sparkctl 0.5.0

Orchestrates Spark standalone clusters on HPCs.

sparkctl

This package implements configuration and orchestration of Spark clusters with standalone cluster managers. This is useful in environments like HPCs where the infrastructure implemented by cloud providers, such as AWS, is not available. It is particularly helpful when users want to deploy Spark but do not have administrative control of the servers.

Example usage

There are two main ways to use this package:

First, allocate compute nodes. For example, with Slurm (1 compute node for the Spark master and 4 compute nodes for Spark workers):

$ salloc -t 01:00:00 -n4 --partition=shared --mem=30G : -N4 --account=<your-account> --mem=240G

1. Configure a Spark cluster and run Spark jobs with spark-submit or pyspark.

$ sparkctl configure
$ sparkctl start
$ spark-submit --master spark://$(hostname):7077 my-job.py
$ sparkctl stop

2. Run Spark jobs in a Python script using the sparkctl library to manage the cluster.

from sparkctl import ClusterManager, make_default_spark_config

config = make_default_spark_config()
mgr = ClusterManager(config)
with mgr.managed_cluster() as spark:
    df = spark.createDataFrame([(x, x + 1) for x in range(1000)], ["a", "b"])
    df.show()

Refer to the user documentation for a description of features and detailed usage instructions.

Project Status

The package is actively maintained and used at the National Laboratory of the Rockies (NLR). The software is primarily geared toward HPCs that use Slurm. It also supports a generic list of servers as long as the servers have access to a shared filesystem and are accessible via SSH without password login.

It would be straightforward to extend the functionality to support other HPC resource managers. Please submit an issue or idea or discussion if you have interest in this package but need that support.

Contributions are welcome.

Development

This project uses uv for environment management. Install the package with its development dependencies:

$ uv sync --extra dev

Lint, format, and type-check the code with ruff and ty:

$ uv run ruff check .
$ uv run ruff format --check .
$ uv run ty check

These checks also run as Git hooks via prek. Install the hooks once and then run them on demand:

$ uv run prek install
$ uv run prek run --all-files

Run the unit tests. These are fast, require no special resources, and are what CI runs:

$ uv run pytest -m "not integration"

The integration tests download a real Spark and Java distribution into tests/data/ and start a real single-node Spark cluster, so they are slower and require network access and sufficient memory. They are excluded from CI; run them locally with:

$ uv run pytest -m integration

Run the complete suite (unit and integration tests) with uv run pytest.

License

sparkctl is released under a BSD 3-Clause license.

Software Record

This package is developed under NLR Software Record SWR-25-109.