Stark: Distributed Inference with Stan & Spark
The idea of stark is to adapt a powerful single-node modeling tool (Stan) for running on a distributed platform (Spark). The naive way of doing this is to copy the data to each worker, and each worker produces poster samples. This emberassingly parallel model could be useful, but running in a distributed setting facilitates other possibilities, such as running on much larger datasets.
One strategy for handling more data is to run the same model where each one is provided a subset of data. Each model will then generate samples from a subposterior. Combining subposteriors is not trivial, and seems to be a current area of research. See the reading section for more resources.
Other possibilties could be running the same model with different datasets or hyperparameters (e.g. for a sensitivity analysis). If MAP estimates were found instead of full posteriors then one might have something akin to a distributed bootstrap.
This project is experimental-quality
Modes
The current status is that only the naive mode and simple weighted averaging are implemented.
Example
Currently there is just a single example, built on the 8 schools dataset. Spark is not useful for such a small dataset, it is used just for illustration.
From the stark/example directory, try:
spark-submit stark_ex.py
(I've only tried this with Spark 1.6.2)
Reading
- Patterns of Scalable Bayesian Inference
- Bayes and Big Data: The Consensus Monte Carlo Algorithm
- Asymptotically exact, embarrassingly parallel MCMC