As shown above, Angel's core API classes, ordered by when (in general) they are called during model training, include:

1. MLRunner
   • MLRunner creates AngelClient with factory class based on conf, and calls AngelClient's interfaces in order according to the standard train process

• AngelClient

  • Starts PSServer
  • Initializes PSServer and loads empty model
  • After training, saves the model to HDFS from multiple PSServers

• TrainTask

  • Starts train process when called by AngelClient

• DataBlock

  • TrainTask calls parse and preProcess methods to read data from HDFS, and assemble data into DataBlock that contains multiple LabeledData
  • TrainTask calls train method to create, and pass DataBlock to, the MLLearner object

• MLLearner

  • MLLearner calls its own learn method, reads DataBlock, computes the model delta, and pushes to / pull from PSServer through PSModel inside MLModel, eventually obtaining a complete MLModel

• MLModel

  • According to the algorithm's need, creates and holds multiple PSModels

• PSModel

  • Encapsulates all the interfaces in AngelClient that communicate with PSServer, facilitating MLLearner calls

Understanding these core classes and processes will be quite helpful for implementing high performance machine-learning algorithms that can run on Angel.