Improving Python and Spark Performance and Interoperability: Spark Summit East talk by: Wes McKinney

www.twosigma.com
Improving Python and Spark
Performance and Interoperability
February 9, 2017All Rights Reserved
Wes McKinney @wesmckinn
Spark Summit East 2017
February 9, 2017

Me
February 9, 2017
•  Currently: Software Architect at Two Sigma Investments
•  Creator of Python pandas project
•  PMC member for Apache Arrow and Apache Parquet
•  Other Python projects: Ibis, Feather, statsmodels
•  Formerly: Cloudera, DataPad, AQR
•  Author of Python for Data Analysis
All Rights Reserved 2

Important Legal Information
The information presented here is offered for informational purposes only and should not be
used for any other purpose (including, without limitation, the making of investment decisions).
Examples provided herein are for illustrative purposes only and are not necessarily based on
actual data. Nothing herein constitutes: an offer to sell or the solicitation of any offer to buy any
security or other interest; tax advice; or investment advice. This presentation shall remain the
property of Two Sigma Investments, LP (“Two Sigma”) and Two Sigma reserves the right to
require the return of this presentation at any time.
Some of the images, logos or other material used herein may be protected by copyright and/or
trademark. If so, such copyrights and/or trademarks are most likely owned by the entity that
created the material and are used purely for identification and comment as fair use under
international copyright and/or trademark laws. Use of such image, copyright or trademark does
not imply any association with such organization (or endorsement of such organization) by Two
Sigma, nor vice versa.
Copyright © 2017 TWO SIGMA INVESTMENTS, LP. All rights reserved

This talk
4February 9, 2017
•  Why some parts of PySpark are “slow”
•  Technology that can help make things faster
•  Work we have done to make improvements
•  Future roadmap
All Rights Reserved

Python and Spark
February 9, 2017
•  Spark is implemented in Scala, runs on the Java virtual machine (JVM)
•  Spark has Python and R APIs with partial or full coverage for many parts of
the Scala Spark API
•  In some Spark tasks, Python is only a scripting front-end.
•  This means no interpreted Python code is executed once the Spark
job starts
•  Other PySpark jobs suffer performance and interoperability issues that we’re
going to analyze in this talk

Spark DataFrame performance
Source: https://databricks.com/blog/2015/02/17/introducing-dataframes-in-spark-for-large-scale-data-science.html
6

Spark DataFrame performance can be misleading
February 9, 2017
•  Spark DataFrames are an example of Python as a DSL / scripting front end
•  Excepting UDFs (.map(…) or sqlContext.registerFunction), no Python code is
evaluated in the Spark job
•  Python API calls create SQL query plans inside the JVM — so Scala and
Python versions are computationally identical

Spark DataFrames as deferred DSL
February 9, 2017

young = users[users.age < 21]
young.groupBy(“gender”).count()

February 9, 2017
SELECT gender, COUNT(*)
FROM users
WHERE age < 21
GROUP BY 1

February 9, 2017
Aggregation[table]
table:
Table: users
metrics:
count = Count[int64]
Table: ref_0
by:
gender = Column[array(string)] 'gender' from users
predicates:
Less[array(boolean)]
age = Column[array(int32)] 'age' from users
Literal[int8]
21

Where Python code and Spark meet
February 9, 2017
•  Unfortunately, many PySpark jobs cannot be expressed entirely as
DataFrame operations or other built-in Scala constructs
•  Spark-Scala interacts with in-memory Python in key ways:
•  Reading and writing in-memory datasets to/from the Spark driver
•  Evaluating custom Python code (user-defined functions)

How PySpark lambda functions work
February 9, 2017
•  The anatomy of
All Rights Reserved
rdd.map(lambda x: … )
df.withColumn(py_func(...))
Scala
RDD
Python worker
Python worker
Python worker
Python worker
Python worker
see PythonRDD.scala
12

PySpark lambda performance problems
February 9, 2017
•  See 2016 talk “High Performance Python on Apache Spark”
•  http://www.slideshare.net/wesm/high-performance-python-on-apache-spark
•  Problems
•  Inefficient data movement (serialization / deserialization)
•  Scalar computation model: object boxing and interpreter overhead
•  General summary: PySpark is not currently designed to achieve high
performance in the way that pandas and NumPy are.

