Pipelines as jobs

Problem

Currently we have tasks, jobs, and pipelines. While this corresponds to a common pattern for building bioinfomatics analysis, in practice we are finding that this design is overly rigid with several unintended consequences:

  1. arv-run-pipeline-instance currently gets special treatment. Pipelines instances are run by crunch-dispatch system automatically using the resources of the dispatcher node, and creating pipelines from pipeline templates is the only way to run jobs using workbench. However, there are other third party pipeline frameworks that our users are currently using, such as bcbio-nextgen, rmake, snakemake, Nextflow; rewriting these to use arv-run-pipeline-instance is time consuming (due to the need to analyze the original pipeline) or infeasible (because they make user of features that don't map cleanly onto existing Arvados features).
  2. Need to work with batches and pipelines of pipelines. If we have a pipeline that processes a single sample, and we want to run it 100 times, we need to create 100 pipelines by hand or by a script.
  3. Currently, we can create jobs which either 1. submits stages as subtasks or 2. submits stages as additional jobs.
    1. In the first approach, job reuse features are not available with tasks, and all subtasks must be able to run out of the same docker image. There is also (by design) reduced visibility into the inner working of tasks as compared to jobs. (But see Tasks as jobs)
    2. In the second approach, the controller job currently ties up a whole node, even though it is mostly idle. Additionally (and unlike tasks and pipelines) we do not track which job submissions were made by which other jobs, so there's a loss of provenance information.

Proposed solution

  1. Improve job scheduling so that we can have more than one job on a node, with jobs can be allocated to a single core (possibly even fractions of a core).
  2. Run arv-run-pipeline-instance as a job in a container just like everything else.
  3. Run pipeline stages as task-like jobs. See Tasks as jobs
  4. ?? Workbench summary pages such as dashboard just display jobs that were submitted by a user (spawned_by_job_uuid is null). Unify the display of pipelines and jobs so that a pipeline is just a job that creates other jobs, and permit drilling down through the process tree.
  5. Improve SDK APIs to make it easy to spawn a a bunch of jobs as "futures" and then wait for them all to finish.

Another benefit: supports the proposed v2 Python SDK by enabling users to orchestrate pipelines where "python program.py" is the same whether it runs locally, runs locally and submits jobs, or runs as a crunch job itself and submits jobs.

Related ideas

Currently, porting tools like bcbio or rmake still requires the tool be modified so that it schedules jobs on the cluster instead of running locally. We could use LD_PRELOAD to intercept a whitelist of exec() calls and redirect them to a script that causes the tool to run on the cluster.