OpenSAFELY job runner

A job runner is a service that encapsulates:

• the task of checking out an OpenSAFELY study repo;
• executing actions defined in its project.yaml configuration file when requested via a jobs queue; and
• storing its results in a particular locations.

End users will find more information in the OpenSAFELY documentation.

Operating principles

In production, this software runs as a loop on a secure server within the infrastructure of the primary data provider. It polls an OpenSAFELY job server, looking for requests to run jobs.

Jobs belong to a workspace. This describes the git repo containing the OpenSAFELY-compliant project under execution; the git branch, and kind of database to use. The workspace also acts as a kind of namespace for partitioning outputs of its jobs.

An OpenSAFELY-compliant repo must provide a project.yaml file which describes how a requested job should be converted into a command (& arguments) that can be run in a subprocess on the secure server. It incorporates the idea of dependencies, so an action that generates a chart might depend on an action that extracts data from the database for that chart. See the Actions reference for more information.

An action can define outputs; these are persisted on disk and made available to subsequent actions in the workspace, and users who have permission to log into the server and view the raw files.

The runner takes care of executing dependencies in order. By default, it skips re-running a dependency whose previous run produced output that still exists in the production environment. The runner also reports status back to the job server, redacting possibly-sensitive information.

The runner is bundled as part of the opensafely-cli tool so users can test their actions locally.

Job structure

The job server serves jobs as JSON in the following format. First, a job must belong to a workspace:

{
    "workspace": {
        "name": "my workspace",
        "repo": "https://github.com/opensafely/job-integration-tests",
        "branch": "master",
        "db": "full"
    }
}

Possible values for "db" are "full", "slice", and "dummy".

A workspace is a way of associating jobs related to a given combination of branch, repository and database. To enqueue a job, a client POSTs JSON like this:

{
    "backend": "tpp",
    "action_id": "do_thing",
    "workspace_id": 1
}

Consuming jobs

A job runner is service installed on a machine that has access to a given backend. It receives jobs from the server and consumes those whose backend value matches the value of the current BACKEND environment variable.

It must also define three environment variables which are an RFC1838 connection URL; these correspond to the db requested in the job's workspace definition, and as such are named FULL_DATABASE_URL, SLICE_DATABASE_URL, and DUMMY_DATABASE_URL.

When a job is found, the following happens:

• The corresponding repo is fetched. Private repos are accessed using the PRIVATE_REPO_ACCESS_TOKEN supplied in the environment.
• Its project.yaml is parsed:
  • Individual actions are extracted from this file
  • A dependency graph is calculated for the requested action; for example, an action might depend on three previous actions before it can be run
  • Each action in the graph is checked to see if it needs to be run
    • Actions that either: (a) already have output generated from a previous run; (b) are currently running; (c) failed on their last run do not need to be run
  • If a dependency has failed, then the requested action fails
  • If the dependency needs to be run, a new job is pushed to the queue, and the current job is postponed
  • If an action has no dependencies needing to be run, then its docker run is executed
  • On completion, a status code and message are reported back to the job server. On success, a list of output file locations are also posted. On failure, the message has any potentially-sensitive information redacted, and is associated with a unique string so that a user with requisite permissions can log into the production environment and examine the docker logs for the full error.

Output locations

Every action defines a list of outputs which are persisted to a permanent storage location. The project author must categorise these outputs as either highly_sensitive or moderately_sensitive. Any pseudonymised data which may be highly disclosive (e.g. without low number redaction) should be classed as highly_sensitive; data which the author believes could be released following review should be classed as moderately_sensitive. This design allows tiered levels of permissions for collaborators to review data outputs. For example, the study author would usually have access to highly_sensitive material for debugging; but other collaborators could have access to moderately_sensitive data to prepare it for release (for which it is planned to add a minimally_sensitive category).

Outputs are therefore persisted to filesystem paths according to the following environment variables:

# A location where cohort CSVs (one row per patient) should be
# stored. This folder must exist.
HIGH_PRIVACY_STORAGE_BASE=/home/opensafely/high_security

# A location where script outputs (some for publication) should be
# stored
MEDIUM_PRIVACY_STORAGE_BASE=/tmp/outputs/medium_security

Project.yaml

A valid project file looks like this:

version: "3.0"

expectations:
  population_size: 1000

actions:

  generate_study_population:
    run: cohortextractor:latest generate_cohort --study-definition study_definition
    outputs:
      highly_sensitive:
        cohort: output/input.csv

  run_model:
    run: stata-mp:latest analysis/model.do
    needs: [generate_study_population]
    outputs:
      moderately_sensitive:
        model: models/cox-model.txt
        figure: figures/survival-plot.png

See the project pipeline documentation for a detailed description of the project.yaml setup.

Local actions development

The cohortextractor command-line tool imports this library, and implements the action-parsing-and-running functionality as a series of synchronous docker commands, rather than asynchronously via the job queue.

For developers

Please see the additional information.