Other issues with PySpark lambdas
February 9, 2017
•  Computation model unlike what pandas users are used to
•  In dataframe.map(f), the Python function f only sees one Row at a time
•  A more natural and efficient vectorized API would be:
•  dataframe.map_pandas(lambda df: …)

Apache
Arrow
15

Apache Arrow: Process and Move Data Fast
February 9, 2017
•  New Top-level Apache project as of February 2016
•  Collaboration amongst broad set of OSS projects around shared needs
•  Language-independent columnar data structures
•  Metadata for describing schemas / chunks of data
•  Protocol for moving data between processes with minimal serialization
overhead

High performance data interchange
Today With Arrow
Source: Apache Arrow
17

What does Apache Arrow give you?
February 9, 2017
•  Zero-copy columnar data: Complex table and array data structures that can
reference memory without copying it
•  Ultrafast messaging: Language-agnostic metadata, batch/file-based and
streaming binary formats
•  Complex schema support: Flat and nested data types
•  C++, Python, and Java Implementations: with integration tests

Arrow binary wire formats
February 9, 2017All Rights Reserved 19

Extreme performance to pandas from Arrow streams
February 9, 2017All Rights Reserved 20

PyArrow file and streaming API
February 9, 2017
from pyarrow import StreamReader

reader = StreamReader(stream)

# pyarrow.Table
table = reader.read_all()

# Convert to pandas
df = table.to_pandas()

For illustration purposes only. Not an offer to buy or sell securities. Two Sigma may modify its investment approach and portfolio parameters in the future in any manner that it believes is consistent with its fiduciary duty to its clients. There is no
guarantee that Two Sigma or its products will be successful in achieving any or all of their investment objectives. Moreover, all investments involve some degree of risk, not all of which will be successfully mitigated. Please see the last page of
this presentation for important disclosure information.
Making DataFrame.toPandas faster
February 9, 2017
•  Background
•  Spark’s toPandas transfers in-memory from the Spark driver to Python and
converts it to a pandas.DataFrame. It is very slow
•  Joint work with Bryan Cutler (IBM), Li Jin (Two Sigma), and Yin Xusen
(IBM). See SPARK-13534 on JIRA
•  Test case: transfer 128MB Parquet file with 8 DOUBLE columns

conda install pyarrow -c conda-forge
23

df = sqlContext.read.parquet('example2.parquet')
df = df.cache()
df.count()
Then
%%prun -s cumulative
dfs = [df.toPandas() for i in range(5)]
24

94483943 function calls (94478223 primitive calls) in 62.492 seconds

ncalls tottime percall cumtime percall filename:lineno(function)
5 1.458 0.292 62.492 12.498 dataframe.py:1570(toPandas)
5 0.661 0.132 54.759 10.952 dataframe.py:382(collect)
10485765 0.669 0.000 46.823 0.000 rdd.py:121(_load_from_socket)
715 0.002 0.000 46.139 0.065 serializers.py:141(load_stream)
710 0.002 0.000 45.950 0.065 serializers.py:448(loads)
10485760 4.969 0.000 32.853 0.000 types.py:595(fromInternal)
1391 0.004 0.000 7.445 0.005 socket.py:562(readinto)
18 0.000 0.000 7.283 0.405 java_gateway.py:1006(send_command)
5 0.000 0.000 6.262 1.252 frame.py:943(from_records)
25

