Closed tomerk closed 9 years ago
@dcrankshaw
Awesome. This looks good to me. Can I assume that it actually works?
The part of using spark as a filesystem to write to/from HDFS definitely works, and it works end to end w/ a local spark master and a local data directory, so it should be working end to end on a remote spark cluster (unclear how quickly though, there's a sizable delay because of needing to ship the jar)
Cool. Merged.
Stripped the entire pluggable KV storage backend system replacing it w/ versioned broadcasting and changed up the model implementation interfaces to lessen the load on model implementers.
Relates to everything but most strongly issue #39