fields will be ["department","num_employees"].

If you are also adding up the salaries associated with each "name" in each "department", the output Fields will be ["department","num_employees","total_salaries"].

This is only true for chains of Aggregator Operations - you are not allowed to chain Buffer operations, as explained below.

When the Every pipe is used with a Buffer operation, instead of an Aggregator, the behavior is different. Instead of being associated with the current grouping tuple, the operation results are associated with the current values tuple. This is analogous to how an Each pipe works with a Function. This approach may seem slightly unintuitive, but provides much more flexibility. To put it another way, the results of the buffer operation are not appended to the current keys being grouped on. It is up to the buffer to emit them if they are relevant. It is also possible for a Buffer to emit more than one result Tuple per unique grouping. That is, a Buffer may or may not emulate an Aggregator, where an Aggregator is just a special optimized case of a Buffer.

For more information on how operations process fields, see Operations and Field-processing .

Pipe merge = new Merge( lhs, rhs );

Fields groupFields = new Fields( "group1", "group2" );
Fields sortFields = new Fields( "value1", "value2" );
Pipe groupBy = new GroupBy( assembly, groupFields, sortFields );

Fields groupFields = new Fields( "group1", "group2" );
Fields sortFields = new Fields( "value1", "value2" );

sortFields.setComparator( "value1", Collections.reverseOrder() );

Pipe groupBy = new GroupBy( assembly, groupFields, sortFields );

Fields groupFields = new Fields( "year", "month", "day" );
Fields sortFields = new Fields( "hour", "minute", "second" );

sortFields.setComparators(
  Collections.reverseOrder(),   // hour
  Collections.reverseOrder(),   // minute
  Collections.reverseOrder() ); // second

Pipe groupBy = new GroupBy( assembly, groupFields, sortFields );

CoGroup

The CoGroup pipe is similar to GroupBy, but instead of a merge, performs a join. That is, CoGroup accepts two or more input streams and groups them on one or more specified keys, and performs a join operation on equal key values, similar to a SQL join.

The output stream contains all the fields from all the input streams.

As with SQL, the join can be inner, outer, left, or right. Self-joins are permitted, as well as mixed joins (for three or more streams) and custom joins. Null fields in the input streams become corresponding null fields in the output stream.

Since the output is grouped, it is suitable for the Every pipe, which performs Aggregator and Buffer operations - such as counting, totalling, or averaging groups of tuples that have a common value (e.g., the same date).

The output stream is sorted by the natural order of the grouping fields. To control this order, at least the first groupingFields value given should be an instance of Fields containing Comparator instances for the appropriate fields. This allows fine-grained control of the sort grouping order.

Field names

In a join operation, all the field names used in any of the input tuples must be unique; duplicate field names are not allowed. If the names overlap there is a collision, as shown in the following diagram.

In this figure, two streams are to be joined on the "url" field, resulting in a new Tuple that contains fields from the two input tuples. However, the resulting tuple would include two fields with the same name ("url"), which is unworkable. To handle the conflict, developers can use the declaredFields argument (described in the Javadoc) to declare unique field names for the output tuple, as in the following example.

Example 3.6. Joining Two Tuple Streams with Duplicate Field Names

Fields common = new Fields( "url" );
Fields declared = new Fields(
  "url1", "word", "wd_count", "url2", "sentence", "snt_count"
);
Pipe join =
  new CoGroup( lhs, common, rhs, common, declared, new InnerJoin() );

This revised figure demonstrates the use of declared field names to prevent a planning failure.

It might seem preferable for Cascading to automatically recognize the duplication and simply merge the identically-named fields, saving effort for the developer. However, consider the case of an outer type join in which one field (or set of fields used for the join) for a given join side happens to be null. Discarding one of the duplicate fields would lose this information.

Further, the internal implementation relies on field position, not field names, when reading tuples; the field names are a device for the developer. This approach allows the behavior of the CoGroup to be deterministic and predictable.

The Joiner class

In the example above, we explicitly specified a Joiner class (InnerJoin) to perform a join on our data. There are five Joiner subclasses, as shown in this diagram.

In CoGroup, the join is performed after all the input streams are first co-grouped by their common keys. Cascading must create a "bag" of data for every grouping in the input streams, consisting of all the Tuple instances associated with that grouping.

It's already been mentioned that joins in Cascading are analogous to joins in SQL. The most commonly-used type of join is the inner join, the default in CoGroup. An inner join tries to match each Tuple on the "lhs" with every Tuple on the "rhs", based on matching field values. With an inner join, if either side has no tuples for a given value, no tuples are joined. An outer join, conversely, allows for either side to be empty and simply substitutes a Tuple containing null values for the non-existent tuple.

This sample data is used in the discussion below to explain and compare the different types of join:

LHS = [0,a] [1,b] [2,c]
            RHS = [0,A] [2,C] [3,D]

In each join type below, the values are joined on the first tuple position (the join key), a numeric value. Note that, when Cascading joins tuples, the resulting Tuple contains all the incoming values from in incoming tuple streams, and does not discard the duplicate key fields. As mentioned above, on outer joins where there is no equivalent key in the alternate stream, null values are used.

For example using the data above, the result Tuple of an "inner" join with join key value of 2 would be [2,c,2,C]. The result Tuple of an "outer" join with join key value of 1 would be [1,b,null,null].

InnerJoin

An inner join only returns a joined Tuple if neither bag for the join key is empty.

[0,a,0,A] [2,c,2,C]

OuterJoin

An outer join performs a join if one bag (left or right) for the join key is empty, or if neither bag is empty.

[0,a,0,A] [1,b,null,null] [2,c,2,C] [null,null,3,D]

LeftJoin

A left join can also be stated as a left inner and right outer join, where it is acceptable for the right bag to be empty (but not the left).

[0,a,0,A] [1,b,null,null] [2,c,2,C]

RightJoin

A right join can also be stated as a left outer and right inner join, where it is acceptable for the left bag to be empty (but not the right).

[0,a,0,A] [2,c,2,C] [null,null,3,D]

MixedJoin

A mixed join is where 3 or more tuple streams are joined, using a small Boolean array to specify each of the join types to use. For more information, see the cascading.pipe.cogroup.MixedJoin class in the Javadoc.

Custom

Developers can subclass the cascading.pipe.cogroup.Joiner class to create custom join operations.

Scaling

CoGroup attempts to store the entire current unique keys tuple "bag" from the right-hand stream in memory for rapid joining to the left-hand stream. If the bag is very large, it may exceed a configurable threshold and be spilled to disk, reducing performance and potentially causing a memory error (if the threshold value is too large). Thus it's usually best to put the stream with the largest groupings on the left-hand side and, if necessary, adjust the spill threshold as described in the Javadoc.

Fields lhsFields = new Fields( "fieldA", "fieldB" );
Fields rhsFields = new Fields( "fieldC", "fieldD" );
Pipe join =
  new HashJoin( lhs, lhsFields, rhs, rhsFields, new InnerJoin() );

Pipe join =
  new HashJoin( lhs, common, rhs, common, declared, new InnerJoin() );

SpillableProps props = SpillableProps.spillableProps()
  .setCompressSpill( true )
  .setMapSpillThreshold( 50 * 1000 );

props.setProperties( join.getConfigDef(), ConfigDef.Mode.REPLACE );

Pipe join =
  new HashJoin( lhs, common, rhs, common, declared, new InnerJoin() );

SpillableProps props = SpillableProps.spillableProps()
  .setCompressSpill( true )
  .setMapSpillThreshold( 50 * 1000 );

props.setProperties( join.getStepConfigDef(), ConfigDef.Mode.DEFAULT );

Cascading supports pluggable planners that allow it to execute on differing platforms. Planners are invoked by an associated FlowConnector subclass. Currently, only two planners are provided, as described below:

Cascading's support for pluggable planners allows a pipe assembly to be executed on an arbitrary platform, using platform-specific Tap and Scheme classes that hide the platform-related I/O details from the developer. For example, Hadoop uses org.apache.hadoop.mapred.InputFormat to read data, but local mode is happy with a java.io.FileInputStream. This detail is hidden from developers unless they are creating custom Tap and Scheme classes.

All input data comes in from, and all output data goes out to, some instance of cascading.tap.Tap. A tap represents a data resource - such as a file on the local file system, on a Hadoop distributed file system, or on Amazon S3. A tap can be read from, which makes it a source, or written to, which makes it a sink. Or, more commonly, taps act as both sinks and sources when shared between flows.

The platform on which your application is running (Cascading local or Hadoop) determines which specific classes you can use. Details are provided in the sections below.

Tap tap = new Hfs( new TextLine( new Fields( "line" ) ), path );

Tap tap =
  new Hfs( new TextLine( new Fields( "line" ) ), path, SinkMode.REPLACE );

All applications created with Cascading read data from one or more sources, process it, then write data to one or more sinks. This is done via the various Tap classes, where each class abstracts different types of back-end systems that store data as files, tables, blobs, and so on. But in order to sink data, some systems require that the resource (e.g., a file) not exist before processing thus must be removed (deleted) before the processing can begin. Other systems may allow for appending or updating of a resource (typical with database tables).

When creating a new Tap instance, a SinkMode may be provided so that the Tap will know how to handle any existing resources. Note that not all Taps support all SinkMode values - for example, Hadoop does not support appends (updates) from a MapReduce job.

