Distributed Execution on Beam

[alxmrs]

Figure out a way to distribute all layers of SQL execution #10 on Apache Beam.

[alxmrs]

Dataframes: https://beam.apache.org/documentation/dsls/dataframes/overview/ Xarray: Xarray-Beam

[alxmrs]

Beam's dataframes library supports multi indexes.

https://beam.apache.org/releases/pydoc/current/apache_beam.dataframe.io.html

This alone makes beam worthy of an exploration sooner rather than later.

[cisaacstern]

Interesting!

[alxmrs]

Some general thoughts on this issue in no particular order:

• I think this would make a good 0.1 release
• Users can call xarray_sql.beam.read_xarray()
• Beam supports the Pandas API, with small differences, primarily specific to the fact that PCollections are unordered.
• Unknown: is there a from_map-like interface to make implementing this easy?

[alxmrs]

This may not be feasible after all. It looks like hdf5 is intentionally not supported because it is a random access format. I think Xarray would follow this characteristic, too.

https://beam.apache.org/releases/pydoc/current/_modules/apache_beam/dataframe/io.html

Maybe this warrants the creation of an xarray-beam-like library for pandas or dask? Can a pd.(multi)index mimic an xbeam key?

[alxmrs]

A core question to answer: do we really need random access?

[alxmrs]

I think Ray is a better, easier fit (see #68). Closing.