Getting Started

You’ll need:

Running Hail locally with a pre-compiled distribution

Hail uploads distributions to Google Storage as part of our continuous integration suite. You can download a pre-built distribution from the below links. Make sure you download the distribution that matches your Spark version!

Unzip the distribution after you download it. Next, edit and copy the below bash commands to set up the Hail environment variables. You may want to add the export lines to the appropriate dot-file (we recommend ~/.profile) so that you don’t need to rerun these commands in each new session.

Un-tar the Spark distribution.

tar xvf <path to spark.tgz>

Here, fill in the path to the un-tarred Spark package.

export SPARK_HOME=<path to spark>

Unzip the Hail distribution.

unzip <path to hail.zip>

Here, fill in the path to the unzipped Hail distribution.

export HAIL_HOME=<path to hail>
export PATH=$PATH:$HAIL_HOME/bin/

To install Python dependencies, create a conda environment for Hail:

conda env create -n hail -f $HAIL_HOME/python/hail/environment.yml
source activate hail

Once you’ve set up Hail, we recommend that you run the Python tutorials to get an overview of Hail functionality and learn about the powerful query language. To try Hail out, run the below commands to start a Jupyter Notebook server in the tutorials directory.

cd $HAIL_HOME/tutorials
jhail

You can now click on the “01-genome-wide-association-study” notebook to get started!

In the future, if you want to run:

  • Hail in Python use hail
  • Hail in IPython use ihail
  • Hail in a Jupyter Notebook use jhail

Hail will not import correctly from a normal Python interpreter, a normal IPython interpreter, nor a normal Jupyter Notebook.

Running on a Spark cluster

Hail can run on any cluster that has Spark 2 installed. The Hail team publishes ready-to-use JARs for Google Cloud Dataproc, see Running in the cloud. For Cloudera specific instructions see Running on a Cloudera Cluster.

For all other Spark clusters, you will need to build Hail from the source code.

Hail should be built on the master node of the Spark cluster with the following command, replacing 2.2.0 with the version of Spark available on your cluster:

./gradlew -Dspark.version=2.2.0 shadowJar archiveZip

Python and IPython need a few environment variables to correctly find Spark and the Hail jar. We recommend you set these environment variables in the relevant profile file for your shell (e.g. ~/.bash_profile).

export SPARK_HOME=/path/to/spark-2.2.0/
export HAIL_HOME=/path/to/hail/
export PYTHONPATH="${PYTHONPATH:+$PYTHONPATH:}$HAIL_HOME/build/distributions/hail-python.zip"
export PYTHONPATH="$PYTHONPATH:$SPARK_HOME/python"
export PYTHONPATH="$PYTHONPATH:$SPARK_HOME/python/lib/py4j-*-src.zip"
## PYSPARK_SUBMIT_ARGS is used by ipython and jupyter
export PYSPARK_SUBMIT_ARGS="\
  --conf spark.driver.extraClassPath=\"$HAIL_HOME/build/libs/hail-all-spark.jar\" \
  --conf spark.executor.extraClassPath=./hail-all-spark.jar \
  pyspark-shell"

If the previous environment variables are set correctly, an IPython shell which can run Hail backed by the cluster can be started with the following command:

ipython

When using ipython, you can import hail and start interacting directly

>>> import hail as hl
>>> mt = hl.balding_nichols_model(3, 100, 100)
>>> mt.aggregate_entries(hl.agg.mean(mt.GT.n_alt_alleles()))

You can also interact with hail via a pyspark session, but you will need to pass the configuration from PYSPARK_SUBMIT_ARGS directly as well as adding extra configuration parameters specific to running Hail through pyspark:

pyspark \
  --conf spark.driver.extraClassPath=$HAIL_HOME/build/libs/hail-all-spark.jar \
  --conf spark.executor.extraClassPath=./hail-all-spark.jar \
  --conf spark.sql.files.openCostInBytes=1099511627776 \
  --conf spark.sql.files.maxPartitionBytes=1099511627776 \
  --conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
  --conf spark.kryo.registrator=is.hail.kryo.HailKryoRegistrator