The available SinkModes are:

Note that Cascading itself only uses these labels internally to know when to automatically call deleteResource() on the Tap or to leave the Tap alone. It is up the the Tap implementation to actually perform a write or update when processing starts. Thus, when start() or complete() is called on a Flow, any sink Tap labeled SinkMode.REPLACE will have its deleteResource() method called.

Conversely, if a Flow is in a Cascade and the Tap is set to SinkMode.KEEP or SinkMode.REPLACE, deleteResource() will be called if and only if the sink is stale (i.e., older than the source). This allows a Cascade to behave like a "make" or "ant" build file, only running Flows that should be run. For more information, see Skipping Flows.

It's also important to understand how Hadoop deals with directories. By default, Hadoop cannot source data from directories with nested sub-directories, and it cannot write to directories that already exist. However, the good news is that you can simply point the Hfs tap to a directory of data files, and they are all used as input - there's no need to enumerate each individual file into a MultiSourceTap. If there are nested directories, use GlobHfs.

Cascading applications can perform complex manipulation or "field algebra" on the fields stored in tuples, using Fields sets, a feature of the Fields class that provides a sort of wildcard tool for referencing sets of field values.

These predefined Fields sets are constant values on the Fields class. They can be used in many places where the Fields class is expected. They are:

The chart below shows common ways to merge input and result fields for the desired output fields. A few minutes with this chart may help clarify the discussion of fields, tuples, and pipes. Also see Each and Every Pipes for details on the different columns and their relationships to the Each and Every pipes and Functions, Aggregators, and Buffers.

When pipe assemblies are bound to source and sink taps, a Flow is created. Flows are executable in the sense that, once they are created, they can be started and will execute on the specified platform. If the Hadoop platform is specified, the Flow will execute on a Hadoop cluster.

A Flow is essentially a data processing pipeline that reads data from sources, processes the data as defined by the pipe assembly, and writes data to the sinks. Input source data does not need to exist at the time the Flow is created, but it must exist by the time the Flow is executed (unless it is executed as part of a Cascade - see Cascades for more on this).

The most common pattern is to create a Flow from an existing pipe assembly. But there are cases where a MapReduce job (if running on Hadoop) has already been created, and it makes sense to encapsulate it in a Flow class so that it may participate in a Cascade and be scheduled with other Flow instances. Alternatively, via the Riffle annotations, third-party applications can participate in a Cascade, and complex algorithms that result in iterative Flow executions can be encapsulated as a single Flow. All patterns are covered here.

HadoopFlowConnector flowConnector = new HadoopFlowConnector();

Flow flow =
  flowConnector.connect( "flow-name", source, sink, pipe );

// the "left hand side" assembly head
Pipe lhs = new Pipe( "lhs" );

lhs = new Each( lhs, new SomeFunction() );
lhs = new Each( lhs, new SomeFilter() );

// the "right hand side" assembly head
Pipe rhs = new Pipe( "rhs" );

rhs = new Each( rhs, new SomeFunction() );

// joins the lhs and rhs
Pipe join = new CoGroup( lhs, rhs );

join = new Every( join, new SomeAggregator() );

Pipe groupBy = new GroupBy( join );

groupBy = new Every( groupBy, new SomeAggregator() );

// the tail of the assembly
groupBy = new Each( groupBy, new SomeFunction() );

Tap lhsSource = new Hfs( new TextLine(), "lhs.txt" );
Tap rhsSource = new Hfs( new TextLine(), "rhs.txt" );

Tap sink = new Hfs( new TextLine(), "output" );

FlowDef flowDef = new FlowDef()
  .setName( "flow-name" )
  .addSource( rhs, rhsSource )
  .addSource( lhs, lhsSource )
  .addTailSink( groupBy, sink );

Flow flow = new HadoopFlowConnector().connect( flowDef );

Properties properties = new Properties();

// pass in the class name of your application
// this will find the parent jar at runtime
AppProps.setApplicationJarClass( properties, Main.class );

// ALTERNATIVELY ...

// pass in the path to the parent jar
AppProps.setApplicationJarPath( properties, pathToJar );


// pass properties to the connector
FlowConnector flowConnector = new HadoopFlowConnector( properties );

A Cascade allows multiple Flow instances to be executed as a single logical unit. If there are dependencies between the Flows, they are executed in the correct order. Further, Cascades act like Ant builds or Unix make files - that is, a Cascade only executes Flows that have stale sinks (i.e., output data that is older than the input data). For more on this, see Skipping Flows.

CascadeConnector connector = new CascadeConnector();
Cascade cascade = connector.connect( flowFirst, flowSecond, flowThird );

When passing Flows to the CascadeConnector, order is not important. The CascadeConnector automatically identifies the dependencies between the given Flows and creates a scheduler that starts each Flow as its data sources become available. If two or more Flow instances have no interdependencies, they are submitted together so that they can execute in parallel.

For more information, see the section on Topological Scheduling.

If an instance of cascading.flow.FlowSkipStrategy is given to a Cascade instance (via the Cascade.setFlowSkipStrategy() method), it is consulted for every Flow instance managed by that Cascade, and all skip strategies on those Flow instances are ignored. For more information on skip strategies, see Skipping Flows.

This section covers some of the operational mechanics of running an application that uses Cascading with the Hadoop platform, including building the application jar file and configuring the operating mode.

To use the HadoopFlowConnector (i.e., to run in Hadoop mode), Cascading requires that Apache Hadoop be installed and correctly configured. Hadoop is an Open Source Apache project, freely available for download from the Hadoop website, http://hadoop.apache.org/core/.

Cascading ships with several jars and dependencies in the download archive. Alternatively, Cascading is available over Maven and Ivy through the Conjars repository, along with a number of other Cascading-related projects. See http://conjars.org for more information.

The core Cascading artifacts include the following:

Cascading works with either of the Hadoop processing modes - the default local standalone mode and the distributed cluster mode. As specified in the Hadoop documentation, running in cluster mode requires the creation of a Hadoop job jar that includes the Cascading jars, plus any needed third-party jars, in its lib directory. This is true regardless of whether they are Cascading Hadoop-mode applications or raw Hadoop MapReduce applications.

During runtime, Hadoop must be told which application jar file should be pushed to the cluster. Typically, this is done via the Hadoop API JobConf object.

Cascading offers a shorthand for configuring this parameter, demonstrated here:

Properties properties = new Properties();

// pass in the class name of your application
// this will find the parent jar at runtime
AppProps.setApplicationJarClass( properties, Main.class );

// ALTERNATIVELY ...

// pass in the path to the parent jar
AppProps.setApplicationJarPath( properties, pathToJar );


// pass properties to the connector
FlowConnector flowConnector = new HadoopFlowConnector( properties );

Above we see two ways to set the same property - via the setApplicationJarClass() method, and via the setApplicationJarPath() method. One is based on a Class name, and the other is based on a literal path.

The first method takes a Class object that owns the "main" function for this application. The assumption here is that Main.class is not located in a Java Jar that is stored in the lib folder of the application Jar. If it is, that Jar is pushed to the cluster, not the parent application jar.

The second method simply sets the path to the parent Class as a property.

In your application, only one of these methods needs to be called, but one of them must be called to properly configure Hadoop.

JobConf jobConf = new JobConf();

// pass in the class name of your application
// this will find the parent jar at runtime
jobConf.setJarByClass( Main.class );

// ALTERNATIVELY ...

// pass in the path to the parent jar
jobConf.setJar( pathToJar );

// build the properties object using jobConf as defaults
Properties properties = AppProps.appProps()
  .setName( "sample-app" )
  .setVersion( "1.2.3" )
  .buildProperties( jobConf );

// pass properties to the connector
FlowConnector flowConnector = new HadoopFlowConnector( properties );

Above we are starting with an existing Hadoop JobConf instance and building a Properties object with it as the default.

Running a Cascading application is the same as running any Hadoop application. After packaging your application into a single jar (see Building), you must use bin/hadoop to submit the application to the cluster.

For example, to execute an application stuffed into your-application.jar, call the Hadoop shell script:

$HADOOP_HOME/bin/hadoop jar your-application.jar [some params]

If the configuration scripts in $HADOOP_CONF_DIR are configured to use a cluster, the Jar is pushed into the cluster for execution.

Cascading does not rely on any environment variables like $HADOOP_HOME or $HADOOP_CONF_DIR, only bin/hadoop does.

It should be noted that even though your-application.jar is passed on the command line to bin/hadoop, this in no way configures Hadoop to push this jar into the cluster. You must still call one of the property setters mentioned above to set the proper path to the application jar. If misconfigured, it's likely that one of the internal libraries (found in the lib folder) will be pushed to the cluster instead, and "Class Not Found" exceptions will be thrown.

So far we've talked about setting up sources and sinks, shaping the data streams, referencing the data fields, and so on. Within this Pipe framework, Operations are used to act upon the data - e.g., alter it, filter it, analyze it, or transform it. You can use the standard Operations in the Cascading library to create powerful and robust applications by combining them in chains (much like Unix operations such as sed, grep, sort, uniq, and awk). And if you want to go further, it's also very simple to develop custom Operations in Cascading.

There are four kinds of Operations: Function, Filter, Aggregator, and Buffer.

Operations typically require an input argument Tuple to act on. And all Operations can return zero or more Tuple object results - except Filter, which simply returns a Boolean indicating whether to discard the current Tuple. A Function, for instance, can parse a string passed by an argument Tuple and return a new Tuple for every value parsed (i.e., one Tuple for each "word"), or it may create a single Tuple with every parsed value included as an element in one Tuple object (e.g., one Tuple with "first-name" and "last-name" fields).

