[jrevels]

Cofounder here :) Happy to answer any questions about scaling Julia on K8s, data architecture for science-first teams, and our weird/wild journey to get to this point.

We are, of course, hiring - check out https://beacon.bio/careers!

[ssivark]

Congratulations! Some basic questions, if you’ll indulge: How exactly does Julia fit into your software architecture? Is your product a cloud offering and/or does it have a client side application? Do you use Julia for only the algorithmic parts, or much more than that?

Also, what do you even mean by data architecture for science-first teams?

Feel free to just point me to other content :-)

[jrevels]

Good questions!

> How exactly does Julia fit into your software architecture?

In a variety of ways:

- We have a bunch of external/internal Julia packages; Julia's package manager is really great at facilitating the development of "tooling ecosystems" comprised of lightweight libraries that compose well together. For example, we use Legolas.jl [1] in conjunction with a well-curated Arrow-in-S3 lake to help teams define lightweight, self-serviceable schemas for Arrow tables in a manner that integrates well with the wider Tables.jl ecosystem [2], interactive analysis workflows, and our own ETL/ELT-ish patterns.

- Julia powers some interesting services within Beacon's Platform. For example, one of our Julia services provides dynamic streaming DSP (multiplexing, filtering, statistics) for biosignal data, atop which we build other applications/pipelines for both product development and internal analysis work.

- We use Julia for exploratory distributed computing on K8s [3], which is awesome because Julia has a lot of potential in the distributed computing landscape (IMO [4]).

> Is your product a cloud offering and/or does it have a client side application?

We work with our clients to do neurobiomarker discovery, clinical trial design, deploy our analysis pipelines into clinical trials, and a few other interesting things :) One of the critical differentiators of Beacon is that we can precisely target and harness key EEG features to a degree that isn't possible without the kind of algorithms/tools we've developed.

> what do you even mean by data architecture for science-first teams

I want to do a blog post on this at some point, but a core value for us - across all of our processes, tooling, and data interactions - is self-serviceability and composability. IMO, the two are inextricably linked. Our goal is to empower each Beaconeer to perform analyses in an afternoon atop terabytes of data that would take them months in a lab atop gigabytes of data.

To achieve this, we treat large-scale data curation/manipulation as an activity that we're all empowered to participate in and contribute to, as opposed to an environment where separate data engineering teams have to administrate siloed systems. Tools like K8s/Julia/Arrow are key enablers here, by surfacing capabilities to domain experts that let them to iterate fast without needing to "throw problems over the wall" to other teams/systems.

It's not a perfect match, and it's a bit abstract, but I remember reading this post about "data meshes" [5] a while back and thinking "Hey, that's similar to what we're chasing after!"

[1] https://github.com/beacon-biosignals/Legolas.jl

[2] https://github.com/JuliaData/Tables.jl

[3] https://github.com/beacon-biosignals/K8sClusterManagers.jl

[4] https://news.ycombinator.com/item?id=24842084

[5] https://martinfowler.com/articles/data-mesh-principles.html