← Back to news

Replacing the Job Router with Zero Downtime on EU

Mothballing the Sorting Hat – how we migrated our dynamic job routing

Mothballing the Sorting Hat

An old hat

usegalaxy.eu runs more than 1 million jobs per month. All these jobs have different demands and use different resources and technologies, like Docker, Apptainer etc., which need all sorts of parameters and environment variables. Galaxy handles jobs with resource managers like SLURM, HTCondor or the remote job execution system Pulsar. But how do these resource managers know which resources each individual job needs? You could specify that for every single tool, but as you can guess, with over 8,000 tools and many different possibilities to compute a job, this approach leads to chaotic and unmaintainable static lists.

In order to determine job execution parameters at runtime and based on factors such as the job inputs, the user submitting the job, cluster status, etc, we needed a different approach Galaxy lets you do that with Dynamic Destination Mapping.

In May of 2018, UseGalaxy.eu started a new project: the Sorting Hat. The name comes from the character in Harry Potter, so chosen because at the time UseGalaxy.eu had multiple independent clusters backing its computation. The hat sorted those jobs into their "houses": the clusters on which they would run. It was initially conceived of as "Job Configuration as a Service" and we thought about making it an API endpoint available for all Galaxies, but we decided to scale back our ambitions.

This python script dynamically modified destinations based on rules that evaluate the specific job characteristics, for example the dataset input size. This allows you to allocate more memory for your job based on the input file size.

In light of J.K. Rowling's controversial personal views and statements, we would certainly have chosen the name differently today[^naming]

Why change a running system?

So far so good – this python script did its job well for some years, but it also came with some limitations. Even though it supported individual rules, defining them was more complicated, not clearly structured and also had limited functionality. The other goal is to save time by sharing a unified job routing system across all Galaxy instances. When every Galaxy instance maintains their own scripts and rules, you also get different results on different instances and admins have to spend a big share of their time communicating and implementing new rules in their custom scripts.

The Total Perspective Vortex

Galaxy Australia came up with a more sophisticated approach that is available as a PyPi package and lets you define tags and rules on all entities, like users, roles, tools and destinations. These tags can be specified as accepted, preferred, required and rejected. Inheritance makes it easy to give tags for example to groups of tools like tools that need to run in Docker. During the matching process all tags from the job-side (user, role, tool) and from the destination-side are evaluated. And because it evaluates really everything it was named Total Perspective Vortex (TPV) after Douglas Adams’ The Hitchhiker's Guide to the Galaxy.

With all that functionality, TPV is not tied to a specific Galaxy instance, but rather is a generalized and flexible approach that fits all needs. With the help of Galaxy Australia's admin Catherine Bromhead, we managed to migrate our Galaxy instance from Sorting Hat to TPV last week. Despite minor scheduling issues that affected mainly Docker tools, we can consider this migration smooth and user friendly.

How can you contribute?

TPV makes use of a shared library that is maintained in GitHub, making it easy for everyone in the community to contribute. This library defines resources and parameters that are valid across all Galaxy instances. You can install TPV to your machine and format, lint and even dry-run the configuration, before deploying in production. Once the tests are passing, the pull request gets accepted and all Galaxy servers that are using the shared library will automatically fetch the updated rules the next time those Galaxies restart.

Small Scale Admins Take Note

The TPV Shared Database is a key feature for small admins and will save you significant amounts of time maintaining a small Galaxy instance. For most tools you might want to install, it will already have an entry for the recommended CPU and Memory requirements. It's just up to you to define the maximum memory and cores that are appropriate for your compute resources, and a small TPV config will take care of the rest! If you want to read more about setting up the TPV and TPV Shared Database, the Galaxy Training Materials.

Footnotes:

[^mothball]: An English expression for taking something out of service

[^naming]: The Harry Potter reference was chosen by a trans sysadmin; she regrets not changing it earlier but it would have involved some migration and risked downtime for a non-user facing component which resulted in a low priority.