In theory, a Function can be used as a Filter by not emitting a Tuple result. However, the Filter type is optimized for filtering, and can be combined with logical Operations such as Not, And, Or, etc.

During runtime, Operations actually receive arguments as one or more instances of the TupleEntry object. The TupleEntry object holds the current Tuple and a Fields object that defines field names for positions within the Tuple.

Except for Filter, all Operations must declare result Fields, and if the actual output does not match the declaration, the process will fail. For example, consider a Function written to parse words out of a String and return a new Tuple for each word. If it declares that its intended output is a Tuple with a single field named "word", and then returns more values in the Tuple beyond that single "word", processing will halt. However, Operations designed to return arbitrary numbers of values in a result Tuple may declare Fields.UNKNOWN.

The Cascading planner always attempts to "fail fast" where possible by checking the field name dependencies between Pipes and Operations, but there may be some cases the planner can't account for.

All Operations must be wrapped by either an Each or an Every pipe instance. The pipe is responsible for passing in an argument Tuple and accepting the resulting output Tuple.

Operations by default are assumed by the Cascading planner to be "safe". A safe Operation is idempotent; it can safely execute multiple times on the exact same record or Tuple; it has no side-effects. If a custom Operation is not idempotent, the method isSafe() must return false. This value influences how the Cascading planner renders the Flow under certain circumstances.

A Function expects a stream of individual argument Tuples, and returns zero or more result Tuples for each of them. Like a Filter, a Function is used with an Each pipe, which may follow any pipe type.

To create a custom Function, subclass the class cascading.operation.BaseOperation and implement the interface cascading.operation.Function. Since the BaseOperation has been subclassed, the operate method, as defined on the Function interface, is the only method that must be implemented.

public class SomeFunction extends BaseOperation implements Function
  {
  public void operate( FlowProcess flowProcess, FunctionCall functionCall )
    {
    // get the arguments TupleEntry
    TupleEntry arguments = functionCall.getArguments();

    // create a Tuple to hold our result values
    Tuple result = new Tuple();

    // insert some values into the result Tuple

    // return the result Tuple
    functionCall.getOutputCollector().add( result );
    }
  }

Whenever possible, functions should declare both the number of argument values they expect and the field names of the Tuple they return. However, these declarations are optional, as explained below.

For input, functions must accept one or more values in a Tuple as arguments. If not specified, the default is to accept any number of values (Operation.ANY). Cascading verifies during planning that the number of arguments selected matches the number of arguments expected.

For output, it's a good practice to declare the field names that a function returns. If not specified, the default is Fields.UNKNOWN, meaning that an unknown number of fields are returned in each Tuple.

Both declarations - the number of input arguments and declared result fields - must be done on the constructor, either by passing default values to the super constructor, or by accepting the values from the user via a constructor implementation.

public class AddValuesFunction extends BaseOperation implements Function
  {
  public AddValuesFunction()
    {
    // expects 2 arguments, fail otherwise
    super( 2, new Fields( "sum" ) );
    }

  public AddValuesFunction( Fields fieldDeclaration )
    {
    // expects 2 arguments, fail otherwise
    super( 2, fieldDeclaration );
    }

  public void operate( FlowProcess flowProcess, FunctionCall functionCall )
    {
    // get the arguments TupleEntry
    TupleEntry arguments = functionCall.getArguments();

    // create a Tuple to hold our result values
    Tuple result = new Tuple();

    // sum the two arguments
    int sum = arguments.getInteger( 0 ) + arguments.getInteger( 1 );

    // add the sum value to the result Tuple
    result.add( sum );

    // return the result Tuple
    functionCall.getOutputCollector().add( result );
    }
  }

The example above implements a Function that accepts two values in the argument Tuple, adds them together, and returns the result in a new Tuple.

The first constructor above assumes a default field name for the field that this Function returns. In practice, it's good to give the user the option of overriding the declared field names, allowing them to prevent possible field name collisions that might cause the planner to fail.

This line is especially important:

int sum = arguments.getInteger( 0 ) +
          arguments.getInteger( 1 );
        

Note that ordinal numbers, not field names, are used here to get argument values. If field names had been used, the AddValuesFunction would have been coupled to the incoming stream.

public class EfficientAddValuesFunction
  extends BaseOperation<Tuple> implements Function<Tuple>
  {
  public EfficientAddValuesFunction()
    {
    // expects 2 arguments, fail otherwise
    super( 2, new Fields( "sum" ) );
    }

  public EfficientAddValuesFunction( Fields fieldDeclaration )
    {
    // expects 2 arguments, fail otherwise
    super( 2, fieldDeclaration );
    }

  @Override
  public void prepare( FlowProcess flowProcess, OperationCall<Tuple> call )
    {
    // create a reusable Tuple of size 1
    call.setContext( Tuple.size( 1 ) );
    }

  public void operate( FlowProcess flowProcess, FunctionCall<Tuple> call )
    {
    // get the arguments TupleEntry
    TupleEntry arguments = call.getArguments();

    // get our previously created Tuple
    Tuple result = call.getContext();

    // sum the two arguments
    int sum = arguments.getInteger( 0 ) + arguments.getInteger( 1 );

    // set the sum value on the result Tuple
    result.set( 0, sum );

    // return the result Tuple
    call.getOutputCollector().add( result );
    }

  @Override
  public void cleanup( FlowProcess flowProcess, OperationCall<Tuple> call )
    {
    call.setContext( null );
    }
  }

This example, a minor variation on the previous one, introduces the use of a "context" object and prepare() and cleanup() methods.

All Operations allow for a context object, simply a user-defined object that holds state between calls to the operate() method. This allows for a given instance of the Operation to be thread safe on a platform that may use multiple threads of execution versus multiple processes. It also allows deferring initialization of complex resources until the Operation is engaged.

The prepare() and cleanup() methods are invoked once per thread of execution, and in the case of the Hadoop platform, only on the cluster side, never on the client.

In the above example, a Tuple is used as the context; a more complex type isn't necessary. Also note that the Tuple isn't storing state, but is re-used to reduce the number of new Object instances created. In Cascading, it is perfectly safe to output the same Tuple instance from operate(). The method functionCall.getOutputCollector().add( result ) will not return until the result Tuple has been processed or persisted downstream.

A Filter expects a stream of individual argument Tuples and returns a Boolean value for each one, stating whether it should be discarded. Like a Function, a Filter is used with an Each pipe, which may follow any pipe type.

To create a custom Filter, subclass the class cascading.operation.BaseOperation and implement the interface cascading.operation.Filter. Because BaseOperation has been subclassed, the isRemove method, as defined on the Filter interface, is the only method that must be implemented.

public class SomeFilter extends BaseOperation implements Filter
  {
  public boolean isRemove( FlowProcess flowProcess, FilterCall call )
    {
    // get the arguments TupleEntry
    TupleEntry arguments = call.getArguments();

    // initialize the return result
    boolean isRemove = false;

    // test the argument values and set isRemove accordingly

    return isRemove;
    }
  }

Filters must accept one or more values in a Tuple as arguments, and should declare the number of argument values they expect. If not specified, the default is to accept any number of values (Operation.ANY). Cascading verifies during planning that the number of arguments selected matches the number of arguments expected.

The number of arguments declaration must be done on the constructor, either by passing a default value to the super constructor, or by accepting the value from the user via a constructor implementation.

public class StringLengthFilter extends BaseOperation implements Filter
  {
  public StringLengthFilter()
    {
    // expects 2 arguments, fail otherwise
    super( 2 );
    }

  public boolean isRemove( FlowProcess flowProcess, FilterCall call )
    {
    // get the arguments TupleEntry
    TupleEntry arguments = call.getArguments();

    // filter out the current Tuple if the first argument length is greater
    // than the second argument integer value
    return arguments.getString( 0 ).length() > arguments.getInteger( 1 );
    }
  }

The example above implements a Filter that accepts two arguments and filters out the current Tuple if the first argument, String length, is greater than the integer value of the second argument.

An Aggregator expects a stream of tuple groups (the output of a GroupBy or CoGroup pipe), and returns zero or more result tuples for every group. An Aggregator may only be used with an Every pipe - which may follow a GroupBy, a CoGroup, or another Every pipe, but not an Each.

To create a custom Aggregator, subclass the class cascading.operation.BaseOperation and implement the interface cascading.operation.Aggregator. Because BaseOperation has been subclassed, the start, aggregate, and complete methods, as defined on the Aggregator interface, are the only methods that must be implemented.

public class SomeAggregator extends BaseOperation<SomeAggregator.Context>
  implements Aggregator<SomeAggregator.Context>
  {
  public static class Context
    {
    Object value;
    }

  public void start( FlowProcess flowProcess,
                     AggregatorCall<Context> aggregatorCall )
    {
    // get the group values for the current grouping
    TupleEntry group = aggregatorCall.getGroup();

    // create a new custom context object
    Context context = new Context();

    // optionally, populate the context object

    // set the context object
    aggregatorCall.setContext( context );
    }

  public void aggregate( FlowProcess flowProcess,
                         AggregatorCall<Context> aggregatorCall )
    {
    // get the current argument values
    TupleEntry arguments = aggregatorCall.getArguments();

    // get the context for this grouping
    Context context = aggregatorCall.getContext();

    // update the context object
    }

  public void complete( FlowProcess flowProcess,
                        AggregatorCall<Context> aggregatorCall )
    {
    Context context = aggregatorCall.getContext();

    // create a Tuple to hold our result values
    Tuple result = new Tuple();

    // insert some values into the result Tuple based on the context

    // return the result Tuple
    aggregatorCall.getOutputCollector().add( result );
    }
  }

Whenever possible, Aggregators should declare both the number of argument values they expect and the field names of the Tuple they return. However, these declarations are optional, as explained below.