Now, using pyarrow

%%prun -s cumulative
dfs = [df.toPandas(useArrow) for i in range(5)]
26

38585 function calls (38535 primitive calls) in 9.448 seconds

5 0.000 0.000 9.358 1.872 dataframe.py:394(collectAsArrow)
6271 9.330 0.001 9.330 0.001 {method 'recv_into' of '_socket.socket‘}
15 0.000 0.000 9.229 0.615 java_gateway.py:860(send_command)
10 0.000 0.000 0.123 0.012 serializers.py:141(load_stream)
5 0.085 0.017 0.089 0.018 {method 'to_pandas' of 'pyarrow.Table‘}
27

pip install memory_profiler
%%memit -i 0.0001
pdf = None
pdf = df.toPandas()
gc.collect()

peak memory: 1223.16 MiB,
increment: 1018.20 MiB

28

Plot thickens: memory use
%%memit -i 0.0001
pdf = None
pdf = df.toPandas(useArrow=True)
gc.collect()

peak memory: 334.08 MiB,
increment: 258.31 MiB

29

Summary of results
February 9, 2017
•  Current version: average 12.5s (10.2 MB/s)
•  Deseralization accounts for 88% of time; the rest is waiting for Spark to
send the data
•  Peak memory use 8x (~1GB) the size of the dataset
•  Arrow version
•  Average wall clock time of 1.89s (6.61x faster, 67.7 MB/s)
•  Deserialization accounts for 1% of total time
•  Peak memory use 2x the size of the dataset (1 memory doubling)
•  Time for Spark to send data 25% higher (1866ms vs 1488 ms)

Aside: reading Parquet directly in Python
February 9, 2017
import pyarrow.parquet as pq

%%timeit
df = pq.read_table(‘example2.parquet’).to_pandas()
10 loops, best of 3: 175 ms per loop


Digging deeper
February 9, 2017
•  Why does it take Spark ~1.8 seconds to send 128MB of data over the wire?
val collectedRows = queryExecution.executedPlan.executeCollect()
cnvtr.internalRowsToPayload(collectedRows, this.schema)
All Rights Reserved
Array[InternalRow]
32

Digging deeper
February 9, 2017
•  In our 128MB test case, on average:
•  75% of time is being spent collecting Array[InternalRow] from the task
executors
•  25% of the time is spent on a single-threaded conversion of all the data
from Array[InternalRow] to ArrowRecordBatch
•  We can go much faster by performing the Spark SQL -> Arrow
conversion locally on the task executors, then streaming the batches to
Python

Future architecture
Task executor
Task executor
Task executor
Task executor
Arrow RecordBatch
Arrow RecordBatch
Arrow RecordBatch
Arrow RecordBatch
Arrow Schema
Spark driver
Python
34

Hot off the presses
5 0.000 0.000 5.838 1.168 dataframe.py:394(collectAsArrow)
5919 0.005 0.000 5.824 0.001 socket.py:561(readinto)
5919 5.809 0.001 5.809 0.001 {method 'recv_into' of '_socket.socket‘}
...
5 0.086 0.017 0.091 0.018 {method 'to_pandas' of 'pyarrow.Table‘}
35
Patch from February 8: 38% perf improvement

The work ahead
February 9, 2017
•  Luckily, speeding up toPandas and speeding up Lambda / UDF functions is
architecturally the same type of problem
•  Reasonably clear path to making toPandas even faster
•  How can you get involved?
•  Keep an eye on Spark ASF JIRA
•  Contribute to Apache Arrow (Java, C++, Python, other languages)
•  Join the Arrow and Spark mailing lists

Thank you
February 9, 2017
•  Bryan Cutler, Li Jin, and Yin Xusen, for building the integration Spark-Arrow
integration
•  Apache Arrow community
•  Spark Summit organizers
•  Two Sigma and IBM, for supporting this work

Improving Python and Spark Performance and Interoperability: Spark Summit East talk by: Wes McKinney

Recommended

More Related Content

What's hot (20)

Similar to Improving Python and Spark Performance and Interoperability: Spark Summit East talk by: Wes McKinney (20)

More from Spark Summit (20)

Recently uploaded (20)

Improving Python and Spark Performance and Interoperability: Spark Summit East talk by: Wes McKinney