Set your GPS to DevOps
Tuesday March 12 I attended the Amsterdam DevOps meetup at TomTom in Amsterdam, which had a focus on AI and Kubernetes.
Martijn Siereveld, Director Software Engineering at TomTom, opened saying that TomTom long moved on from putting navigation devices in “your dad’s car”, focusing on maps, autonomous driving, and automotive and enterprise customers. TomTom powers Microsoft cloud offerings (Azure) with location-based services.
A DevOps approach to AI
Thiago de Faria (@thiagoavadore) is the Head of Solutions Engineer for full-service IT provider LINKIT.
A report from MMC, a London-based venture capital firm, recently made the claim that 40% of European firms that are classified as an “AI startup” don’t actually exploit the field of study in any material way for their business:
Thiago’s mission is to make us all feel more confident in supporting AI and Machine Learning applications. Those that are in fact real.
You can’t catch bugs in your machine learning algo with console.log(). You can’t “patch” ML models. Instead, Thiago suggests to borrow best practices from DevOps.
Instead of
1. Ops object to Dev using untested frameworks
2. Dev adds technical debt to circumvent Ops
3. Ops object a bunch more
… Thiago says Product, Dev, IT Ops, QA, InfoSec, Data Science should all be holding hands and collaborating. That means undergoing a similar culture shift we went through with DevOps:
- Data Scientists want to be “on-call” even less than you and me do. Thiago thinks they should be on the rotation, and this should be a top-down decision.
- Acknowledge we’re dealing with ethics issues (and biased models)
- … and explainability problems. Shouldn’t you be able to explain why someone is denied a loan, even or especially when an algorithm made that decision?
- Creating a safe env where people dare to share their models.
- Creating proxy metrics, because it’s hard to measure efficiency otherwise — you can hardly ask users if your model correctly predicted hotdog / not hotdog, every time.
“Local Data Science is not a thing, just like works-on-my-machine died with containers.” Beyond CI/CD for ML, Thiago advocates for “continuous evaluation”. Bias, drift & fragility are ML’s bugs, says Thiago. Bias, like Amazon’s sexist hiring algo. Drift, as in ML models need to be (re)calibrated, (re)trained. And then there’s the fragility of the input data. And the latter two only surface in Prod.
Thiago demo-ed mlflow (by DataBricks) and its use for evaluating, packaging, deployment and model serving, and hopes that Uber will open source their machine learning-as-a-service system Michelangelo.
Agnostic CI/CD: the end of the line for DevOps?
Ruben Visser is a full-time “IT guy” since 1995. He recently started n-ableconsulting.nl and offers Kubernetes / cloud-native consulting.
CI/CD should cover your application from cradle to grave, says Ruben, be automated, and should be environment neutral (hence, “Agnostic CI/CD” ). Ruben showed how with Gitlab & Kubernetes the whole DevOps cycle can be automated for Microservices projects. He walked us through his namespaces (and dev/prod parity), Git and branching, Helm charts, and other personal tweaks.
Ruben ultimately ends up with a setup that is indeed agnostic-ish. It still utilizes GitLab and Kubernetes, but admittedly doesn’t depend on it.
Announcements and discount codes
Use “devopsmeetup” (or “dutchazuremeetup” if they limited it) to get 25% off of your ServerlessDays Amsterdam March 29 ticket. With “devopsgogoto” you shave 10% off of your GOTO Amsterdam entree fee.
And DevOps Days Amsterdam is still looking for speakers (go to their CfP), and Early Bird tickets are still — but limitly — available.