For input, Aggregators must accept one or more values in a Tuple as arguments. If not specified, the default is to accept any number of values (Operation.ANY). Cascading verifies during planning that the number of arguments selected is the same as the number of arguments expected.

For output, it's good practice for Aggregators to declare the field names they return. If not specified, the default is Fields.UNKNOWN, meaning that an unknown number of fields are returned in each Tuple.

Both declarations - the number of input arguments and declared result fields - must be done on the constructor, either by passing default values to the super constructor, or by accepting the values from the user via a constructor implementation.

public class AddTuplesAggregator
    extends BaseOperation<AddTuplesAggregator.Context>
    implements Aggregator<AddTuplesAggregator.Context>
  {
  public static class Context
    {
    long value = 0;
    }

  public AddTuplesAggregator()
    {
    // expects 1 argument, fail otherwise
    super( 1, new Fields( "sum" ) );
    }

  public AddTuplesAggregator( Fields fieldDeclaration )
    {
    // expects 1 argument, fail otherwise
    super( 1, fieldDeclaration );
    }

  public void start( FlowProcess flowProcess,
                     AggregatorCall<Context> aggregatorCall )
    {
    // set the context object, starting at zero
    aggregatorCall.setContext( new Context() );
    }

  public void aggregate( FlowProcess flowProcess,
                         AggregatorCall<Context> aggregatorCall )
    {
    TupleEntry arguments = aggregatorCall.getArguments();
    Context context = aggregatorCall.getContext();

    // add the current argument value to the current sum
    context.value += arguments.getInteger( 0 );
    }

  public void complete( FlowProcess flowProcess,
                        AggregatorCall<Context> aggregatorCall )
    {
    Context context = aggregatorCall.getContext();

    // create a Tuple to hold our result values
    Tuple result = new Tuple();

    // set the sum
    result.add( context.value );

    // return the result Tuple
    aggregatorCall.getOutputCollector().add( result );
    }
  }

The example above implements an Aggregator that accepts a value in the argument Tuple, adds all the argument tuples in the current grouping, and returns the result as a new Tuple.

The first constructor above assumes a default field name that this Aggregator returns. In practice, it's good to give the user the option of overriding the declared field names, allowing them to prevent possible field name collisions that might cause the planner to fail.

There are several constraints on the use of Aggregators that may not be self-evident. These are detailed in the Javadoc

A Buffer expects set of argument tuples in the same grouping, and may return zero or more result tuples.

A Buffer is very similar to an Aggregator, except that it receives the current Grouping Tuple, and an iterator of all the arguments it expects, for every value Tuple in the current grouping - all on the same method call. This is very similar to the typical Reducer interface in MapReduce, and is best used for operations that need visibility to the previous and next elements in the stream - such as smoothing a series of time-stamps where there are missing values.

A Buffer may only be used with an Every pipe, and it may only follow a GroupBy or CoGroup pipe type.

To create a custom Buffer, subclass the class cascading.operation.BaseOperation and implement the interface cascading.operation.Buffer. Because BaseOperation has been subclassed, the operate method, as defined on the Buffer interface, is the only method that must be implemented.

public class SomeBuffer extends BaseOperation implements Buffer
  {
  public void operate( FlowProcess flowProcess, BufferCall bufferCall )
    {
    // get the group values for the current grouping
    TupleEntry group = bufferCall.getGroup();

    // get all the current argument values for this grouping
    Iterator<TupleEntry> arguments = bufferCall.getArgumentsIterator();

    // create a Tuple to hold our result values
    Tuple result = new Tuple();

    while( arguments.hasNext() )
      {
      TupleEntry argument = arguments.next();

      // insert some values into the result Tuple based on the arguemnts
      }

    // return the result Tuple
    bufferCall.getOutputCollector().add( result );
    }
  }

Buffers should declare both the number of argument values they expect and the field names of the Tuple they return.

For input, Buffers must accept one or more values in a Tuple as arguments. If not specified, the default is to accept any number of values (Operation.ANY). During the planning phase, Cascading verifies that the number of arguments selected is the same as the number of arguments expected.

For output, it's good practice for Buffers to declare the field names they return. If not specified, the default is Fields.UNKNOWN, meaning that an unknown number of fields are returned in each Tuple.

Both declarations - the number of input arguments and declared result fields - must be done on the constructor, either by passing default values to the super constructor, or by accepting the values from the user via a constructor implementation.

public class AverageBuffer extends BaseOperation implements Buffer
  {

  public AverageBuffer()
    {
    super( 1, new Fields( "average" ) );
    }

  public AverageBuffer( Fields fieldDeclaration )
    {
    super( 1, fieldDeclaration );
    }

  public void operate( FlowProcess flowProcess, BufferCall bufferCall )
    {
    // init the count and sum
    long count = 0;
    long sum = 0;

    // get all the current argument values for this grouping
    Iterator<TupleEntry> arguments = bufferCall.getArgumentsIterator();

    while( arguments.hasNext() )
      {
      count++;
      sum += arguments.next().getInteger( 0 );
      }

    // create a Tuple to hold our result values
    Tuple result = new Tuple( sum / count );

    // return the result Tuple
    bufferCall.getOutputCollector().add( result );
    }
  }

The example above implements a buffer that accepts a value in the argument Tuple, adds all these argument tuples in the current grouping, and returns the result divided by the number of argument tuples counted in a new Tuple.

The first constructor above assumes a default field name for the field that this Buffer returns. In practice, it's good to give the user the option of overriding the declared field names, allowing them to prevent possible field name collisions that might cause the planner to fail

Note that this example is somewhat artificial. In actual practice, an Aggregator would be a better way to compute averages for an entire dataset. A Buffer is better suited for calculating running averages across very large spans, for example.

There are several constraints on the use of Buffers that may not be self-evident. These are detailed in the Javadoc.

As with the Function example above, a Buffer may define a custom context object and implement the prepare() and cleanup() methods to maintain state, or re-use outgoing Tuple instances for efficiency.

In all of the above sections, the cascading.operation.BaseOperation class was subclassed. This class is an implementation of the cascading.operation.Operation interface, and provides a few default method implementations. It is not strictly required to extend BaseOperation when implementing this interface, but it is very convenient to do so.

When developing custom operations, the developer may need to initialize and destroy a resource. For example, when doing pattern matching, you might need to initialize a java.util.regex.Matcher and use it in a thread-safe way. Or you might need to open, and eventually close, a remote connection. But for performance reasons, the operation should not create or destroy the connection for each Tuple or every Tuple group that passes through.

For this reason, the interface Operation declares two methods: prepare() and cleanup(). In the case of Hadoop and MapReduce, the prepare() and cleanup() methods are called once per Map or Reduce task. The prepare() method is called before any argument Tuple is passed in, and the cleanup() method is called after all Tuple arguments have been operated on. Within each of these methods, the developer can initialize a "context" object that can hold an open socket connection or Matcher instance. This context is user defined, and is the same mechanism used by the Aggregator operation - except that the Aggregator is also given the opportunity to initialize and destroy its context, via the start() and complete() methods.

Note that if a "context" object is used, its type should be declared in the subclass class declaration using the Java Generics notation.

Cascading is designed to be easily configured and enhanced by developers. In addition to creating custom Operations, developers can create custom Tap and Scheme classes that let applications connect to external systems or read/write data to proprietary formats.

A Tap represents something physical, like a file or a database table. Accordingly, Tap implementations are responsible for life-cycle issues around the resource they represent, such as tests for resource existence, or to perform resource deletion (dropping a remote SQL table).

A Scheme represents a format or representation - such as a text format for a file, the columns in a table, etc. Schemes are used to convert between the source data's native format and a cascading.tuple.Tuple instance.

Creating custom taps and schemes can be an involved process. When using the Cascading Hadoop mode, it requires some knowledge of Hadoop and the Hadoop FileSystem API. If a flow needs to support a new file system, passing a fully-qualified URL to the Hfs constructor may be sufficient - the Hfs tap will look up a file system based on the URL scheme via the Hadoop FileSystem API. If not, a new system is commonly constructed by subclassing the cascading.tap.Hfs class.

Delegating to the Hadoop FileSystem API is not a strict requirement. But if not using it, the developer must implement Hadoop org.apache.hadoop.mapred.InputFormat and/or org.apache.hadoop.mapred.OutputFormat classes so that Hadoop knows how to split and handle the incoming/outgoing data. The custom Scheme is responsible for setting the InputFormat and OutputFormat on the JobConf, via the sinkConfInit and sourceConfInit methods.

For examples of how to implement a custom tap and scheme, see the Cascading Modules page.

All custom Tap classes must subclass the cascading.tap.Tap abstract class and implement the required methods. The method getIdentifier() must return a String that uniquely identifies the resource the Tap instance is managing. Any two Tap instances with the same fully-qualified identifier value will be considered equal.

