Improving PySpark Performance - Spark Beyond the JVM @ PyData DC 2016

Improving PySpark
Performance
(Spark beyond the JVM)
PyDataDC 2016

Who am I?
● My name is Holden Karau
● Prefered pronouns are she/her
● I’m a Principal Software Engineer at IBM’s Spark Technology Center
● previously Alpine, Databricks, Google, Foursquare & Amazon
● co-author of Learning Spark & Fast Data processing with Spark
○ co-author of a new book focused on Spark performance coming this next year*
● @holdenkarau
● Slide share http://www.slideshare.net/hkarau
● Linkedin https://www.linkedin.com/in/holdenkarau
● Github https://github.com/holdenk
● Spark Videos http://bit.ly/holdenSparkVideos

What is going to be covered:
● What I think I might know about you
● A quick background of how PySpark works
● RDD re-use (caching, persistence levels, and checkpointing)
● Working with key/value data
○ Why group key is evil and what we can do about it
● When Spark SQL can be amazing and wonderful
● A brief introduction to Datasets (new in Spark 1.6)
● Calling Scala code from Python with Spark
● How we can make PySpark go fast in the future (vroom vroom)
Torsten Reuschling

Who I think you wonderful humans are?
● Nice people - we are at PyData conference :)
● Don’t mind pictures of cats
● Might know some Apache Spark
● Want to scale your Apache Spark jobs
● Don’t overly mind a grab-bag of topics
Lori Erickson

What is Spark?
● General purpose distributed system
○ With a really nice API including Python :)
● Apache project (one of the most
active)
● Must faster than Hadoop
Map/Reduce
● Good when too big for a single
machine
● Built on top of two abstractions for
distributed data: RDDs & Datasets

The different pieces of Spark
Apache Spark
SQL &
DataFrames
Streaming
Language
APIs
Scala,
Java,
Python, &
R
Graph
Tools
Spark ML
bagel &
Graph X
MLLib
Community
Packages
Jon Ross

SparkContext: entry to the world
● Can be used to create distributed data from many input
sources
○ Native collections, local & remote FS
○ Any Hadoop Data Source
● Also create counters & accumulators
● Automatically created in the shells (called sc)
● Specify master & app name when creating
○ Master can be local[*], spark:// , yarn, etc.
○ app name should be human readable and make sense
● etc.
Petfu
l

RDDs: Spark’s Primary abstraction
RDD (Resilient Distributed Dataset)
● Distributed collection
● Recomputed on node failure
● Distributes data & work across the cluster
● Lazily evaluated (transformations & actions)
Helen Olney

What’s new for PySpark in 2.0?
● Newer Py4J bridge
● SparkSession now replaces SQLContext & HiveContext
● DataFrame/SQL speedups
● Better filter push downs in SQL
● Much better ML interop
● Streaming DataFrames* (ALPHA)
● WARNING: Slightly Different Persistence Levels
● And a bunch more :)

Improving PySpark Performance - Spark Beyond the JVM @ PyData DC 2016

A detour into PySpark’s internals
Photo by Bill Ward

Spark in Scala, how does PySpark work?
● Py4J + pickling + magic
○ This can be kind of slow sometimes
● RDDs are generally RDDs of pickled objects
● Spark SQL (and DataFrames) avoid some of this
kristin klein

So what does that look like?
Driver
py4j
Worker 1
Worker K
pipe
pipe

So how does that impact PySpark?
● Data from Spark worker serialized and piped to Python
worker
○ Multiple iterator-to-iterator transformations are still pipelined :)
● Double serialization cost makes everything more
expensive
● Python worker startup takes a bit of extra time
● Python memory isn’t controlled by the JVM - easy to go
over container limits if deploying on YARN or similar
● Error messages make ~0 sense
● etc.

