Kafka topic should be consumed only once to create the rocksdb state.
Once created, state is check pointed and each time job is cancelled with a save point
each time job is run from a save point.

• What's the best way to achieve this
  Find a way to load models from disk into the operator partition memory

Implement SVM Model serving using queryable state

Migrate the code base to Flink 1.3.1
Check if the existing code of Queryable client needs any change
Build Flink 1.3.1 from source with RocksJava patch on cluster

So far:

1. Updated the SVMpredictRandom logic
2. Added Rebalance option in kafka Consumers

Work load

Recommender System

Large Vector Model (NLP maybe)

Model Deployment

Model outside of the processing system (Parameter Server Style)

Model inside processing system

Model Distribution

Partitioned Model

Boardcast

• Underlying storage (K/V store)
  Find out what are the requirement for the parameter server (storage requirement, amount of traffic, ...)

• Flink architecture
  How does the processing part look like

• Start with Recommender System use case
  Try to make it robust enough for other types of models (Sparse Vectors)