Moreover, unlike in ipython, pyspark provides a Spark Context via the global variable sc. For Hail to interact properly with the Spark cluster, you must tell hail about this special Spark Context

>>> import hail as hl
>>> hl.init(sc) 

After this initialization step, you can interact as you would in ipython

>>> mt = hl.balding_nichols_model(3, 100, 100)
>>> mt.aggregate_entries(hl.agg.mean(mt.GT.n_alt_alleles()))

It is also possible to run Hail non-interactively, by passing a Python script to spark-submit. Again, you will need to explicitly pass several configuration parameters to spark-submit

spark-submit \
  --jars "$HAIL_HOME/build/libs/hail-all-spark.jar" \
  --py-files "$HAIL_HOME/build/distributions/hail-python.zip" \
  --conf spark.driver.extraClassPath="$HAIL_HOME/build/libs/hail-all-spark.jar" \
  --conf spark.executor.extraClassPath=./hail-all-spark.jar \
  --conf spark.sql.files.openCostInBytes=1099511627776 \
  --conf spark.sql.files.maxPartitionBytes=1099511627776 \
  --conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
  --conf spark.kryo.registrator=is.hail.kryo.HailKryoRegistrator \
  your-hail-python-script-here.py

Running on a Cloudera Cluster

These instructions explain how to install Spark 2 on a Cloudera cluster. You should work on a gateway node on the cluster that has the Hadoop and Spark packages installed on it.

Once Spark is installed, building and running Hail on a Cloudera cluster is exactly the same as above, except:

  • On a Cloudera cluster, when building a Hail JAR, you must specify a Cloudera version of Spark. The following example builds a Hail JAR for Cloudera’s 2.2.0 version of Spark:

    ./gradlew shadowJar -Dspark.version=2.2.0.cloudera
    
  • On a Cloudera cluster, SPARK_HOME should be set as: SPARK_HOME=/opt/cloudera/parcels/SPARK2/lib/spark2,

  • On Cloudera, you can create an interactive Python shell using pyspark:

    pyspark --jars build/libs/hail-all-spark.jar \
            --py-files build/distributions/hail-python.zip \
            --conf spark.driver.extraClassPath="build/libs/hail-all-spark.jar" \
            --conf spark.executor.extraClassPath=./hail-all-spark.jar \
            --conf spark.sql.files.openCostInBytes=1099511627776 \
            --conf spark.sql.files.maxPartitionBytes=1099511627776 \
            --conf spark.serializer=org.apache.spark.serializer.KryoSerializer \
            --conf spark.kryo.registrator=is.hail.kryo.HailKryoRegistrator \
            --conf spark.hadoop.parquet.block.size=1099511627776
    

Running in the cloud

Google and Amazon offer optimized Spark performance and exceptional scalability to tens of thousands of cores without the overhead of installing and managing an on-prem cluster.

Hail publishes pre-built JARs for Google Cloud Platform’s Dataproc Spark clusters. We recommend running Hail on GCP via an interactive Jupyter notebook, as described in Liam’s forum post. If you prefer to submit your own JARs or python files rather than use a Jupyter notebook, see Laurent’s forum post.

Building with other versions of Spark 2

Hail should work with other versions of Spark 2. To build against a different version, such as Spark 2.3.0, modify the above instructions as follows:

  • Set the Spark version in the gradle command

    ./gradlew -Dspark.version=2.3.0 shadowJar
    
  • SPARK_HOME should point to an installation of the desired version of Spark, such as spark-2.3.0-bin-hadoop2.7

  • The version of the Py4J ZIP file in the hail alias must match the version in $SPARK_HOME/python/lib in your version of Spark.

BLAS and LAPACK

Hail uses BLAS and LAPACK optimized linear algebra libraries. These should load automatically on recent versions of Mac OS X and Google Dataproc. On Linux, these must be explicitly installed; on Ubuntu 14.04, run

apt-get install libatlas-base-dev

If natives are not found, hail.log will contain the warnings

Failed to load implementation from: com.github.fommil.netlib.NativeSystemLAPACK
Failed to load implementation from: com.github.fommil.netlib.NativeSystemBLAS

See netlib-java for more information.