Managing ML operations: when does it make sense to build and when to buy?
When to build your own MLOps platform and when to buy it?
Here’s the framework:
This depends on the level of maturity and size of the organization.
For large enterprises that have no internal resources whatsoever to build, the only option is to buy.
But for most enterprises, the typical answer here is to do both.
Throughout the MLOps lifecycle, some areas are unique to your business.
This is typically around:
- data pipelines
- deployment of a mobile or web application
- batch prediction.
For these use cases, it better to build the components internally.
However, it often doesn’t make sense to build in areas that are more generalizable.
- experiment management,
- model management, and
- model monitoring.
For example, Uber has an amazing MLOps platform called Michaelangelo.
They came to us as they recognized that their engineering cycles were better spent in other places.
Comet is integrated harmoniously with their Michelangelo platform.
This way, their internal customers can get value from day 1, and they can spend their resources in other critical areas.