Deepinsight can provide a premade configuration with great open source tools for cloud deployment. This enables you to have machine learning in production very, very fast.
The only requirement is having access to deploy Kubernetes clusters. This is easily available on public clouds such as AWS, Google Cloud and Azure, usually done in about 5 minutes. Kubernetes can also be deployed on premise.
The last mile of data science doesn't take forever
A common trend in today’s corporate world is that many machine learning models with great business impact never make it to production. In silos, analysts create the models, the business side evaluates them and IT deploys them. Heavy handovers can take years and can be expensive.
Typical drivers of costs are:
- Heavy handovers on understanding requirements
- Costly fights on prioritization
- Having to translate data science code to “production code”
These problems might appear organizational, but maybe they aren’t?
What if we could find a compromise where data science code is deployable, testable and scalable?
Flexibility matters. Kubernetes and open source tools gives balance between flexibility and access to out-of-the-box options that enables rapid deployment. Packaging code in the right way enables us to push prototypes to production immediately, and show those results in dashboards or APIs.
Data science is more fun when it has impact
In 2019, Deepinsight worked on a very interesting consulting project with significant end-to-end responsibility. Our client had several open analytical problems and a product manager who only believed in end-to-end results.
To solve this, we started to deploy our models to Kubernetes. Results from our algorithms were highly visible, also in early prototypes of our work, so of course testing was important from the very beginning.
Continuous integration and deployment paid off
Some very interesting patterns emerged after a few months.
- With better testing, we deployed changes to production a few times a week.
- Frequent deployments enabled experimental feature to be deployed to Kubernetes as part of the “definition of done”.
- Average time from user story to production dropped from several weeks to around 16-40 working hours.
- With early deployments, our models were up and running for a few weeks before used by the client. This increased our confidence that our models would work as expected.
- Seeing results early gave us faster feedback loops and enabled us to make improvements based on data.
The right technology was however crucial in achieving this tempo. For 2020 we will improve on what we learned from last year. Get in touch if you would like to learn more about how we enable companies deliver better services.