[SPARK-40038] spark.sql.files.maxPartitionBytes does not observe on-disk compression

• Type: Question
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.2.0
• Component/s: Input/Output, Optimizer, PySpark, SQL

Environment

files:

• ORC with snappy compression
• 232 GB files on disk 
• 1800 files on disk (pretty sure no individual file is over 200MB)
• 9 partitions on disk

cluster:

• EMR 6.6.0 (spark 3.2.0)
• cluster: 288 vCPU (executors), 1.1TB memory (executors)

OS info:
LSB Version:    :core-4.1-amd64:core-4.1-noarch:cxx-4.1-amd64:cxx-4.1-noarch:desktop-4.1-amd64:desktop-4.1-noarch:languages-4.1-amd64:languages-4.1-noarch:printing-4.1-amd64:printing-4.1-noarch
Distributor ID:    Amazon
Description:    Amazon Linux release 2 (Karoo)
Release:    2
Codename:    Karoo

• Assignee: Unassigned
• Reporter: RJ Marcus
• Created: 10/Aug/22 22:00
• Updated: 15/May/24 07:38

Description

Why does `spark.sql.files.maxPartitionBytes` estimate the number of partitions based on file size on disk instead of the uncompressed file size?

For example I have a dataset that is 213GB on disk. When I read this in to my application I get 2050 partitions based on the default value of 128MB for maxPartitionBytes. My application is a simple broadcast index join that adds 1 column to the dataframe and writes it out. There is no shuffle.

Initially the size of input /output records seem ok, but I still get a large amount of memory "spill" on the executors. I believe this is due to the data being highly compressed and each partition becoming too big when it is deserialized to work on in memory.

(If I try to do a repartition immediately after reading I still see the first stage spilling memory to disk, so that is not the right solution or what I'm interested in.) 

Instead, I attempt to lower maxPartitionBytes by the (average) compression ratio of my files (about 7x, so let's round up to 8). So I set maxPartitionBytes=16MB.  At this point  I see that spark is reading in from the file in 12-28 MB chunks. Now it makes 14316 partitions on the initial file read and completes with no spillage. 

 
Is there something I'm missing here? Is this just intended behavior? How can I tune my partition size correctly for my application when I do not know how much the data will be compressed ahead of time?

Issue Links

• is related to: SPARK-24914 totalSize is not a good estimate for broadcast joins