Cat photo from http://galato901.deviantart.com/art/Cat-on-Work-Break-173043455
Photo from Cocoa Dream

Lets look at some old stand bys:
words = rdd.flatMap(lambda x: x.split(" "))
wordPairs = words.map(lambda w: (w, 1))
grouped = wordPairs.groupByKey()
grouped.mapValues(lambda counts: sum(counts))
grouped.saveAsTextFile("counts")
warnings = rdd.filter(lambda x: x.lower.find("warning") !=
-1).count()
Tomomi

RDD re-use - sadly not magic
● If we know we are going to re-use the RDD what should we do?
○ If it fits nicely in memory caching in memory
○ persisting at another level
■ MEMORY, MEMORY_AND_DISK
○ checkpointing
○ The options changed in Spark 2.0 (we can’t easily specify serialized anymore since there is no
benefit on RDDs - but things get complicated when sharing RDDs or working with
DataFrames)
● Noisey clusters
○ _2 & checkpointing can help
● persist first for checkpointing
Richard Gillin

What is key skew and why do we care?
● Keys aren’t evenly distributed
○ Sales by zip code, or records by city, etc.
● groupByKey will explode (but it's pretty easy to break)
● We can have really unbalanced partitions
○ If we have enough key skew sortByKey could even fail
○ Stragglers (uneven sharding can make some tasks take much longer)
Mitchell
Joyce

groupByKey - just how evil is it?
● Pretty evil
● Groups all of the records with the same key into a single record
○ Even if we immediately reduce it (e.g. sum it or similar)
○ This can be too big to fit in memory, then our job fails
● Unless we are in SQL then happy pandas
PROgeckoam

So what does that look like?
(94110, A, B)
(94110, A, C)
(10003, D, E)
(94110, E, F)
(94110, A, R)
(10003, A, R)
(94110, D, R)
(94110, E, R)
(94110, E, R)
(67843, T, R)
(94110, T, R)
(94110, T, R)
(67843, T, R)(10003, A, R)
(94110, [(A, B), (A, C), (E, F), (A, R), (D, R), (E, R), (E, R), (T, R) (T, R)]
Tomomi

“Normal” Word count w/RDDs
lines = sc.textFile(src)
words = lines.flatMap(lambda x: x.split(" "))
word_count =
(words.map(lambda x: (x, 1))
.reduceByKey(lambda x, y: x+y))
word_count.saveAsTextFile(output)
No data is read or
processed until after
this line
This is an “action”
which forces spark to
evaluate the RDD
These are still
pipelined
inside of the
same python
executor
Trish Hamme

So what did we do instead?
● reduceByKey
○ Works when the types are the same (e.g. in our summing version)
● aggregateByKey
○ Doesn’t require the types to be the same (e.g. computing stats model or similar)
Allows Spark to pipeline the reduction & skip making the list
We also got a map-side reduction (note the difference in shuffled read)

Can just the shuffle cause problems?
● Sorting by key can put all of the records in the same partition
● We can run into partition size limits (around 2GB)
● Or just get bad performance
● So we can handle data like the above we can add some “junk” to our key
(94110, A, B)
(94110, A, C)
(10003, D, E)
(94110, E, F)
(94110, A, R)
(10003, A, R)
(94110, D, R)
(94110, E, R)
(94110, E, R)
(67843, T, R)
(94110, T, R)
(94110, T, R)
PROTodd
Klassy

Shuffle explosions :(
(94110, A, B)
(94110, A, C)
(10003, D, E)
(94110, E, F)
(94110, A, R)
(10003, A, R)
(94110, D, R)
(94110, E, R)
(94110, E, R)
(67843, T, R)
(94110, T, R)
(94110, T, R)
(94110, A, B)
(94110, A, C)
(94110, E, F)
(94110, A, R)
(94110, D, R)
(94110, E, R)
(94110, E, R)
(94110, T, R)
(94110, T, R)
(67843, T, R)(10003, A, R)
(10003, D, E)
javier_artiles

