[CTAKES-374] Scaleout of cTAKES pipeline

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: future enhancement
• Fix Version/s: 3.2.1
• Assignee: Unassigned
• Reporter: Selina Chu
• Created: 18/Aug/15 18:09
• Updated: 19/Aug/15 03:14

Description

Currently, cTAKES can't be easily deployed in an asynchronous manner. UIMA components aren't serializable (and thus cTAKES' components as well). Would like to come up with better ways to allow cTAKES to be easily run in a distributed fashion.

For example, for processing a long document (e.g. 10+ pages), cTAKES would take a long time to process.

I would like to see a feature where we can partition the input to cTAKES, in a way that won't affect the cTAKES annotation performance, allowing us to process through a cluster running in distributed mode (e.g. Spark streaming cTAKES). And then recombine the results such that the word/phrase token positions will be sequentially ordered.

We have a simple implementation of the ClinicalPipelineFactory with Spark Streaming. Currently our initial attempt in partitioning is by paragraphs. For example, we are doing something like:
RDD.map(a_single_paragraph.process_in_ctakes())

I also wanted to see if there are any better ways of doing this.

Issue Links

• is related to: CTAKES-314 BigTop/Hadoop cTAKES integration