Spark: save DataFrame partitioned by “virtual” column

Question

I'm using PySpark to do classic ETL job (load dataset, process it, save it) and want to save my Dataframe as files/directory partitioned by a "virtual" column; what I mean by "virtual" is that I have a column Timestamp which is a string containing an ISO 8601 encoded date, and I'd want to partition by Year / Month / Day; but I don't actually have either a Year, Month or Day column in the DataFrame; I have this Timestamp from which I can derive these columns though, but I don't want my resultat items to have one of these columns serialized.

File structure resulting from saving the DataFrame to disk should look like:

/ 
    year=2016/
        month=01/
            day=01/
                part-****.gz

Is there a way to do what I want with Spark / Pyspark ?

Answer 1

Columns which are used for partitioning are not included in the serialized data itself. For example if you create DataFrame like this:

df = sc.parallelize([
    (1, "foo", 2.0, "2016-02-16"),
    (2, "bar", 3.0, "2016-02-16")
]).toDF(["id", "x", "y", "date"])

and write it as follows:

import tempfile
from pyspark.sql.functions import col, dayofmonth, month, year
outdir = tempfile.mktemp()

dt = col("date").cast("date")
fname = [(year, "year"), (month, "month"), (dayofmonth, "day")]
exprs = [col("*")] + [f(dt).alias(name) for f, name in fname]

(df
    .select(*exprs)
    .write
    .partitionBy(*(name for _, name in fname))
    .format("json")
    .save(outdir))

individual files won't contain partition columns:

import os

(sqlContext.read
    .json(os.path.join(outdir, "year=2016/month=2/day=16/"))
    .printSchema())

## root
##  |-- date: string (nullable = true)
##  |-- id: long (nullable = true)
##  |-- x: string (nullable = true)
##  |-- y: double (nullable = true)

Partitioning data is stored only in a directory structure and not duplicated in serialized files. It will be attached only when your read complete or partial directory tree:

sqlContext.read.json(outdir).printSchema()

## root
##  |-- date: string (nullable = true)
##  |-- id: long (nullable = true)
##  |-- x: string (nullable = true)
##  |-- y: double (nullable = true)
##  |-- year: integer (nullable = true)
##  |-- month: integer (nullable = true)
##  |-- day: integer (nullable = true)

sqlContext.read.json(os.path.join(outdir, "year=2016/month=2/")).printSchema()

## root
##  |-- date: string (nullable = true)
##  |-- id: long (nullable = true)
##  |-- x: string (nullable = true)
##  |-- y: double (nullable = true)
##  |-- day: integer (nullable = true)