Skip to content

Latest commit

 

History

History
140 lines (108 loc) · 5.15 KB

README.md

File metadata and controls

140 lines (108 loc) · 5.15 KB
Raw

RepartiPy

RepartiPy helps you to elaborately handle PySpark DataFrame partition size.

Possible Use Cases

  • Repartition your DataFrame precisely, without knowing the whole DataFrame size (i.e. Dynamic Repartition)
  • Estimate your DataFrame size with more accuracy

Why RepartiPy

Although Spark SizeEstimator can be used to estimate a DataFrame size, it is not accurate sometimes. RepartiPy uses Spark's execution plan statistics in order to provide a roundabout way. It suggests two approaches to achieve this:

  • reaprtipy.SizeEstimator
  • reaprtipy.SamplingSizeEstimator

reaprtipy.SizeEstimator

Recommended when your executor resource (memory) is affordable to cache the whole DataFrame. SizeEstimator just simply caches the whole Dataframe into the memory and extract the execution plan statistics.

repartipy.SamplingSizeEstimator

Recommended when your executor resource (memory) is NOT affordable to cache the whole dataframe. SamplingSizeEstimator uses 'disk write and re-read (HDFS)' approach behind the scene for two reasons:

  1. Prevent double read from the source like S3, which might be inefficient -> better performance
  2. Reduce partition skewness by reading data again on purpose (leverage MaxPartitionBytes) -> better sampling result

Therefore, you must have HDFS settings on your cluster and enough disk space.

This may not be accurate compared to SizeEstimator due to sampling. If you want more accurate results, tune the sample_count option properly. Additionally, this approach will be slower than SizeEstimator as SamplingSizeEstimator requires disk I/O and additional logics.

How To Use

Setup

pip install repartipy

Prerequisite

from pyspark.sql import SparkSession

spark = SparkSession.builder.getOrCreate()
input_data = [
        (1, "Seoul"),
        (2, "Busan"),
    ]
df = spark.createDataFrame(data=input_data, schema=["id", "location"])

get_desired_partition_count()

Calculate ideal number of partitions for a DataFrame

SizeEstimator will suggest desired_partition_count, so that each partition can have desired_partition_size_in_bytes (default: 1GiB) after repartition. reproduce() produces exactly the same df, but internally reproduced by SizeEstimator for better performance. SizeEstimator reproduces df from Memory (Cache). SamplingSizeEstimator reproduces df from Disk (HDFS).

with SizeEstimator

import repartipy

one_gib_in_bytes = 1073741824

with repartipy.SizeEstimator(spark=spark, df=df) as se:
    desired_partition_count = se.get_desired_partition_count(desired_partition_size_in_bytes=one_gib_in_bytes)
    
    se.reproduce().repartition(desired_partition_count).write.save("your/write/path")
    # or 
    se.reproduce().coalesce(desired_partition_count).write.save("your/write/path")

with SamplingSizeEstimator

import repartipy
    
one_gib_in_bytes = 1073741824

with repartipy.SamplingSizeEstimator(spark=spark, df=df, sample_count=10) as se:
    desired_partition_count = se.get_desired_partition_count(desired_partition_size_in_bytes=one_gib_in_bytes)
    
    se.reproduce().repartition(desired_partition_count).write.save("your/write/path")
    # or 
    se.reproduce().coalesce(desired_partition_count).write.save("your/write/path")

estimate()

Estimate size of a DataFrame

with SizeEstimator

import repartipy

with repartipy.SizeEstimator(spark=spark, df=df) as se:
    df_size_in_bytes = se.estimate()

with SamplingSizeEstimator

import repartipy

with repartipy.SamplingSizeEstimator(spark=spark, df=df, sample_count=10) as se:
    df_size_in_bytes = se.estimate()

Benchmark

Overall, there appears to be a slight performance loss when employing RepartiPy. This benchmark compares the running time of spark jobs in the following two cases to give a rough estimate:

  • Static Repartition (repartition without RepartiPy)
# e.g.
df.repartition(123).write.save("your/write/path")
  • Dynamic Repartition (repartition with RepartiPy)
# e.g.
with repartipy.SizeEstimator(spark=spark, df=df) as se:
    desired_partition_count = se.get_desired_partition_count(desired_partition_size_in_bytes=one_gib_in_bytes)
    se.reproduce().repartition(desired_partition_count).write.save("your/write/path")

All the other conditions remain the same except the usage of RepartiPy.

Note

Benchmark results provided are for brief reference only, not absolute. Actual performance metrics can vary depending on your own circumstances (e.g. your data, your spark code, your cluster resources, ...).

SizeEstimator

  • DataFrame Size ~= 256 MiB (decompressed size)
Static Dynamic
Running Time 8.5 min 8.6 min

SamplingSizeEstimator

  • DataFrame Size ~= 241 GiB (decompressed size)
Static Dynamic
Running Time 14 min 16 min