100% less explosions
(94110, A, B)
(94110, A, C)
(10003, D, E)
(94110, E, F)
(94110, A, R)
(10003, A, R)
(94110, D, R)
(94110, E, R)
(94110, E, R)
(67843, T, R)
(94110, T, R)
(94110, T, R)
(94110_A, A, B)
(94110_A, A, C)
(94110_A, A, R)
(94110_D, D, R)
(94110_T, T, R)
(10003_A, A, R)
(10003_D, D, E)
(67843_T, T, R)
(94110_E, E, R)
(94110_E, E, R)
(94110_E, E, F)
(94110_T, T, R)
Jennifer Williams

Well there is a bit of magic in the shuffle….
● We can reuse shuffle files
● But it can (and does) explode*
Sculpture by Flaming Lotus Girls
Photo by Zaskoda

Our saviour from serialization: DataFrames
● For the most part keeps data in the JVM
○ Notable exception is UDFs written in Python
● Takes our python calls and turns it into a query plan
● If we need more than the native operations in Spark’s
DataFrames
● be wary of Distributed Systems bringing claims of
usability….
Andy
Blackledge

So what are Spark DataFrames?
● More than SQL tables
● Not Pandas or R DataFrames
● Semi-structured (have schema information)
● tabular
● work on expression instead of lambdas
○ e.g. df.filter(df.col(“happy”) == true) instead of rdd.filter(lambda x:
x.happy == true))
● Not a subset of “Datasets” - since Dataset API isn’t
exposed in Python yet :(
Quinn Dombrowski

Why are DataFrames good for performance?
● Space efficient columnar cached representation
● Able to push down operations to the data store
● Reduced serialization/data transfer overhead
● Able to perform some operations on serialized data
● Optimizer is able to look inside of our operations
○ Regular spark can’t see inside our operations to spot the difference between (min(_, _)) and
(append(_, _))

How much faster can it be? (Python)
Andrew Skudder

Loading with sparkSQL & spark-csv
df = sqlContext.read
.format("com.databricks.spark.csv")
.option("header", "true")
.option("inferSchema", "true")
.load("resources/adult.data")
Jess Johnson

What about other data formats?
● Built in
○ Parquet
○ JDBC
○ Json (which is amazing!)
○ Orc
○ Hive
● Available as packages
○ csv*
○ Avro, Redshift, Mongo, Cassandra, Cloudant, Couchbase, etc.
○ +34 at
http://spark-packages.org/?q=tags%3A%22Data%20Sources%22
Michael Coghlan
*pre-2.0 package, 2.0+ built in hopefully

Ok so we’ve got our Data, what now?
● We can inspect the Schema
● We can start to apply some transformations (relational)
● We can do some machine learning
● We can jump into an RDD or a Dataset for functional
transformations
● We could wordcount - again!

Getting the schema
● printSchema() for human readable
● schema for machine readable

Sample json record
{"name":"mission",
"pandas":[{"id":1,"zip":"94110","pt":"giant",
"happy":true, "attributes":[0.4,0.5]}]}
Xiahong Chen

Word count w/Dataframes
df = sqlCtx.read.load(src)
# Returns an RDD
words = df.select("text").flatMap(lambda x: x.text.split(" "))
words_df = words.map(
lambda x: Row(word=x, cnt=1)).toDF()
word_count = words_df.groupBy("word").sum()
word_count.write.format("parquet").save("wc.parquet")
Still have the double
serialization here :(
*(Also in 2.0 have to
explicitly switch to
RDD)

Buuuut….
● UDFs / custom maps will be “slow” (e.g. require data
copy from executor and back)
Nick
Ellis

Mixing Python & JVM code FTW:
● DataFrames are an example of pushing our processing
to the JVM
● Python UDFS & maps lose this benefit
● But we can write Scala UDFS and call them from
Python
○ py4j error messages can be difficult to understand :(
● Work to make JVM UDFs easier to register in PR #9766
● Trickier with RDDs since stores pickled objects

Exposing functions to be callable from
Python:
// functions we want to be callable from python
object functions {
def kurtosis(e: Column): Column = new
Column(Kurtosis(EvilSqlTools.getExpr(e)))
def registerUdfs(sqlCtx: SQLContext): Unit = {
sqlCtx.udf.register("rowKurtosis", helpers.rowKurtosis _)
}
}
Fiona
Henderson