Every Tap is presented an opportunity to set any custom properties the underlying platform requires, via the methods sourceConfInit() (for a Tuple source tap) and sinkConfInit() (for a Tuple sink tap). These two methods may be called more than once with new configuration objects, and should be idempotent.

A Tap is always sourced from the openForRead() method via a TupleEntryIterator - i.e., openForRead() is always called in the same process that will read the data. It is up to the Tap to return a TupleEntryIterator that will iterate across the resource, returning a TupleEntry instance (and Tuple instance) for each "record" in the resource. TupleEntryIterator.close() is always called when no more entries will be read. For more on this topic, see TupleEntrySchemeIterator in the Javadoc.

On some platforms, openForRead() is called with a pre-instantiated Input type. Typically this Input type should be used instead of instantiating a new instance of the appropriate type.

In the case of the Hadoop platform, a RecordReader is created by Hadoop and passed to the Tap. This RecordReader is already configured to read data from the current InputSplit.

Similarly, a Tap is always used to sink data from the openForWrite() method via the TupleEntryCollector. Here again, openForWrite() is always called in the process in which data will be written. It is up to the Tap to return a TupleEntryCollector that will accept and store any number of TupleEntry or Tuple instances for each record that is processed or created by a given Flow. TupleEntryCollector.close() is always called when no more entries will be written. See TupleEntrySchemeCollector in the Javadoc.

Again, on some platforms, openForWrite() will be called with a pre-instantiated Output type. Typically this Output type should be used instead of instantiating a new instance of the appropriate type.

In the case of the Hadoop platform, an OutputCollector is created by Hadoop and passed to the Tap. This OutputCollector is already configured to to write data to the current resource.

Both the TupleEntrySchemeIterator and TupleEntrySchemeCollector should be used to hold any state or resources necessary to communicate with any remote services. For example, when connecting to a SQL database, any JDBC drivers should be created on the constructor and cleaned up on close().

Note that the Tap is not responsible for reading or writing data to the Input or Output type. This is delegated to the Scheme passed on the constructor of the Tap. Consequently, the Scheme is responsible for configuring the Input and Output types it will be reading and writing.

All custom Scheme classes must subclass the cascading.scheme.Scheme abstract class and implement the required methods.

A Scheme is ultimately responsible for sourcing and sinking Tuples of data. Consequently it must know what Fields it presents during sourcing, and what Fields it accepts during sinking. Thus the constructors on the base Scheme type must be set with the source and sink Fields.

A Scheme is allowed to source different Fields than it sinks. The TextLine Scheme does just this. (The TextDelimited Scheme, on the other hand, forces the source and sink Fields to be the same.)

The retrieveSourceFields() and retrieveSinkFields() methods allow a custom Scheme to fetch its source and sink Fields immediately before the planner is invoked - for example, from the header of a file, as is the case with TextDelimited. Also the presentSourceFields() and presentSinkFields() methods notify the Scheme of the Fields that the planner expects the Scheme to handle - for example, to write the field names as a header, as is the case with TextDelimited.

Every Scheme is presented the opportunity to set any custom properties the underlying platform requires, via the methods sourceConfInit() (for a Tuple source tap) and sinkConfInit() (for a Tuple sink tap). These methods may be called more than once with new configuration objects, and should be idempotent.

On the Hadoop platform, these methods should be used to configure the appropriate org.apache.hadoop.mapred.InputFormat and org.apache.hadoop.mapred.OutputFormat.

A Scheme is always sourced via the source() method, and is always sunk to via the sink() method.

Prior to a source() or sink() call, the sourcePrepare() and sinkPrepare() methods are called. After all values have been read or written, the s ourceCleanup() and sinkCleanup() methods are called.

The *Prepare() methods allow a Scheme to initialize any state necessary - for example, to create a new java.util.regex.Matcher instance for use against all record reads). Conversely, the *Cleanup() methods allow for clearing up any resources.

These methods are always called in the same process space as their associated source() and sink() calls. In the case of the Hadoop platform, this will likely be on the cluster side, unlike calls to *ConfInit() which will likely be on the client side.

In Cascading, SubAssemblies are reusable pipe assemblies that are linked into larger pipe assemblies. They function much like subroutines in a larger program. SubAssemblies are a good way to organize complex pipe assemblies, and they allow for commonly-used pipe assemblies to be packaged into libraries for inclusion in other projects by other users.

To create a SubAssembly, subclass the cascading.pipe.SubAssembly class.

public class SomeSubAssembly extends SubAssembly
  {
  public SomeSubAssembly( Pipe lhs, Pipe rhs )
    {
    // continue assembling against lhs
    lhs = new Each( lhs, new SomeFunction() );
    lhs = new Each( lhs, new SomeFilter() );

    // continue assembling against rhs
    rhs = new Each( rhs, new SomeFunction() );

    // joins the lhs and rhs
    Pipe join = new CoGroup( lhs, rhs );

    join = new Every( join, new SomeAggregator() );

    join = new GroupBy( join );

    join = new Every( join, new SomeAggregator() );

    // the tail of the assembly
    join = new Each( join, new SomeFunction() );

    // must register all assembly tails
    setTails( join );
    }
  }

In the example above, we pass in (as parameters via the constructor) the pipes that we wish to continue assembling against, and in the last line we register the "join" pipe as a tail. This allows SubAssemblies to be nested within larger pipe assemblies or other SubAssemblies.

// the "left hand side" assembly head
Pipe lhs = new Pipe( "lhs" );

// the "right hand side" assembly head
Pipe rhs = new Pipe( "rhs" );

// our custom SubAssembly
Pipe pipe = new SomeSubAssembly( lhs, rhs );

pipe = new Each( pipe, new SomeFunction() );

The example above demonstrates how to include a SubAssembly into a new pipe assembly.

Note that in a SubAssembly that represents a split - that is, a SubAssembly with two or more tails - you can use the getTails() method to access the array of tails set internally by the setTails() method.

public class SplitSubAssembly extends SubAssembly
  {
  public SplitSubAssembly( Pipe pipe )
    {
    // continue assembling against pipe
    pipe = new Each( pipe, new SomeFunction() );

    Pipe lhs = new Pipe( "lhs", pipe );
    lhs = new Each( lhs, new SomeFunction() );

    Pipe rhs = new Pipe( "rhs", pipe );
    rhs = new Each( rhs, new SomeFunction() );

    // must register all assembly tails
    setTails( lhs, rhs );
    }
  }

// the "left hand side" assembly head
Pipe head = new Pipe( "head" );

// our custom SubAssembly
SubAssembly pipe = new SplitSubAssembly( head );

// grab the split branches
Pipe lhs = new Each( pipe.getTails()[ 0 ], new SomeFunction() );
Pipe rhs = new Each( pipe.getTails()[ 1 ], new SomeFunction() );

To rephrase, if a SubAssembly does not split the incoming Tuple stream, the SubAssembly instance can be passed directly to the next Pipe instance. But, if the SubAssembly splits the stream into multiple branches, handles will be needed to access them. The solution is to pass each branch tail to the setTails() method, and call the getTails() method to get handles for the desired branches, which can be passed to subsequent instances of Pipe.

Stream assertions are simply a mechanism for asserting that one or more values in a tuple stream meet certain criteria. This is similar to the Java language "assert" keyword, or a unit test. An example would be "assert not null" or "assert matches".

Assertions are treated like any other function or aggregator in Cascading. They are embedded directly into the pipe assembly by the developer. By default, if an assertion fails, the processing fails. As an alternative, an assertion failure can be caught by a failure Trap.

Assertions may be more, or less, desirable in different contexts. For this reason, stream assertions can be treated as either "strict" or "validating". Strict assertions make sense when running tests against regression data - which should be small, and should represent many of the edge cases that the processing assembly must robustly support. Validating assertions, on the other hand, make more sense when running tests in staging, or when using data that may vary in quality due to an unmanaged source.

And of course there are cases where assertions are unnecessary and only impede processing, and it would be best to just bypass them altogether.

To handle all three of these situations, Cascading can be instructed to plan out (i.e., omit) strict assertions, validation assertions, or both when building the Flow. To create optimal performance, Cascading implements this by actually leaving the undesired assertions out of the final Flow (not merely switching them off).

// incoming -> "ip", "time", "method", "event", "status", "size"

AssertNotNull notNull = new AssertNotNull();
assembly = new Each( assembly, AssertionLevel.STRICT, notNull );

AssertSizeEquals equals = new AssertSizeEquals( 6 );
assembly = new Each( assembly, AssertionLevel.STRICT, equals );

AssertMatchesAll matchesAll = new AssertMatchesAll( "(GET|HEAD|POST)" );
assembly = new Each( assembly, new Fields( "method" ),
  AssertionLevel.STRICT, matchesAll );

// outgoing -> "ip", "time", "method", "event", "status", "size"

Again, assertions are added to a pipe assembly like any other operation, except that the AssertionLevel must be set to tell the planner how to treat the assertion during planning.

// FlowDef is a fluent way to define a Flow
FlowDef flowDef = new FlowDef();

// bind the taps and pipes
flowDef
  .addSource( assembly.getName(), source )
  .addSink( assembly.getName(), sink )
  .addTail( assembly );

// removes all assertions from the Flow
flowDef
  .setAssertionLevel( AssertionLevel.NONE );

Flow flow = new HadoopFlowConnector().connect( flowDef );

To configure the planner to remove some or all assertions, a property can be set via the FlowConnectorProps.setAssertionLevel() method or directly on the FlowDef instance, as shown above. Setting AssertionLevel.NONE removes all assertions. AssertionLevel.VALID keeps VALID assertions but removes STRICT ones. And AssertionLevel.STRICT keeps all assertions - the planner default value.

