[
https://issues.apache.org/jira/browse/DRILL-7370?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17003741#comment-17003741
]
Charles Givre commented on DRILL-7370:
--------------------------------------
[~arina]
I have a question for you. The basic idea here is to add a UDF which would
allow Drill to be a direct part of a ML pipeline.
Just as an aside, H2O enables users to export a trained model as a POJO or
MOJO.[1] H2O has provided templates for Hive UDFs [2]
Users could build and train a model using H2O and then write queries like:
{code:sql}
SELECT *
FROM <data>
WHERE PREDICT( <model jar>, field1, field2, field3 ) = 1
{code}
My question is that including this into Drill would require the user to add
external JARs for their models. I had the idea that we could use a similar
structure to the dynamic UDF where by the user defines a model directory and
simply copies the JAR files there. Then in the query, they specify the model
name, and the UDF would look in that directory for the model.
Does that make sense or how would you suggest implementing that?
[1]: [http://docs.h2o.ai/h2o/latest-stable/h2o-docs/productionizing.html]
[2]:
[https://github.com/h2oai/h2o-tutorials/tree/master/tutorials/hive_udf_template]
> Add Generic Predict UDF for H20 ML Models
> -----------------------------------------
>
> Key: DRILL-7370
> URL: https://issues.apache.org/jira/browse/DRILL-7370
> Project: Apache Drill
> Issue Type: Improvement
> Components: Functions - Drill
> Affects Versions: 1.17.0
> Reporter: Charles Givre
> Assignee: Charles Givre
> Priority: Major
>
> h20.ai enables a user to export a trained machine learning model as either a
> POJO or MOJO. The proposed UDF will implement the `predict()` function and
> enable ML predictions to be included in Drill queries.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
