Skip to content

WorkFlow Launcher's Role in Terra⚓︎

Summary⚓︎

The Data Sciences Platform (DSP) is building a new system (around Terra) for storing and processing biological data. The system design includes a Data Repository where data is stored, and a Methods Repository that executably describes transformations on that data.

In the new system, the DSP needs something to fulfill the role that Zamboni currently plays in DSP's current infrastructure to support the Genomics Platform (GP).

Zamboni watches various queues for messages describing new data and how to process it. Zamboni interprets those messages to dispatch work to workflow engines (running on the premises or in the cloud) and monitors the progress of those workflows. The Zamboni web UI allows users to track the progress of workflows, and enables Ops engineers to debug problems and resume or restart failed workflows. Zamboni can run workflows on both a local Sun Grid Engine (SGE), and on Cromwell on premises and in the cloud.

We think that WFL can fill the role of Zamboni in the new data storage and processing system that DSP is developing now.

History⚓︎

WFL began as a project to replace a Zamboni starter, with the old name "Zero". A starter is a Zamboni component that brokers messages among the queues that Zamboni watches. It interprets messages queued from a Laboratory Information Management System (LIMS), such as the Mercury web service, and demultiplexes them to other Zamboni queues.

Zero was later adapted to manage the reprocessing of the first batch of UK Biobank exomes. It has since been adapted to drive workflows for other projects at the Broad.

Zero is unusual in that it usually runs as a stateless command line program without special system privilege, and interfaces with services running both on premises and in Google Cloud. It also manages a processing workload as a set of inputs mapped to outputs instead of tracking the progress of individual sample workflows.

A Zero user need only specify a source of inputs, a workflow to run, an execution environment, and an output location. Then each time it is invoked, Zero ensures that workflows are started and retried as needed until an output exists for every input.

Zero has recently been adapted again to deploy as a web service under Google App Engine (GAE) though most of the value of Zero is still not available to the server. And now it has the new name WFL.

The role of WFL in Terra⚓︎

Diagrams of the new DSP processing system show a WFL service subscribed to event streams from the Data Repository (DR), with interfaces to both the Data and the Method Repositories.

The implication is that something notifies WFL of new data in the Data Repository and WFL determines how to process it somehow. WFL then looks up whatever is required from the Method Repository, calls on other services as necessary to process the data and writes the results back to the DR. There is also, presumably, a web UI to track and debug the workflows managed by WFL.

Many details are yet to be worked out.

WFL Concepts⚓︎

WFL is designed around several novel concepts.

  1. Manage workloads instead of workflows.

    This is the biggest difference between Zamboni and Zero (WFL). Zamboni manages workflows whereas WFL manages workloads.

    Zamboni's unit of work is the workflow. Zamboni manages each workflow separately.

    A workflow is a transformation specified in WDL or Scala code that succeeds or fails to produce a result. The input to a workflow and its result may consist of multiple files, but they represent a single unit of work managed by a workflow engine such as Cromwell. Zamboni prepares a new workflow for each message it receives by packaging up the input and submitting it to a workflow engine. It then monitors that workflow and reports on its success or failure.

    WFL manages a workload, which indirectly comprises multiple workflows. Each workflow maps an input to some output, but WFL generally tracks only the inputs and outputs instead of the workflows themselves.

    Think of a workload as a set of inputs transformed via a workflow engine into a set of outputs. Call that set of outputs the result set. WFL generally does not care whether any individual workflow succeeds or fails. It merely considers all possible inputs specified by the workload, and looks for inputs whose outputs are missing from the result set. If some input lacks an output in the result set, WFL starts a new workflow to process that input.

    Note: This characterization is unfair to Zamboni. Zamboni also had to manage multiple workflows before the advent of Cromwell and still does when running workflows on SGE. But WFL can take advantage of Cromwell's job management to simplify its implementation.

  2. Specify inputs and outputs by general predicates.

    Each Zamboni message explicitly specifies an input to be processed. Zamboni then starts a workflow for that input and reports its status. Zamboni reports failure so a user can debug and manually succeed, reconsider, or restart the workflow. The output of a successful workflow is not Zamboni's concern.

    WFL finds inputs by applying a predicate specified by the user subject to some run-time constraint. Then WFL applies a function to each input to find how that input maps to the result set. Another predicate applied to the input, and its output in the result set, determines whether WFL will launch a workflow on that input.

    Those predicates and function can be anything expressed in a programming language. The run-time constraint is some strings passed on the command line.

  3. Minimize user input and decisions at run time.

    WFL gathers the predicates and mapping functions described above into a module that also knows how to generate everything a workflow engine needs to launch the workflow to process an input into a result output.