The following diagram shows the use of Failure Traps in a pipe assembly.

Failure Traps are similar to tap sinks (as opposed to tap sources) in that they allow data to be stored. The difference is that Tap sinks are bound to a particular tail pipe in a pipe assembly and are the primary outlet of a branch in a pipe assembly. Traps can be bound to intermediate pipe assembly branches - just like Stream Assertions - yet they only capture data that causes an Operation to fail (throw an Exception).

Whenever an operation fails and throws an exception, if there is an associated trap, the offending Tuple is saved to the resource specified by the trap Tap. This allows the job to continue processing without any data loss.

By design, clusters are hardware fault-tolerant - lose a node, and the cluster continues working. But fault tolerance for software is a little different. Failure Traps provide a means for the processing to continue without losing track of the data that caused the fault. For high fidelity applications, this may not be very useful, since you likely will want any errors during processing to cause the application to stop. But for low fidelity applications such as webpage indexing, where skipping a page or two out of a few million is acceptable, this can dramatically improve processing reliability.

// ...some useful pipes here

// name this pipe assembly segment
assembly = new Pipe( "assertions", assembly );

AssertNotNull notNull = new AssertNotNull();
assembly = new Each( assembly, AssertionLevel.STRICT, notNull );

AssertSizeEquals equals = new AssertSizeEquals( 6 );
assembly = new Each( assembly, AssertionLevel.STRICT, equals );

AssertMatchesAll matchesAll = new AssertMatchesAll( "(GET|HEAD|POST)" );
Fields method = new Fields( "method" );
assembly =
  new Each( assembly, method, AssertionLevel.STRICT, matchesAll );

// ...some more useful pipes here

FlowDef flowDef = new FlowDef();

flowDef
  .setName( "log-parser" )
  .addSource( "logs", source )
  .addTailSink( assembly, sink );

// set the trap on the "assertions" branch
flowDef
  .addTrap( "assertions", trap );

FlowConnector flowConnector = new HadoopFlowConnector();
Flow flow =
  flowConnector.connect( flowDef );

The example above binds a trap tap to the pipe assembly segment named "assertions". Note how we can name branches and segments by using a single Pipe instance, and that the naming applies to all subsequent Pipe instances.

Traps are for exceptional cases, in the same way that Java Exception handling is. Traps are not intended for application flow control, and not a means to filter some data into other locations. Applications that need to filter out bad data should do so explicitly, using filters. For more on this, see Handling Good and Bad Data.

New to Cascading 2, and only supported by the Hadoop planner, is the ability to "checkpoint" data within a Flow by using the cascading.pipe.Checkpoint Pipe. That is, a Tuple stream can be persisted to disk at most any arbitrary point. Doing so forces a new FlowStep (MapReduce job when using Hadoop) after the checkpoint position.

By default the checkpoint is anonymous and is cleaned up immediately after the Flow completes. This feature is useful when used in conjunction with a HashJoin where the small side of the join starts out extremely large but is filtered down to fit into memory before being read into the HashJoin. By forcing a checkpoint before the HashJoin, only the small filtered version of the data is replicated over the cluster. Without the checkpoint, it is likely the full unfiltered file will be replicated to every node the FlowStep is executing.

Alternatively, checkpointing is useful for debugging when used with a checkpoint Tap, where the Tap has specified a TextDelimited Scheme without any declared Fields.

// the "left hand side" assembly head
Pipe lhs = new Pipe( "lhs" );

lhs = new Each( lhs, new SomeFunction() );
lhs = new Each( lhs, new SomeFilter() );

// the "right hand side" assembly head
Pipe rhs = new Pipe( "rhs" );

rhs = new Each( rhs, new SomeFunction() );

// joins the lhs and rhs
Pipe join = new CoGroup( lhs, rhs );

join = new Every( join, new SomeAggregator() );

// we want to see the data passing through this point
Checkpoint checkpoint = new Checkpoint( "checkpoint", join );

Pipe groupBy = new GroupBy( checkpoint );

groupBy = new Every( groupBy, new SomeAggregator() );

// the tail of the assembly
groupBy = new Each( groupBy, new SomeFunction() );

Tap lhsSource = new Hfs( new TextLine(), "lhs.txt" );
Tap rhsSource = new Hfs( new TextLine(), "rhs.txt" );

Tap sink = new Hfs( new TextLine(), "output" );

// write all data as a tab delimited file, with headers
Tap checkpointTap =
  new Hfs( new TextDelimited( true, "\t" ), "checkpoint" );

FlowDef flowDef = new FlowDef()
  .setName( "flow-name" )
  .addSource( rhs, rhsSource )
  .addSource( lhs, lhsSource )
  .addTailSink( groupBy, sink )
  .addCheckpoint( checkpoint, checkpointTap ); // bind the checkpoint tap

Flow flow = new HadoopFlowConnector().connect( flowDef );

As can be seen above, we instantiate a new Checkpoint tap by passing it the previous Every Pipe. This will be the point at which data is persisted. Since we wish to keep the data after the Flow has completed, we create a checkpointTap that saves the data as a TAB delimited text file. We also specify that field names should be written out into a header file on the TextDelimited constructor. Finally the Tap is bound to the Checkpoint Pipe using the FlowDef.

When using Checkpoint pipes in a Flow and the Flow fails, a future execution of the Flow can be restarted after the last successful FlowStep writing to a checkpoint file. That is, a Flow will only restart from the last Checkpoint Pipe location.

This feature requires that the failed Flow be planned with a runID set on the FlowDef, and the retry Flow use the same runID value. It goes without saying, the retry Flow should be (roughly) equivant to the previous failed attempt.

FlowDef flowDef = new FlowDef()
  .setName( "flow-name" )
  .addSource( rhs, rhsSource )
  .addSource( lhs, lhsSource )
  .addTailSink( groupBy, sink )
  .addCheckpoint( checkpoint, checkpointTap )
  .setRunID( "some-unique-value" ); // re-use this id to restart this flow

Flow flow = new HadoopFlowConnector().connect( flowDef );

Caution should be used when using restarted checkpoint Flows. If the input data has changed, or the pipe assembly has significantly been altered, the Flow may fail or there may be undetectable errors.

Note that when using a runID, all Flow instances must use a unique value unless they are intended as a retry attempt. The runID value is used to scope the directories for the temporary checkpoint files to prevent file name collisions.

On successful completion of a Flow with a runID, all temporary checkpoint files will be removed, if any.

Each Flow and Cascade has the ability to execute callbacks via an event listener. This ability is useful when an external application needs to be notified that either a Flow or Cascade has started, halted, completed, or either has thrown an exception.

For instance, at the completion of a flow that runs on an Amazon EC2 Hadoop cluster, an Amazon SQS message can be sent to notify another application to fetch the job results from S3 or begin the shutdown of the cluster.

Flows support event listeners through the cascading.flow.FlowListener interface and Cascades support event listeners through the cascading.cascade.CascadeListener, which supports four events:

The TemplateTap Tap class provides a simple means to break large datasets into smaller sets based on data item values. This is commonly called partitioning or binning the data, where each "bin" of data is named after some data value(s) shared by the members of that bin. For example, this is a simple way to organize log files by month and year.

TextDelimited scheme =
  new TextDelimited( new Fields( "year", "month", "entry" ), "\t" );
Hfs tap = new Hfs( scheme, path );

String template = "%s-%s"; // dirs named "year-month"
Tap months = new TemplateTap( tap, template, SinkMode.REPLACE );

In the example above, we construct a parent Hfs tap and pass it to the constructor of a Templatetap instance, along with a String format "template". This format template is populated in the order in which values are declared via the Scheme class. If more complex path formatting is necessary, you may subclass the Templatetap.

Note that you can only create sub-directories to bin data into. Hadoop must still write "part" files into each bin directory, and there is no safe mechanism for manipulating part file names.

One last thing to keep in mind is whether binning happens during the Map phase or the Reduce phase. By doing a GroupBy on the values used to populate the template, binning will happen during the Reduce phase, and will likely scale much better in cases where there are a very large number of unique values used in the template resulting in a large number of directories.

In Hadoop mode, Cascading does not support MapReduce "Combiners". Combiners are a simple optimization allowing some Reduce functions to run on the Map side of MapReduce. Combiners are very powerful in that they reduce the I/O between the Mappers and Reducers - why send all of your Mapper data to Reducers when you can compute some values on the Map side and combine them in the Reducer? But Combiners are limited to Associative and Commutative functions only, such as "sum" and "max". And the process requires that the values emitted by the Map task must be serialized, sorted (which involves deserialization and comparison), deserialized again, and operated on - after which the results are again serialized and sorted. Combiners trade CPU for gains in I/O.

Cascading takes a different approach. It provides a mechanism to perform partial aggregations on the Map side and combine the results on the Reduce side, but trades memory, instead of CPU, for I/O gains by caching values (up to a threshold limit). This bypasses the redundant serialization, deserialization, and sorting. Also, Cascading allows any aggregate function to be implemented - not just Associative and Commutative functions.

Cascading supports a few built-in partial aggregate operations, including AverageBy, CountBy, SumBy, and FirstBy. These are actually SubAssemblies, not Operations, and are subclasses of the AggregateBy SubAssembly. For more on this, see the section on AggregateBy.

