sparkle: Apache Spark applications in Haskell

sparkle [spär′kəl]: a library for writing resilient analytics applications in Haskell that scale to thousands of nodes, using Spark and the rest of the Apache ecosystem under the hood. See this blog post for the details.

Getting started

The tl;dr using the hello app as an example on your local machine:

$ nix-shell --pure --run "bazel build //apps/hello:sparkle-example-hello_deploy.jar"
$ nix-shell --pure --run "bazel run spark-submit -- --packages com.amazonaws:aws-java-sdk:1.11.920,org.apache.hadoop:hadoop-aws:2.10.2 $PWD/bazel-bin/apps/hello/sparkle-example-hello_deploy.jar"

You'll need Nix for the above to work.

How it works

sparkle is a tool for creating self-contained Spark applications in Haskell. Spark applications are typically distributed as JAR files, so that's what sparkle creates. We embed Haskell native object code as compiled by GHC in these JAR files, along with any shared library required by this object code to run. Spark dynamically loads this object code into its address space at runtime and interacts with it via the Java Native Interface (JNI).

How to use

To run a Spark application the process is as follows:

1. create an application in the apps/ folder, in-repo or as a submodule;
2. build the app;
3. submit it to a local or cluster deployment of Spark.

If you run into issues, read the Troubleshooting section below first.

Build

Linux

Include the following in a BUILD.bazel file next to your source code.

package(default_visibility = ["//visibility:public"])

load(
  "@rules_haskell//haskell:defs.bzl",
  "haskell_library",
)

load("@io_tweag_sparkle//:sparkle.bzl", "sparkle_package")

# hello-hs needs to contain a Main module with a main function.
# This main function will be invoked by spark.
haskell_library (
  name = "hello-hs",
  srcs = ...,
  deps = ...,
  ...
)

sparkle_package(
  name = "sparkle-example-hello",
  src = ":hello-hs",
)

You might want to add the following settings to your .bazelrc.local file.

common --repository_cache=~/.bazel_repo_cache
common --disk_cache=~/.bazel_disk_cache
common --local_cpu_resources=4

And then ask Bazel to build a deploy jar file.

$ nix-shell --pure --run "bazel build //apps/hello:sparkle-example-hello_deploy.jar"

Other platforms

sparkle builds in Mac OS X, but running it requires installing binaries for Spark and maybe Hadoop (See .github/workflows/build.yml.

Another alternative is to build and run sparkle via Docker in non-Linux platforms, using a docker image provisioned with Nix.

Integrating sparkle in another project

As sparkle interacts with the JVM, you need to tell ghc where JVM-specific headers and libraries are. It needs to be able to locate jni.h, jni_md.h and libjvm.so.

sparkle uses inline-java to embed fragments of Java code in Haskell modules, which requires running the javac compiler, which must be available in the PATH of the shell. Moreover, javac needs to find the Spark classes that inline-java quotations refer to. Therefore, these classes need to be added to the CLASSPATH when building sparkle. Dependending on your build system, how to do this might vary. In this repo, we use gradle to install Spark, and we query gradle to get the paths we need to add to the CLASSPATH.

Additionally, the classes need to be found at runtime to load them. The main thread can find them, but other threads need to invoke initializeSparkThread or runInSparkThread from Control.Distributed.Spark.

If the main function terminates with unhandled exceptions, they can be propagated to Spark with Control.Distributed.Spark.forwardUnhandledExceptionsToSpark. This allows spark both to report the exception and to cleanup before termination.

Submit

Finally, to run your application, for example locally:

$ nix-shell --pure --run "bazel run spark-submit -- /path/to/$PWD/<app-target-name>_deploy.jar"

The <app-target-name> is the name of the Bazel target producing the jar file. See apps in the apps/ folder for examples.

RTS options can be passed as a java property

$ nix-shell --pure --run "bazel run spark-submit -- --driver-java-options=-Dghc_rts_opts='+RTS\ -s\ -RTS' <app-target-name>_deploy.jar

or as command line arguments

$ nix-shell --pure --run "bazel run spark-submit -- <app-target-name>_deploy.jar +RTS -s -RTS

See here for other options, including launching a whole cluster from scratch on EC2. This blog post shows you how to get started on the Databricks hosted platform and on Amazon's Elastic MapReduce.

Troubleshooting

JNI calls in auxiliary threads fail with ClassNotFoundException

The context class loader of threads needs to be set appropriately before JNI calls can find classes in Spark. Calling initializeSparkThread or runInSparkThread from Control.Distributed.Spark should set it.

Anonymous classes in inline-java quasiquotes fail to deserialize

When using inline-java, it is recommended to use the Kryo serializer, which is currently not the default in Spark but is faster anyways. If you don't use the Kryo serializer, objects of anonymous class, which arise e.g. when using Java 8 function literals,

foo :: RDD Int -> IO (RDD Bool)
foo rdd = [java| $rdd.map((Integer x) -> x.equals(0)) |]

won't be deserialized properly in multi-node setups. To avoid this problem, switch to the Kryo serializer by setting the following configuration properties in your SparkConf:

do conf <- newSparkConf "some spark app"
   confSet conf "spark.serializer" "org.apache.spark.serializer.KryoSerializer"
   confSet conf "spark.kryo.registrator" "io.tweag.sparkle.kryo.InlineJavaRegistrator"

See #104 for more details.

java.lang.UnsatisfiedLinkError: /tmp/sparkle-app...: failed to map segment from shared object

Sparkle unzips the Haskell binary program in a temporary location on the filesystem and then loads it from there. For loading to succeed, the temporary location must not be mounted with the noexec option. Alternatively, the temporary location can be changed with

spark-submit --driver-java-options="-Djava.io.tmpdir=..." \
             --conf "spark.executor.extraJavaOptions=-Djava.io.tmpdir=..."

java.io.IOException: No FileSystem for scheme: s3n

Spark 2.4 requires explicitly specifying extra JAR files to spark-submit in order to work with AWS. To work around this, add an additional 'packages' argument when submitting the job:

spark-submit --packages com.amazonaws:aws-java-sdk:1.11.920,org.apache.hadoop:hadoop-aws:2.8.4

License

Copyright (c) 2015-2016 EURL Tweag.

All rights reserved.

sparkle is free software, and may be redistributed under the terms specified in the LICENSE file.

Sponsors

sparkle is maintained by Tweag I/O.

Have questions? Need help? Tweet at @tweagio.