Calling the functions with py4j*:
● The SparkContext has a reference to the jvm (_jvm)
● Many Python objects which are wrappers of JVM
objects have _j[objtype] to get the JVM object
○ rdd._jrdd
○ df._jdf
○ sc._jsc
● These are private and may change
*The py4j bridge only exists on the driver**
** Not exactly true but close enough
Fiona
Henderson

e.g.:
def register_sql_extensions(sql_ctx):
scala_sql_context = sql_ctx._ssql_ctx
spark_ctx = sql_ctx._sc
(spark_ctx._jvm.com.sparklingpandas.functions
.registerUdfs(scala_sql_context))

More things to keep in mind with DFs (in Python)
● Schema serialized as json from JVM
● toPandas is essentially collect
● joins can result in the cross product
○ big data x big data =~ out of memory
● Pre 2.0: Use the HiveContext
○ you don’t need a hive install
○ more powerful UDFs, window functions, etc.

The “future*”: Awesome UDFs
● Work going on in Scala land to translate simple Scala
into SQL expressions - need the Dataset API
○ Maybe we can try similar approaches with Python?
● Very early work going on to use Jython for simple UDFs
(e.g. 2.7 compat & no native libraries) - SPARK-15369
○ Early benchmarking w/word count 5% slower than native Scala UDF,
close to 2x faster than regular Python
● Willing to share your Python UDFs for benchmarking? -
http://bit.ly/pySparkUDF
*The future may or may not have better performance than today. But bun-bun the bunny has some lettuce so its
ok!

Want to help with reviewing the code?
● https://github.com/apache/spark/pull/13571
Some open questions:
● Do we want to make the Jython dependency optional?
● If so how do we want people to load it?
● Do we want to fall back automatically on Jython failure?
E-mail me: holden@pigscanfly.ca :)

The “future*”: Faster interchange
● Faster interchange between Python and Spark (e.g.
Tungsten + Apache Arrow)? (SPARK-13391 &
SPARK-13534)
● Willing to share your Python UDFs for benchmarking? -
● Dask integration?
*The future may or may not have better performance than today. But bun-bun the bunny has some lettuce so its
ok!

Additional Spark Resources
● Programming guide (along with JavaDoc, PyDoc,
ScalaDoc, etc.)
○ http://spark.apache.org/docs/latest/
● Books
● Videos
● Spark Office Hours
○ Normally in the bay area - will do Google Hangouts ones soon
○ follow me on twitter for future ones - https://twitter.com/holdenkarau
raider of gin

Learning Spark
Fast Data
Processing with
Spark
(Out of Date)
Fast Data
Processing with
Spark
(2nd edition)
Advanced
Analytics with
Spark
Coming soon:
Spark in Action
Coming soon:
High Performance Spark

And the next book…..
First five chapters are available in “Early Release”*:
● Buy from O’Reilly - http://bit.ly/highPerfSpark
● Chapter 9(ish) - Going Beyond Scala
Get notified when updated & finished:
● http://www.highperformancespark.com
● https://twitter.com/highperfspark
* Early Release means extra mistakes, but also a chance to help us make a more awesome
book.

Spark Videos
● Apache Spark Youtube Channel
● My Spark videos on YouTube -
○ http://bit.ly/holdenSparkVideos
● Spark Summit 2014 training
● Paco’s Introduction to Apache Spark

k thnx bye!
If you care about Spark testing and
don’t hate surveys:
http://bit.ly/holdenTestingSpark
Will tweet results
“eventually” @holdenkarau
PySpark Users: Have some simple
UDFs you wish ran faster you are
willing to share?:

Improving PySpark Performance - Spark Beyond the JVM @ PyData DC 2016

More Related Content

What's hot (20)

Similar to Improving PySpark Performance - Spark Beyond the JVM @ PyData DC 2016 (20)

Recently uploaded (20)

Improving PySpark Performance - Spark Beyond the JVM @ PyData DC 2016