    That module name is one of a few run-time constraints specified by strings in a web form or on a command line. Further constraints are usually one or two of the following:

    - a processing environment (`dev` `prod` `pharma5`),
    - a file system directory (`/seq/tng/tbl/`
    `file://home/tbl/`),
    - a cloud object prefix (`gs://bucket/folder/`
    `s3://bucket/`),
    - a pathname suffix (`.cram` `.vcf`),
    - a spreadsheet (or JSON, TSV, CSV, XML) filename
    - or a count to limit the scope of a predicate.
    

    The module interprets the other constraints, determines which processing environments are allowed, and parses any files named accordingly.

  4. Maintain provenance.

    WFL runs out of a single JARfile built entirely from sources pulled from Git repositories. WFL records the Git commit hashes in the JARfile and adds them to every Cromwell workflow it starts. WFL can also preserve the Cromwell metadata alongside any result files generated by the workflow.

  5. Run with minimal privilege.

    Zamboni runs as a service with system account credentials such as picard.

    WFL is designed to run as whoever invokes it, such as tbl@broadinstitute.org. WFL fetches the users credentials from the environment when invoked from the command line. WFL requires authentication when running as a server, and constructs a JSON Web Token (JWT) to authorize other services as needed.

  6. Limit dependencies.

    WFL depends on a Java runtime, and Gnu make and the Clojure CLI to manage dependencies. Of course, it also pulls in numerous Clojure and Java libraries at build time, and sources WDL files from the warp repository. A programmer need only install Clojure, clone the wfl Git repository, and run make to bootstrap WFL from source.

WFL server⚓︎

The WFL client is a command-line batch program that a person runs intermittently on a laptop or virtual machine (VM).

We are working to port the client functions of WFL to a ubiquitous web service (WFL server) running in Google Cloud. That port requires we solve several problems.

  1. State

    The WFL client is a stateless program that relies on consistent command line arguments to provide the constraints needed to drive the input discovery predicates and so on. Each user runs a separate process that lasts only as long as necessary to complete some stage of a workload.

    The WFL server is shared among all its users and runs continually. Therefore it requires some kind of data store (a database) to maintain the state of each workload across successive connections from web browsers.

    We intend to use the hosted Postgres service available to GAE applications for this. This work is already underway (GH-573).

  2. Authorization

    The WFL client assumes it runs in an authenticated context. It can pull credentials from the environment on every invocation that requires authorization to a service.

    The WFL server will also need to authorize services to run as some authenticated user, but cannot assume the credentials are always available, nor that there is a user present to provide them. WFL can already use OAuth2.0 to authenticate users against an identity provider and use the resulting credentials to build a JWT. It can also derive the bearer token required by most of our authorized services from a JWT.

    But WFL also needs some secure JWT store, so tokens are available to authorize services even when there is no active user connection. It also needs some mechanism to refresh tokens as they expire to support long-running workloads.

  3. Workload specification

    The user of a WFL service needs some way to specify a workload. A workload may be some set of inputs and the kind of workflow to run on them.

    A WFL client user now specifies a workload with a module name and a constraint. For example, ukb pharma5 110000 gs://broad-ukb/in/ gs://broad-ukb/out/ means find up to 110000 cloud objects with names prefixed with gs://broad-ukb/in/, process them in the Cromwell set up for pharma5, and store their outputs under gs://broad-ukb/out/ somewhere.

    The ukb module knows how to find .aligned.cram files under the gs://broad-ukb/in/ cloud path and set up the WDL and Cromwell dependencies and options necessary to reprocess them into .cram output files. The ukb module also knows how to find the Cromwell deployed to support pharma5 workloads, how to authorize the user to that Cromwell, and how to read any supporting data from other services. And finally, the ukb module knows how to determine which inputs do not yet have outputs under the gs://broad-ukb/out/ cloud path, and do not have workflows running in the pharma5 Cromwell.

    In an ideal design, this workload specification would integrate conveniently with the Data Repository's subscription or eventing service. In any case though, WFL needs some interface through which a user can specify what needs to be done.

  4. Workload management

    Workloads need to be started, stopped, and monitored somehow. This implies that there is some way to find active or suspended workloads, and affordances for acting on them.

    Users need some way to monitor the progress of a workload, and to find and debug workloads encountering unacceptable workflow failures.

    Monitoring and diagnostic code already exists in various WFL modules, but there is no easy way to use them from a web browser.

  5. Service interface

    WFL should be useful to programs other than web browsers. It is easy to imagine Terra users wanting to query WFL for the status of workloads directly without buggy and tedious screen scraping.

    WFL should at least export a query endpoint for use by other reporting services as well as its own browser interface. It would be nice to provide a familiar JSON or GraphQL query syntax to other services.

  6. Browser interface

    A browser interface should require little in addition to WFL's service interface. Ideally, one should be able to adapt WFL to new workloads via a browser interface without requiring a redeployment.

Back to top