Apache Spark MLlib Model File Format

Question

Apache Spark MLlib algorithms (e.g., Decision Trees) save the model in a location (e.g., myModelPath) where it creates two directories, viz. myModelPath/d

Answer 1

Spark >= 2.4

Since Spark 2.4 provides format agnostic writer interfaces and selected models already implement these. For example LinearRegressionModel:

val lrm: org.apache.spark.ml.regression.LinearRegressionModel = ???
val path: String = ???

lrm.write.format("pmml").save(path)

will create a directory with a single file containing PMML representation.

Spark < 2.4

What are the format of these files?

• data/*.parquet files are in Apache Parquet columnar storage format
• metadata/part-* looks like JSON

Which file/files contain actual model?

• model/*.parquet

Can I save the model to somewhere else, for example in a DB?

I am not aware of any direct method but you can load model as a data frame and store it in a database afterwards:

val modelDf = spark.read.parquet("/path/to/data/")
modelDf.write.jdbc(...)