LinkedInAttic / Cubert

Cubert is a fast and efficient batch computation engine for complex analysis and reporting of massive datasets on Hadoop.

Cubert Documentation hosted at github.

Cubert Users Google Group: cubert-users. Email: [email protected]

Cubert is ideally suited for the following application domains:

1. Statistical Calculations, Joins and Aggregations

   Cubert introduces a new model of computation that allows users to organize data in a format that is ideally suited for scalable execution of subsequent query processing operators, and a set of algorithmically-efficient operators (MeshJoin and CUBE) that exploit the organization to provide significantly improved CPU and resource utilization compared to existing solutions.

2. Cubes and Grouping Set Aggregations

   The power-horse is the new CUBE operator that can efficiently (CPU and memory) compute additive, non-additive (e.g. Count Distinct) and exact percentile rank (e.g. Median) statistics; can roll up inner dimensions on-the-fly and compute multiple measures within a single job.

3. Time range calculation and Incremental computations

   Cubert primitives are specially suited for reporting workflows that employ computation pattern that is both regular and repetitive, allowing for efficiency gains from partial result caching and incremental processing.

4. Graph computations

   Cubert provides a novel sparse matrix multiplication algorithm that is best suited for analytics with large-scale graphs.

5. When performance or resources are a matter of concern

   Cubert Script is a developer-friendly language that takes out the hints, guesswork and surprises when running the script. The script provides the developers complete control over the execution plan (without resorting to low-level programming!), and is extremely extensible by adding new functions, aggregators and even operators.

A flavor of Cubert Script

Cubert script is a physical script where we explicitly define the operators at the Mappers, Reducers and Combiners for the different jobs. Following is an example of the Word Count problem written in cubert script.

JOB "word count job"
    REDUCERS 10;
    MAP {
            // load the input data set as a TEXT file
            input = LOAD "$CUBERT_HOME/examples/words.txt" USING TEXT("schema": "STRING word");
            // add a column to each tuple
            with_count = FROM input GENERATE word, 1 AS count;
    }
    // shuffle the data and also invoke combiner to aggregate on map-side
    SHUFFLE with_count PARTITIONED ON word AGGREGATES COUNT(count) AS count;
    REDUCE {
            // at the reducers, sum the counts for each word
            output = GROUP with_count BY word AGGREGATES SUM(count) AS count;
    }
	// store the output using TEXT format
    STORE output INTO "output" USING TEXT();
END

While the Cubert Script code above is already very concise representation of the Word Count problem; as a matter of interest, the idiomatic way of writing in Cubert is even more concise (and a lot faster)!

JOB "idiomatic word count program (even more concise!)"
        REDUCERS 10;
        MAP {
                input = LOAD "$CUBERT_HOME/examples/words.txt" USING TEXT("schema": "STRING word");
        }
        CUBE input BY word AGGREGATES COUNT(word) AS count GROUPING SETS (word);
        STORE input INTO "output" USING TEXT();
END

Installation

Download or clone the repository (say, into /path/to/cubert) and run the following command:

$ cd /path/to/cubert
$ ./gradlew

This will create a folder /path/to/cubert/release, which is what we will need to run cubert. This folder can be copied to hadoop cluster gateway.

To run cubert, first make sure that Hadoop is installed and the HADOOP_HOME environment variable points to the hadoop installation. Set the CUBERT_HOME environment to the release folder (note: CUBERT_HOME points to the release folder and not the "root" repository folder).

$ export CUBERT_HOME=/path/to/cubert/release
$ $CUBERT_HOME/bin/cubert -h
	Using HADOOP_CLASSPATH=:/path/to/cubert/release/lib/*

	usage: ScriptExecutor <cubert script file> [options]
	 -c,--compile             stop after compilation
	 -D <property=value>      use value for given property
	 -d,--debug               print debuging information
	 ...

Example Scripts: Sample scripts are available in the $CUBERT_HOME/examples folder.

Cubert Operators

Cubert provides a rich suite of operators for processing data. These include:

• Input/Output operators: LOAD, STORE, TEE (storing data at intermediate stages of computation into a separate folder) and LOAD-CACHED (read from Hadoop's distributed cache).
• Transformation operators: FROM .. GENERATE (similar to SQL's SELECT), FILTER, LIMIT, TOPN (select top N rows from a window of rows), DISTINCT, SORT, DUPLICATE and FLATTEN (transform nested rows into flattened rows), MERGE JOIN and HASH JOIN.
• Aggregation operators: GROUP BY and CUBE (OLAP cube for additive and non-additive aggregates).
• Data movement operators: SHUFFLE (standard Hadoop shuffle), BLOCKGEN (create partitioned data units, aka blocks), COMBINE (combine two or more blocks), PIVOT (create sub-blocks out of a block, based on specified partitioning scheme).
• Dictionary operators to encode and decode strings to integers.

Coming (really) soon..

• Tez execution engine: the existing cubert scripts (written in Map-Reduce paradigm) can run unmodified on underlying Tez engine.
• Cubert Script v2: a DAG oriented language to express complex composition of workflows.
• Incremental computations: daily running jobs that read multiple days of data can materialize partial output and incrementally compute the results (thus cutting down on resources and time).
• Analytical window functions.

Documentation

Users Guide and Javadoc available at

http://linkedin.github.io/Cubert

Cubert Users Google Group: cubert-users

Email: [email protected]