Using partial aggregate operations is quite easy. They are actually less verbose than a standard Aggregate operation.

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );
Fields valueField = new Fields( "size" );
Fields sumField = new Fields( "total-size" );
assembly =
  new SumBy( assembly, groupingFields, valueField, sumField, long.class );

For composing multiple partial aggregate operations, things are done a little differently.

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );

// note we do not pass the parent assembly Pipe in
Fields valueField = new Fields( "size" );
Fields sumField = new Fields( "total-size" );
SumBy sumBy = new SumBy( valueField, sumField, long.class );

Fields countField = new Fields( "num-events" );
CountBy countBy = new CountBy( countField );

assembly = new AggregateBy( assembly, groupingFields, sumBy, countBy );

It's important to note that a GroupBy Pipe is embedded in the resulting assemblies above. But only one GroupBy is performed in the case of the AggregateBy, and all of the partial aggregations will be performed simultaneously. It is also important to note that, depending on the final pipe assembly, the Map side partial aggregate functions may be planned into the previous Reduce operation in Hadoop, further improving the performance of the application.

The cascading.operation.Identity function is used to "shape" a tuple stream. Here are some common patterns that illustrate how Cascading "field algebra" works. (Note that, in actual practice, some of these example tasks might be better performed with helper subassemblies such as cascading.pipe.assembly.Rename, cascading.pipe.assembly.Retain, and cascading.pipe.assemblyDiscard.)

The cascading.operation.Debug function is a utility function (actually, it's a Filter) that prints the current argument Tuple to either stdout or stderr. Used with one of the DebugLevel enum values (NONE, DEFAULT, or VERBOSE), different debug levels can be embedded in a pipe assembly.

The example below inserts a Debug operation at the VERBOSE level, but configures the planner to remove all Debug operations from the resulting Flow.

Pipe assembly = new Pipe( "assembly" );

// ...
assembly = new Each( assembly, DebugLevel.VERBOSE, new Debug() );
// ...

// head and tail have same name
FlowDef flowDef = new FlowDef()
  .setName( "debug" )
  .addSource( "assembly", source )
  .addSink( "assembly", sink )
  .addTail( assembly );


// tell the planner to remove all Debug operations
flowDef
  .setDebugLevel( DebugLevel.NONE );

// ...
FlowConnector flowConnector = new HadoopFlowConnector();

Flow flow = flowConnector.connect( flowDef );

Note that if the above Flow is run on a cluster, the stdout on the cluster nodes will be used. Nothing from the debug output will display on the client side. Debug is only useful when testing things in an IDE or if the remote logs are readily available.

The Sample and Limit functions are used to limit the number of tuples that pass through a pipe assembly.

The cascading.operation.Insert function allows for insertion of constant literal values into the tuple stream.

This is most useful when a splitting a tuple stream and one of the branches needs some identifying value, or when some missing parameter or value, like a date String for the current date, needs to be inserted.

Cascading includes a number of text functions in the cascading.operation.text package.

Cascading provides some support for dynamically-compiled Java expressions to be used in either Functions or Filters. This capability is provided by the Janino embedded Java compiler, which compiles the expressions into byte code for optimal processing speed. Janino is documented in detail on its website, http://www.janino.net/.

This capability allows an Operation to evaluate a suitable one-line Java expression, such as a + 3 * 2 or a < 7, where the variable values ( a and b) are passed in as Tuple fields. The result of the Operation thus depends on the evaluated result of the expression - in the first example, some number, and in the second, a Boolean value.

To use XML Operations in a Cascading application, include the cascading-xml-x.y.z.jar in the project. When using the TagSoupParser operation, this module requires the TagSoup library, which provides support for HTML and XML "tidying". More information is available at the TagSoup website, http://home.ccil.org/~cowan/XML/tagsoup/.

Cascading Stream Assertions are used to build robust reusable pipe assemblies. If desired, they can be planned out of a Flow instance at runtime. For more information, see the section on Stream Assertions. Below we describe the Assertions available in the core library.

The logical Filter operators allow you to combine multiple filters to run in a single Pipe, instead of chaining multiple Pipes together to get the same logical result.

// incoming -> "ip", "time", "method", "event", "status", "size"

FilterNull filterNull = new FilterNull();
RegexFilter regexFilter = new RegexFilter( "(GET|HEAD|POST)" );

And andFilter = new And( filterNull, regexFilter );

assembly = new Each( assembly, new Fields( "method" ), andFilter );

// outgoing -> "ip", "time", "method", "event", "status", "size"

The example above performs a logical "and" on the two filters. Both must be satisfied for the data to pass through this one Pipe.

The cascading.pipe.assembly.AggregateBy SubAssembly is an implementation of the Partial Aggregation pattern, and is the base class for built-in and custom partial aggregation implementations like AverageBy or CountBy.

Generally the AggregateBy class is used to combine multiple AggregateBy subclasses into a single Pipe.

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );

// note we do not pass the parent assembly Pipe in
Fields valueField = new Fields( "size" );
Fields sumField = new Fields( "total-size" );
SumBy sumBy = new SumBy( valueField, sumField, long.class );

Fields countField = new Fields( "num-events" );
CountBy countBy = new CountBy( countField );

assembly = new AggregateBy( assembly, groupingFields, sumBy, countBy );

To create a custom partial aggregation, subclass the AggregateBy class and implement the appropriate internal interfaces. See the Javadoc for details.

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );
Fields valueField = new Fields( "size" );
Fields avgField = new Fields( "avg-size" );
assembly = new AverageBy( assembly, groupingFields, valueField, avgField );

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );
Fields countField = new Fields( "count" );
assembly = new CountBy( assembly, groupingFields, countField );

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );
Fields valueField = new Fields( "size" );
Fields sumField = new Fields( "total-size" );
assembly =
  new SumBy( assembly, groupingFields, valueField, sumField, long.class );

Pipe assembly = new Pipe( "assembly" );

// ...
Fields groupingFields = new Fields( "date" );
Fields valueField = new Fields( "size" );

// we want the largest size in this grouping
valueField.setComparator( "size", new LongComparator() );

assembly =
  new FirstBy( assembly, groupingFields, valueField );

The cascading.pipe.assembly.SumBy SubAssembly is used to coerce a set of values from one type to another type - for example, to convert the field age from a String to an Integer.

// incoming -> first, last, age

assembly =
  new Coerce( assembly, new Fields( "age" ), Integer.class );

// outgoing -> first, last, age

The cascading.pipe.assembly.Discard SubAssembly is used to shape the Tuple stream by discarding all fields given on the constructor. All fields not listed are retained.

// incoming -> first, last, age

assembly = new Discard( assembly, new Fields( "age" ) );

// outgoing -> first, last

The cascading.pipe.assembly.Rename SubAssembly is used to rename a field.

// incoming -> first, last, age

assembly =
  new Rename( assembly, new Fields( "age" ), new Fields( "years" ) );

// outgoing -> first, last, years

The cascading.pipe.assembly.Retain SubAssembly is used to shape the Tuple stream by retaining all fields given on the constructor. All fields not listed are discarded.

// incoming -> first, last, age

assembly = new Retain( assembly, new Fields( "first", "last" ) );

// outgoing -> first, last

The cascading.pipe.assembly.Unique SubAssembly is used to remove duplicate values in a Tuple stream. Uniqueness is determined by the values of all fields listed in uniqueFields. Thus to find all distinct Tuples in a Tuple stream, use Fields.ALL as the uniqueFields argument.

// incoming -> first, last

assembly = new Unique( assembly, new Fields( "first", "last" ) );

// outgoing -> first, last

Discrete testing of all Operations, pipe assemblies, and applications is a must. The cascading.CascadingTestCase provides a number of helper methods.

When testing custom Operations, use the invokeFunction(), invokeFilter(), invokeAggregator(), and invokeBuffer() methods.

When testing Flows, use the validateLength() methods. There are quite a few of them, and collectively they offer great flexibility. All of them read the sink tap, validate that it is the correct length and has the correct Tuple size, and check to see whether the values match a given regular expression pattern.

As of Cascading 2, it is possible to write tests that are independent of the underlying platform. Any unit test should subclass cascading.PlatformTestCase located in the cascading-platform-x.y.z-tests.jar jar file. Any platform to be tested against should be added to the classpath as well. PlatformTestCase will search the classpath for all available platforms and run each test on the subclass against each platform found.

You can exclude any platforms you don't want to have the tests run against by omitting from a system property named platform.includes. For example, to only run the local mode tests when both local and Hadoop platforms are in the classpath, call System.setProperty( "platform.includes", "local ); or use an appropriate -D switch to the JVM.

See the Cascading platform unit tests for examples.

For Maven users, be sure to add the tests classifier to any dependencies. Note that the cascading-platform project has no main code, but does have only tests, so it must be retrieved via the tests classifier.

Although using one large Flow may result in slightly more efficient performance, it's advisable to use a more modular and flexible approach, creating medium or small Flows with well-defined responsibilities, and passing all the resulting interdependent Flows to a Cascade to sequence and execute as a single unit. Similarly, using the TextDelimited Scheme (or any custom format for long-term archival) between Flow instances allows you to hand off intermediate data to other systems for reporting or QA purposes, incurring a minimal performance penalty while remaining compatible with other tools.

