analysis-runner 2.43.2

Analysis runner to help make analysis results reproducible

Analysis runner

This tool helps to make analysis results reproducible, by automating the following aspects:

• Allow quick iteration using an environment that resembles production.
• Only allow access to production datasets through code that has been reviewed.
• Link the output data with the exact program invocation of how the data has been generated.

One of our main workflow pipeline systems at the CPG is Hail Batch. By default, its pipelines are defined by running a Python program locally. This tool instead lets you run the "driver" on Hail Batch itself.

Furthermore, all invocations are logged together with the output data, as well as Airtable and the sample-metadata server.

When using the analysis-runner, the batch jobs are not run under your standard Hail Batch service account user (<USERNAME>-trial). Instead, a separate Hail Batch account is used to run the batch jobs on your behalf. There's a dedicated Batch service account for each dataset (e.g. "tob-wgs", "fewgenomes") and access level ("test", "standard", or "full", as documented in the team docs storage policies), which helps with bucket permission management and billing budgets.

Note that you can use the analysis-runner to start arbitrary jobs, e.g. R scripts. They're just launched in the Hail Batch environment, but you can use any Docker image you like.

The analysis-runner is also integrated with our Cromwell server to run WDL based workflows.

CLI

The analysis-runner CLI can be used to start pipelines based on a GitHub repository, commit, and command to run.

First, make sure that your environment provides Python 3.10 or newer:

> python3 --version
Python 3.10.7

If the installed version is too old, on a Mac you can use brew to update. E.g.:

brew install python@3.10

Then install the analysis-runner Python package using pip:

python3 -m pip install analysis-runner

Run analysis-runner --help to see usage information.

Make sure that you're logged into GCP:

gcloud auth application-default login

If you're in the directory of the project you want to run, you can omit the --commit and --repository parameters, which will use your current git remote and commit HEAD.

For example:

analysis-runner \
    --dataset <dataset> \
    --description <description> \
    --access-level <level> \
    --output-dir <directory-within-bucket> \
    script_to_run.py with arguments

<level> corresponds to an access level as defined in the storage policies.

<directory-within-bucket> does not contain a prefix like gs://cpg-fewgenomes-main/. For example, if you want your results to be stored in gs://cpg-fewgenomes-main/1kg_pca/v2, specify --output-dir 1kg_pca/v2.

If you provide a --repository, you MUST supply a --commit <SHA>, e.g.:

analysis-runner \
    --repository my-approved-repo \
    --commit <commit-sha> \
    --dataset <dataset> \
    --description <description> \
    --access-level <level>
    --output-dir <directory-within-bucket> \
    script_to_run.py with arguments

For more examples (including for running an R script and dataproc), see the examples directory.

GitHub Authentication

If you are submitting an analysis-runner job that needs to clone a private repository owned by populationgenomics on GitHub (eg submitting a script to analysis-runner from a private repository), please make sure that your configuration file contains the following section:

[infrastructure]
git_credentials_secret_name = '<ask_software_team_for_secret_name>'
git_credentials_secret_project = '<ask_software_team_for_secret_project>'

If you are specifying multiple configuration files, please make sure that this section appears in the final right-most config to avoid these settings being overwritten.

Custom Docker images

The default driver image that's used to run scripts comes with Hail and some statistics libraries preinstalled (see the corresponding Hail Dockerfile). It's possible to use any custom Docker image instead, using the --image parameter. Note that any such image needs to contain the critical dependencies as specified in the base image.

For R scripts, we add the R-tidyverse set of packages to the base image, see the corresponding R Dockerfile and the R example for more details.

Helper for Hail Batch

The analysis-runner package has a number of functions that make it easier to run reproducible analysis through Hail Batch.

This is installed in the analysis runner driver image, ie: you can access the analysis_runner module when running scripts through the analysis-runner.

Checking out a git repository at the current commit

from cpg_utils.hail_batch import get_batch
from analysis_runner.git import (
  prepare_git_job,
  get_repo_name_from_current_directory,
  get_git_commit_ref_of_current_repository,
)

b = get_batch(name='do-some-analysis')
j = b.new_job('checkout_repo')
prepare_git_job(
  job=j,
  # you could specify a name here, like 'analysis-runner'
  repo_name=get_repo_name_from_current_directory(),
  # you could specify the specific commit here, eg: '1be7bb44de6182d834d9bbac6036b841f459a11a'
  commit=get_git_commit_ref_of_current_repository(),
)

# Now, the working directory of j is the checkout out repository
j.command('examples/bash/hello.sh')

Running a dataproc script

from cpg_utils.hail_batch import get_batch
from analysis_runner.dataproc import setup_dataproc

b = get_batch(name='do-some-analysis')

# starts up a cluster, and submits a script to the cluster,
# see the definition for more information about how you can configure the cluster
# https://github.com/populationgenomics/analysis-runner/blob/main/analysis_runner/dataproc.py#L80
cluster = dataproc.setup_dataproc(
    b,
    max_age='1h',
    packages=['click', 'selenium'],
    cluster_name='My Cluster with max-age=1h',
)
cluster.add_job('examples/dataproc/query.py', job_name='example')

Development

You can ignore this section if you just want to run the tool.

To set up a development environment for the analysis runner using pip, run the following:

pip install -r requirements-dev.txt

pre-commit install --install-hooks

pip install --editable .

Deployment

1. Add a Hail Batch service account for all supported datasets.
2. Copy the Hail tokens to the Secret Manager.
3. Deploy the server by invoking the deploy_server workflow manually.
4. Deploy the Airtable publisher.
5. Publish the CLI tool and library to PyPI.

The CLI tool is shipped as a pip package. To build a new version, we use bump2version. For example, to increment the patch section of the version tag 1.0.0 and make it 1.0.1, run:

git checkout -b add-new-version
bump2version patch
git push --set-upstream origin add-new-version
# Open pull request
open "https://github.com/populationgenomics/analysis-runner/pull/new/add-new-version"

It's important the pull request name start with "Bump version:" (which should happen by default). Once this is merged into main, a GitHub action workflow will build a new package that will be uploaded to PyPI, and become available to install with pip install.