[SPARK-5037] support dynamic loading of input DStreams in pyspark streaming

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Auto Closed
• Affects Version/s: 1.2.0
• Component/s: DStreams, PySpark
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Jascha Swisher
• Created: 31/Dec/14 17:25
• Updated: 06/Jun/19 13:57
• Resolved: 06/Jun/19 13:57

Description

The scala and java streaming APIs support "external" InputDStreams (e.g. the ZeroMQReceiver example) through a number of mechanisms, for instance by overriding ActorReceiver or just subclassing Receiver directly. The pyspark streaming API does not currently allow similar flexibility, being limited at the moment to file-backed text and binary streams or socket text streams.

It would be great to open up the pyspark streaming API to other stream sources, putting it closer to on par with the JVM APIs.

One way of doing this could be to support dynamically loading InputDStream implementations through reflection at the JVM level, analogously to what is currently done for Hadoop InputFormats in the regular pyspark context.py Hadoop methods.

I'll submit a PR momentarily with my shot at this. Comments and alternative approaches more than welcome.

Issue Links

• links to: [Github] Pull Request #3858 (industrial-sloth)