When developing your applications, use SubAssembly subclasses, not "factory" methods. The resulting code is much easier to read and test.

It's worth noting that the Object constructors are "factories", so there isn't much reason to build frameworks to duplicate what a constructor already does. Of course there are exceptional cases in which you don't have the option to use a SubAssembly, but in practice they are rare.

SubAssembies provide a very convenient means to co-locate similar or related responsibilities into a single place. For example, it's simple to use a ParsingSubAssembly and a RulesSubAssembly, where the first is responsible solely for parsing incoming Tuple streams (log files for example), and the second applies rules to decide whether a given Tuple should be discarded or marked as bad.

Additionally, in your unit tests you can create a TestAssertionsSubAssembly that simply inlines various ValueAssertions and GroupAssertions. The practice of inlining Assertions directly in your SubAssemblies is also important, but sometimes it makes sense to have more tests outside of the business logic.

There are a few Operations in Cascading (e.g., ExpressionFunction and ExpressionFilter) that compile and apply Java expressions on the fly. In these expressions, Operation argument field names are used as variable names in the expression. For this reason, take care to create field names that don't contain characters which will cause compilation errors if they are used in an expression. For example, "first-name" is a valid field name for use with Cascading, but might result in the expression first-name.trim(), which will cause a compilation error.

The FlowConnector will sometimes fail when attempting to plan a Flow. If the error message given by PlannerException is vague, use the method PlannerException.writeDOT() to export a representation of the internal pipe assembly. DOT files can be opened by GraphViz and OmniGraffle. These plans are only partial, but you will be able to see where the Cascading planner failed.

Note that you can also create a DOT file from a Flow, by using Flow.writeDOT().

When joining two streams via a CoGroup Pipe, try to put the largest of the streams in the leftmost argument to the CoGroup. The reason for this is that joining multiple streams requires some accumulation of values before the join operator can begin, but the leftmost stream is not accumulated, so this technique should improve the performance of most joins.

When creating complex assemblies, it's safe to embed Debug operations (see Debug Function) at appropriate debug levels as needed. Use the planner to remove them at runtime for production and staging runs, to avoid wasting resources.

It's very common when processing raw data streams to encounter data that is corrupt or malformed in some way. For instance, bad content may be fetched from the web via a crawler upstream, or a bug may have leaked into a browser widget somewhere that sends user behavior information back for analysis. Whatever the cause, it's a good practice to define a set of rules for identifying and discarding questionable records.

It is tempting to simply throw an exception and have a Trap capture the offending Tuple, but Traps were not designed as a filtering mechanism, and consequently much valuable information would be lost.

Instead of traps, use filters. Create a SubAssembly that applies rules to the stream by setting a binary field that marks the tuple as good or bad. After all the rules are applied, split the stream based on the value of the good or bad Boolean value. Consider setting a reason field that states why the Tuple was marked bad.

When creating custom Operations (Function, Filter, Aggregator, or Buffer) do not store operation state in class fields.

For example, if implementing a custom "counter" Aggregator, do not create a field named "count" and increment it on every Aggregator.aggregate() call. There is no guarantee that your Operation will be called from a single thread in a JVM - and future versions of Hadoop or Cascading local mode might execute the same operation from multiple threads.

Passing a custom class through a Tuple stream is generally frowned upon. It leads to the coupling of custom Operations to particular types, and it removes opportunities for reducing the amount of data that passes over the network.

The first objection can be overcome with a little work. When using a custom type that has multiple instance fields, try to provide Functions that can promote a value from the custom object to a position in a Tuple, or demote the Tuple value for a particular field back into the custom type. This lets you use existing operations like ExpressionFunction or RegexFilter to operate on values owned by a custom type.

For example, if you have a Person object, create a Function named GetPersonAge that takes Person as an argument and returns just the age. The next operation can then Filter all Persons based on their age. This may seem more difficult and less effiicient, but it keeps your application flexible and avoids duplicating existing operations. (The only alternative here is to create a PersonAgeFilter, which becomes one more thing to test.)

Instead of having String field names strewn about, create an Interface that holds a constant value for each field name:

public static Fields FIRST_NAME = new Fields( "firstname" );

Using the Fields class, instead of String, allows for building more complex constants:

public static Fields NAME = FIRST_NAME.append( LAST_NAME );

When in doubt, look at the Cascading source code. If something is not documented in this User Guide or Javadoc, and it's a feature of Cascading, the feature source code or unit tests will give you clear instructions on what to do or expect.

The Cascading API was designed with scripting in mind. Any Java-compatible scripting language can import and instantiate Cascading classes, create pipe assemblies and flows, and execute those flows. And if the scripting language in question supports Domain Specific Language (DSL) creation, users can create their own DSLs to handle common idioms.

The Cascading website includes information on scripting language bindings that are publicly available.

The Tuple class is a generic container for all java.lang.Object instances. Thus any primitive value or custom Class can be stored in a Tuple instance - that is, returned by a Function, Aggregator, or Buffer as a result value.

But for this to work when using the Cascading Hadoop mode, any Class that isn't a primitive type or a Hadoop Writable type requires a corresponding Hadoop serialization class registered in the Hadoop configuration files for your cluster. Hadoop Writable types work because there is already a generic serialization implementation built into Hadoop. See the Hadoop documentation for information on registering a new serialization helper or creating Writable types. Registered serialization implementations are automatically inherited by Cascading.

During serialization and deserialization of Tuple instances that contain custom types, the Cascading Tuple serialization framework must store the class name (as a String) before serializing the custom object. This can be very space-inefficient. To overcome this, custom types can add the SerializationToken Java annotation to the custom type class. The SerializationToken annotation expects two arrays - one of integers that are used as tokens, and one of Class name strings. Both arrays must be the same size. The integer tokens must all have values of 128 or greater, since the first 128 values are reserved for internal use.

During serialization and deserialization, the token values are used instead of the String Class names, in order to reduce the amount of storage used.

Serialization tokens may also be stored in the Hadoop config files or set as a property passed to the FlowConnector, with the property name cascading.serialization.tokens. The value of this property is a comma separated list of token=classname values.

Note that Cascading natively serializes/deserializes all primitives and byte arrays (byte[]), if the developer registers the BytesSerialization class by using TupleSerializationProps.addSerialization(properties, BytesSerialization.class.getName(). The token 127 is used for the Hadoop BytesWritable class.

By default, Cascading uses lazy deserialization on Tuple elements during comparisons when Hadoop sorts keys during the "shuffle" phase.

Cascading supports custom serialization for custom types, as well as lazy deserialization of custom types during comparisons. This is accomplished by implementing the StreamComparator interface. See the Javadoc for detailed instructions on implemention, and the unit tests for examples.

Frequently, objects in one Tuple are compared to objects in a second Tuple. This is especially true during the sort phase of GroupBy and CoGroup in Cascading Hadoop mode . By default, Hadoop and Cascading use the native Object methods equals() and hashCode() to compare two values and get a consistent hash code for a given value, respectively.

To override this default behavior, you can create a custom java.util.Comparator class to perform comparisons on a given field in a Tuple. For instance, to secondary-sort a collection of custom Person objects in a GroupBy, use the Fields.setComparator() method to designate the custom Comparator to the Fields instance that specifies the sort fields.

Alternatively, you can set a default Comparator to be used by a Flow, or used locally on a given Pipe instance. There are two ways to do this. Call FlowProps.setDefaultTupleElementComparator() on a Properties instance, or use the property key cascading.flow.tuple.element.comparator.

If the hash code must also be customized, the custom Comparator can implement the interface cascading.tuple.Hasher. For more information, see the Javadoc.

The Hadoop MapReduce Job Planner is an internal feature of Cascading.

When a collection of functions, splits, and joins are all tied up together into a "pipe assembly", the FlowConnector object is used to create a new Flow instance against input and output data paths. This Flow is a single Cascading job.

Internally, the FlowConnector employs an intelligent planner to convert the pipe assembly to a graph of dependent MapReduce jobs that can be executed on a Hadoop cluster.

All this happens behind the scenes - as does the scheduling of the individual MapReduce jobs, and the cleanup of intermediate data sets that bind the jobs together.

The diagram above shows how a typical Flow is partitioned into MapReduce jobs. Every job is delimited by a temporary file that serves as the sink from the first job and the source for the next.

To create a visualization of how your Flows are partitioned, call the Flow#writeDOT() method. This writes a DOT file out to the path specified, which can be viewed in a graphics package like OmniGraffle or Graphviz.

Cascading has a simple class, Cascade , that executes a collection of Cascading Flows on a target cluster in dependency order.

Consider the following example.

A Cascade is constructed through the CascadeConnector class, by building an internal graph that makes each Flow a "vertex", and each file an "edge". A topological walk on this graph will touch each vertex in order of its dependencies. When a vertex has all its incoming edges (i.e., files) available, it is scheduled on the cluster.

In the example above, Flow 1 goes first, Flow 2 goes second, and Flow 3 is last.

If two or more Flows are independent of one another, they are scheduled concurrently.

And by default, if any outputs from a Flow are newer than the inputs, the Flow is skipped. The assumption is that the Flow was executed recently, since the output isn't stale. So there is no reason to re-execute it and use up resources or add time to the job. This is similar behavior a compiler would exhibit if a source file wasn't updated before a recompile.

This is very handy if you have a large set of jobs, with varying interdependencies between them, that needs to be executed as a logical unit. Just pass them to the CascadeConnector and let it sort them all out.