Apache Apex with Apache SAMOA

Traditional Machine Learning

• Batch Oriented
• Supervised - most common
• Training and Scoring
• One time model building
• Data set
• Training: Model building
• Holdout: Paremeter tuning
• Test: Accuracy

Online Machine Learning

• Streaming
• Change
• Dynmaically adapt to new patterns in Data
• Change over time (concept drift)
• Model updates
• Approximation Algorithms
• Single pass: one data item at a time
• sublinear space and time per data item
• small error with high probabilities

Apache SAMOA

• What we need
• Platform for streaming learning Algorithms
• Distributed Scalable

Machine learning classification

               Machine learning
              /                \
     Disribué                   Non distibué
    /        \                   /        \
  Batch   streaming           Batch    streaming
    |         |                 |          |
  Hadoop Apex/flink/storm       |          |
    |         |                 |          |
  Mahout    SAMOA            R, WEKA      MOA

Logical Building Blocks

Each block is a processor (an algorithm), then we create a topology of blocks.

Prequential Eval Tasks in SAMOA

• Interleaved test-then-train
• Evaluates performance for online classifiers
• Basic - Overall
• Sliding Windows based - Most recent

Apex DSPE

Distributed Stream Processing Engine:

• Apex
• Storm
• Flink

Apex Application DAG

• DAG is composed of vertices (Operators) and Edges (Streams)
• Stream is a sequence of data tuples which connects operators

Distribution of tuples

• Calculate Hash of tuple
• Modulo by the number of partition

Iteration support in Apex

• Machine-learning needs iterations
• At the very least, feedback loop
• Apex topology - Acyclic: DAG

Delay Operators

• Increment window id for all outgoing ports
• A note on Fault tolerance

Challenges

Adding DSPE for Apache Apex:

• Differences in the topology builder APIs of SAMOA and Apex
• No concept of Ports in SAMOA
• On demand declaration of streams in SAMOA
• Cycles in topology: Delay Operator
• Serialization of Processor state during checkpointing.

Also serialization of tuples:

• Number of tuples in a single window - Affects number of tuples in future windows coming from the delay operator