Practical MLOps

Probably you should ask Noah Gift about it 😃 But you can read it through OReilly Learning, the early release version is already available there

Ah ok I didn’t realise that you could read the book before it was published!

Same in Germany.
And I think my trial with O’Reilly learning expired. So I’ll have to wait for the book release!

Me too. I can’t wait till October to by this book fromNoah Gift. His previous two books (Python for DevOps and Pragmatic AI) are great and have been a huge help for me. Also, his co author Alfredo Deza has such an inspiring life story.

Yes, you can read online in rough draft form on the O’Reilly website: https://learning.oreilly.com/library/view/practical-mlops/9781098103002/

It also should be in kindle form in around 30 days or so and in print soon after.

A good place to start is by using the tools on the platform you are already on. All major cloud platforms have an MLOps solution and this is a great place to start. AWS Sagemaker, GCP Vertex AI, and Azure ML Studio

1. I do think the process of MLOps should be applied to all projects because it is an extension of DevOps. All software projects should have CI/CD and you can even do this with notebooks: https://github.com/noahgift/myrepo
   
2. For feature stores they have raw materials in a form easily consumed by a ML pipeline. I.E. Containers package the runtime with the code, Feature stores package the raw ingredients for ML into a metadata system. A database is too low level by itself to be a feature store.

There is a strong overlap between Data Engineer and MLOps with perhaps as little as a 5% overlap. The key 5% is that a MLOps practitioner also knows a bit about ML and can train models, diagnose their output and knows about ML Platforms like AWS Sagemaker, MLflow, etc.

I would start with whatever platform is available and use their offerings: i.e. Google, AWS, Azure. Let’s take AWS for example, if you have gigantic data and gigantic teams, say over 250 people in your company then a “big” platform like Sagemaker probably makes sense because of how much it offers.
If you use AWS but have a 3 person team, Sagemaker may or may not be the best easy win. Perhaps AWS App Runner with open source MLOps tools might be a better fit.

Which open source tools can you recommend?

Autonomous driving is a good example. I went to Tesla AI Day last week and 90% of the people I spoke with did MLOps, i.e. tools/infra around computer vision.

Thanks

People with a strong DevOps/Infrastructure skill set can easily make the transition to MLOps. They just need to pick up a bit of ML training. One way to do this is to read the book I wrote and also to get AWS ML Certification certified (or similar). Note, I helped create the AWS ML certificaiton….

Thanks, Noah!

I think MLOps is essentially an evolved DevOps but with the addition of ML.

You can read in any order. Since both Python for DevOps and Cloud Computing are start with either then move on to Practical MLOps. They all have a similar theme with more depth on cloud, devops or mlops depending on the book

How to decide which tools to choose?
whatever is simple to get started with an improves automation and quality.

Should one choose for an open source alternative or choose a tool by a cloud service provider?
I personally prefer to pay a vendor, so I would start with a cloud offering.

[10:03 AM] There are a plethora of tools coming out, how do you make a framework on choosing which tool to choose and how to choose?
If you are on a cloud platform start with what they offer and go from there.

[10:04 AM] How to practically navigate through the MLOps cycle? Some nuggets of wisdom like MLOps isn’t a tech problem but a people problem etc
Make sure you have CI/CD working and iterate from there.

[10:04 AM] Do small startups really need MLOps or is it over engineering?
MLOps is a behavior/methodology that focuses on Kaizen (continuous improvement). So it applies to anything small or big.
A. Automate everything
B. Make it better quality daily

Looks like a 🐕, probably dalmation

We don’t have control of the animals.

To start with I would make sure you have CI/CD, i.e. the foundation of modern software engineering. This is the first step.

I don’t believe there are any disadvantages to MLOps. In a nutshell it just means “Kazien”, i.e. continuous improvement. Make everything better and more automated.

I would start with traditional monitoring/instrumentation for you platform using whatever tools are already in place. Then add additional business logic for ML.
Additionally if you use Cloud Platforms they have default monitoring like for example Azure ML Studio which does model versioning and experiment versioning.

“What is most important MLOps checklist for business critical model serve pipeline?”
Start with CI/CD, if you don’t have this you cannot do MLOps

“Do you believe current set of lowcode/nocode MLOps solutions are good enough to be used for mission critical usecase?”

Yes, in many cases you don’t need to write code. A good example is Azure ML Studio AutoML.

I think it may help to have someone who has some form of MLOps certification. One good example of this is course I just created on Coursera: https://www.coursera.org/specializations/building-cloud-computing-solutions-at-scale

We cover AWS/Azure/GCP very heavily

Thanks for the response ! 🙂

Hi Maja! Thanks for the super kind words. This is a great question! I think that there are a few things from seeing the opposites of the core pillars of operations (DevOps/MLOps in general) like automation, monitoring, testing, and CI/CD. For example: no (or little) automation, doing things manually, no pipelines, no monitoring.
Aside from those, you have other red-flags like over-engineering. Fast, iterative processes are far better than waiting 3 months to design the perfect thing

There is always room for improvement. I keep hearing people say “what if everything is already automated?” - well… there is always stuff to automate and improve. You are asking a critical question here, and not asking critical questions (see critical thinking section at the beginning of the book) is a tremendous problem.

I will read it as soon as I get the book. Thank you Alfredo Deza so much for your guidance!

Automation is not a one time thing that takes months to achieve and is super expensive. Noah Gift taught me the right path years ago: pick any one thing you do manually and automate it by the end of the week. Rinse and repeat, and suddenly a few months later you have several things automated. It is now CHEAPER to run operations because of it and the team can concentrate in even better automation

Always automate

Leveraging the cloud for automation (CI/CD or pipelines doesn’t matter) is good. Leveraging anything that is already solved that is not a core competency of your business is crucial

This is a difficult question to get a straight answer. I don’t think there is a one-size-fits-all problem solver here. If you have biased data, but you have automation, tests, pipelines, etc… you still have a biased model in the end. MLOps can’t solve biased data. There is always the human element in all of this, and critical thinking (see critical thinking section at the beginning of the book) is essential

Alfredo Deza Thank you for your advice.. Can’t wait to read your book and thanks for all your valuable time!

Reusability is the gold standard. Not entirely sure how to abstract data structures, but sharing/reusing artifacts sounds great to me. As to how to do this, well it depends! Perhaps an S3 bucket would suffice if everything is behind AWS. If you need external access, it sounds like an HTTP API is the way to go

There is no silver bullet here where everything can be fully automated. You’ve mentioned one of the caveats which is unpredictable behavior. Human interaction+evaluation has to be possible. Pipelines have to be flexible. Any automation/workflow has to easily allow for changes and updates. When automating, you must think about the pitfalls and how to address them. For example, you have a pipeline that normalizes data in small amounts, what can you do today that will allow batching the normalizing if the data is gigantic?

alfredinsky

if you are starting out then I would pick automation. Anything you can do to start automation is going to be super useful and empowering

The book uses a public Github repository that you can use to see examples GitHub - paiml/practical-mlops-book: [Book-2021] Practical MLOps O’Reilly Book

cookbook in particular is a good recipe

thank you guys 🙂

We don’t have anything related to Kedro (sorry, not sure what that is)

thank you