Here are 3 of the most common challenges tech executives face when creating their MLOps strategy:

1. Picking the right tools:

There’s a lot of noise in the market around different categories (probably around 20 only within MLOps), vendors, and solutions.

Thus, choosing the right tools for your stack becomes difficult.

2. Internal eagerness:

There’s often internal pressure to build everything.

This is a challenging problem space, and your talented engineers are likely excited to try and build their own solutions.

3. Lack of full visibility:

It’s difficult to plan for the future because it’s likely that when designing this strategy, there isn’t enough understanding of the problems you’ll face in production.

We've worked with organizations at various stages of their MLOps adoption— anything from Uber, who invented the concept of an MLOps platform with Michelangelo, to organizations that are much newer in this space.

Here are some suggestions we give to the companies we meet:

1. You do not need to build the entire stack end-to-end from day one.

It's okay to start investing where there’s the most pain at the moment. That typically begins around experimentation management and orchestration.

2. It's critical to work with vendors and tools that allow you to customize their behavior so that when you're ready to bring in another solution, you can easily integrate it.

3. It's important to avoid the hype and ask your data scientists what their pain points are, because your DS team and their pains might not be the same as others.

4. Even though there's a lot of noise and you don't have a full vision of your canonical stack, here are some common tools to look into:
- For orchestration and serving, you can use one of the cloud providers.
- For experimentation management, you can use Comet.
- For deployment, you can use the same CI/CD tools as used for software.
- For monitoring, you can get by with things like Data Dog and New Relic.
- For higher-level metrics, you can use a BI solution like Looker.

5. When the models start affecting the bottom line, you can invest in a dedicated monitoring solution for machine learning.

While you’re at it, pick a monitoring solution that works well with your experimentation because these 2 components of an MLOps stack are closely related, and keeping them separate will lead to additional pain points down the road

Let me know if you found this helpful!

About Comet: Comet provides an MLOps platform that enables data scientists and teams to track, compare, explain and optimize experiments and models.

Learn more about Comet: [ Ссылка ]

Comet Office Hours are powered by "The Artists of Data Science", the only self-development podcast for data scientists, featuring interviews and perspective from some of the leading voices in the industry.

Listen and subscribe to The Artists of Data Science: [ Ссылка ]

#machinelearning #mlops #datascience #linkedin

свернуть