Spark 1 1 and Beyond Patrick Wendell About

Spark 1. 1 and Beyond Patrick Wendell

About Me Work at Databricks leading the Spark team Spark 1. 1 Release manager Committer on Spark since AMPLab days

This Talk Spark 1. 1 (and a bit about 1. 2) A few notes on performance Q&A with myself, Tathagata Das, and Josh Rosen

A Bit about Spark… DStream’s: Streams of RDD’s RDD-Based Graphs RDD-Based Matrices Spark Streaming Graph. X MLLib real-time Graph (alpha) machine learning RDD-Based Tables Spark SQL Spark RDD API HDFS, S 3, Cassandra YARN, Mesos, Standalone

Spark Release Process ~3 month release cycle, time-scoped 2 months of feature development 1 month of QA Maintain older branches with bug fixes Upcoming release: 1. 1. 0 (previous was 1. 0. 2)

Master More stable V 1. 1. 0 For any P. O. C or non branch-1. 1 production cluster, we always recommend running off of the head V 1. 0. 0 V 1. 0. 1 of a release branch. More features branch-1. 0

Spark 1. 1 1, 297 patches 200+ contributors (still counting) Dozens of organizations To get updates – join our dev list: E-mail dev-subscribe@spark. apache. org

Roadmap Spark 1. 1 and 1. 2 have similar themes Spark core: Usability, stability, and performance MLlib/SQL/Streaming: Expanded feature set and performance Around ~40% of mailing list traffic is about these libraries.

Spark Core in 1. 1 Performance “out of the box” Sort-based shuffle Efficient broadcasts Disk spilling in Python YARN usability improvements Usability Task progress and user-defined counters UI behavior failing or large jobs

Spark SQL in 1. 1 1. 0 was the first “preview” release 1. 1 provides upgrade path for Shark Replaced Shark in our benchmarks with 2 -3 X perf gains Can perform optimizations with 10100 X less effort than Hive.

Turning an RDD into a Relation • // Define the schema using a case class Person(name: String, age: Int) // Create an RDD of Person objects, register it as a table. val people = sc. text. File("examples/src/main/resources/people. txt"). map(_. split(", "). map(p => Person(p(0), p(1). trim. to. Int)) people. register. As. Table("people")

Querying using SQL • // SQL statements can be run directly on RDD’s val teenagers = sql("SELECT name FROM people WHERE age >= 13 AND age <= 19") // The results of SQL queries are Schema. RDDs and support // normal RDD operations. val name. List = teenagers. map(t => "Name: " + t(0)). collect() • // Language integrated queries (ala LINQ) val teenagers = people. where('age >= 10). where('age <= 19). select('name)

Spark SQL in 1. 1 JDBC server for multi-tenant access and BI tools Native JSON support Public types API – “make your own” Schema RDD’s Improved operator performance Native Parquet support and optimizations

Spark Streaming Stability improvements across the board Amazon Kinesis support Rate limiting for streams Support for polling Flume streams Streaming + ML: Streaming linear regressions

What’s new in MLlib v 1. 1 • Contributors: 40 (v 1. 0) -> 68 • Algorithms: SVD via Lanczos, multiclass support in decision tree, logistic regression with L-BFGS, nonnegative matrix factorization, streaming linear regression • Feature extraction and transformation: scaling, normalization, tf-idf, Word 2 Vec • Statistics: sampling (core), correlations, hypothesis testing, random data generation • Performance and scalability: major improvement to decision tree, tree aggregation • Python API: decision tree, statistics, linear methods

Performance (v 1. 0 vs. v 1. 1)

Sort-based Shuffle Old shuffle: Each mapper opens a file for each reducer and writes output simultaneously. Files = # mappers * # reducers New Shuffle: Each mapper buffers reduce output in memory, spills, then sort-merges on disk data.

Group. By Operator Spark group. By. Key != SQL group. By NO: people. map(p => (p. zip. Code, p. get. Income)). group. By. Key(). map(incomes => incomes. sum) YES: people. map(p => (p. zip. Code, p. get. Income)). reduce. By. Key(_ + _)

Group. By Operator Spark group. By. Key != SQL group. By NO: people. map(p => (p. zip. Code, p. get. Income)). group. By. Key(). map(incomes => incomes. sum) YES: people. group. By(‘zip. Code). select(sum(‘inc ome))

Group. By Operator Spark group. By. Key != SQL group. By NO: people. map(p => (p. zip. Code, p. get. Income)). group. By. Key(). map(incomes => incomes. sum) YES: SELECT sum(income) FROM people GROUP BY zip. Code;

Other efforts Ooyala Job Server Hive on Spark Pig on Spark DStream’s: Streams of RDD’s RDD-Based Graphs RDD-Based Matrices Spark Streaming Graph. X MLLib real-time Graph (alpha) machine learning RDD-Based Tables Spark SQL Spark RDD API HDFS, S 3, Cassandra YARN, Mesos, Standalone

Looking Ahead to 1. 2+ [Core] Scala 2. 11 support Debugging tools (task progress, visualization) Netty-based communication layer [SQL] Portability across Hive versions Performance optimizations (TPC-DS and Parquet) Planner integration with Cassandra and other sources

Looking Ahead to 1. 2+ [Streaming] Python Support Lower level Kafka API w/ recoverability [MLLib] Multi-model training Many new algorithms Faster internal linear solver

Q and A Josh Rosen Py. Spark and Spark Core Tathagata Das Spark Streaming Lead