[Part 1]
🤷🏻♀️ Centralized vs Distributed MLOps?
🗣️ At times this "debate" can feel like a needless, waste-of-time where the answer is going to be "it depends" (similar to build or buy).
😬 Except for the implications on funding, organizational structure, and even success (depending onthe maturity of the MLOps initiatives).
🧑🤝🧑 One the one hand, "MLOps-as-a-Service" doesn't have to be centralized to a single team.
[Part 2]
🕸️ In organizations that are particularly large and complex, you'll have multiple teams contribute to MLOps systems and practices.
🤼♀️ And when not well-orchestrated, there are a number of reasons why these attempts to try to centralize on a single team (and org) fail:
✔️ Political infighting as teams attempt to compete in land-grabs over the tooling & infrasturcture.
✔️ Legacy systems (& culture) that kill momentum and time to delivery of new components.
[Part 3]
✔️ The deep, relationship building between different teams, against the backdrop of competing priorities, often requires the short-term sacrifice of big "flashy" wins.
🧠 On the other hand, centralization (at least at the beginning of building your company's ML Platform) can help ensure:
✅ Multiple teams aren't doing the same work & reinventing the wheel;
✅ Alignment, especially with regards to communication;
✅ Faster & responsive pivots when new information & requirements come in.
[Part 5]
💭 If you're reading this & you've been involved with developing an ML Platform, how did you approach the "centralize vs distributed" discussion? What worked? What failed?
👇 Let me know in the comments below!
#mlops #mlplatform #ai #strategy #dataops #dataengineering #devops #platformengineering #platformdesign #mlengineer
Mikiko