# Practical MLOps

Noah Gift, Pragmatic AI Labs | MLOps Coffee Sessions | Episode 27 | 58:36
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=GvAyV8m8ICI
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/practical-mlops
Published: 2021-01-26
Tags: deployment, engineering-culture, reliability

## TL;DR
- Noah Gift says machine learning should be treated as one component of a larger product, with user experience, repeatability, data drift, and ethics considered alongside predictive accuracy.
- He teaches students to deploy an existing model in a web or mobile application, create a feedback loop, and set up continuous integration and delivery before tuning the model deeply.
- Noah Gift recommends learning software engineering and cloud practices, using established platforms where possible, and considering the external effects of machine learning systems.

## Summary
Noah Gift argues that practical MLOps starts with treating a model like a component of a software product. Students should first take an existing model, put it into a web or mobile application, create a feedback loop, and set up continuous integration and delivery. He is critical of training that focuses on predictive modeling while leaving deployment and operations to other people. He expects AutoML to automate much of the current model-building work, which will make domain knowledge, solution design, and the ability to create useful products more valuable. Noah recommends cloud certifications and studying how web and mobile developers build reliable systems. He prefers established cloud providers and warns that small or immature tools can create hiring and support problems. The conversation also deals with ethics and externalities. Noah uses recommendation systems, facial recognition, Airbnb, and scooters to argue that teams must consider how systems affect people beyond the immediate product metric.

## Key ideas
### A model is one part of a product, not the whole solution
[03:03](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=183s)
Noah Gift tells his students to think about machine learning like a mobile app, database, or other component. He warns that teams can get trapped in Kaggle scores and predictive value while losing sight of the product they are trying to build. A system can have extremely high accuracy and still be catastrophically bad if the team has not solved user experience, repeatability, data drift, and other operational problems. His practical advice is to give machine learning no special status in the larger system and judge it alongside the other pieces.

### Teams should consider how a model is used as carefully as its accuracy
[05:16](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=316s)
Noah connects machine learning operations to the way a model affects people after deployment. He discusses The Social Dilemma and recommendation engines that directed traffic toward Alex Jones, flat-earth theories, and QAnon. In his view, focusing only on predictive power can produce serious social outcomes. He asks whether customers will like the product and whether it is ethical to build. Later, he applies the same concern to facial recognition trained on biased data and to business models that take shared resources, such as housing or sidewalks, while ignoring the externalities.

### Students should deploy an existing model before studying model internals deeply
[08:14](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=494s)
Noah's teaching approach begins with an already-trained model from Kaggle or a managed service. He mentions downloading a model, using Google AutoML Vision, or dragging data into Apple's Create ML. The first task is to get the model into a web or mobile application and make the feedback loop work. He compares this with marathon training: someone should first find out whether they can run the distance before buying specialized equipment or optimizing their pace. For him, the foundation is a working application with continuous integration and continuous delivery.

### Data scientists need enough DevOps knowledge to work without a surrounding group
[10:28](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=628s)
Noah says a data scientist who lacks foundational DevOps knowledge may need a dozen people around them to get anything done. He calls that a poor way to start a career. His courses therefore treat a model as a file or endpoint and teach students to build and deliver the surrounding application. He is direct about ownership: when someone says deployment is not their job because it is DevOps, he compares that attitude with refusing responsibility for the frame of a house. People building machine learning systems should understand reliability and the software practices already used by other engineering teams.

### Cloud and AutoML are changing which technical skills matter
[12:10](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=730s)
Noah is skeptical that academic tools such as scikit-learn and pandas will fit every company, especially when teams work with data at scale. He mentions AWS SageMaker, Microsoft Azure ML Studio, Google AI Platform, and Spark as examples of platforms that may be more suitable. He also expects more of the current data science workflow to become automated. Hyperparameter tuning and other detailed tasks may increasingly happen through a button or managed service. He advises people to focus on the problem and the solution rather than on techniques that may soon be automated.

### Established cloud platforms reduce operational and hiring risk
[19:17](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=1157s)
When asked to compare providers, Noah calls AWS the safest place to start, Microsoft Azure his second choice, and Google a more niche option. He points to AWS's broad coverage, regular product improvements, documentation, and clearer pricing. He likes Azure ML Studio and some Microsoft products. He also likes individual Google offerings such as BigQuery and App Engine, though he does not expect Google to be the sole platform for most companies. His broader point is that a cloud provider is more predictable than betting the company on one young MLOps startup.

### Certifications and studying other engineering teams can build architecture skills
[28:19](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=1699s)
Noah recommends that people with master's degrees or PhDs still consider certifications such as AWS Solutions Architect. Preparing for the exam exposes concepts including elasticity and high availability. He also suggests building ordinary applications, such as a Flask app, APIs, or an AWS Lambda service. Engineers can learn by watching web and mobile developers work for six months to a year, then applying their build and delivery practices to machine learning. A production mobile application is built automatically rather than by one person on an occasional basis, and machine learning systems should follow the same standard.

### Team character matters more than individual brilliance
[35:54](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=2154s)
Noah says the worst way to think about a software team is to focus on individual talent. His worst hires were people who believed they were smarter than everyone else and did not need to follow the rules. He prefers people with basic competence who are willing to learn other parts of the job and help the team. That can include machine learning, web development, mobile development, and routine work. He connects this hiring approach to product quality: credentials and elitism can create a toxic organization, while teamwork and character make it easier to build and maintain the system.

### MLOps may become ordinary software work as machine learning is commoditized
[53:33](https://www.youtube.com/watch?v=GvAyV8m8ICI&t=3213s)
Noah is not convinced that machine learning will remain a separate discipline from DevOps. He expects classification and similar tasks to become highly automated, with tools reporting measures such as area under the curve and false positives or negatives. He says the valuable skills will be deep knowledge of a particular dataset, the ability to use high-level automation, and the ability to build a working solution that creates revenue for a customer. Choosing a model or improving its accuracy may matter less than understanding the domain and delivering something useful.

## Notable quotes
- Noah Gift: "If you can't solve the other 10 pieces like user experience, repeatability, data drift, all those kind of problems, then you're really building something that even if it does have extremely high accuracy could be catastrophically bad." (06:54)
- Noah Gift: "You need to get continuous integration and continuous delivery set up and you really don't necessarily at first need to know anything other than it's a machine learning model is a file or it's an endpoint." (09:48)
- Noah Gift: "The worst possible way to think about a software team is to focus on individual talent." (35:54)
- Noah Gift: "The externalities and ethics do matter even if you're very, very technical." (43:47)
- Noah Gift: "Those three things, revenue, creating solutions, and then also having domain expertise in the data, those are more important than what model you're picking or how accurate your model is." (56:11)

## Tools & references mentioned
- Pragmatic AI Labs
- The Social Dilemma
- Kaggle
- Google AutoML Vision
- Create ML
- scikit-learn
- pandas
- AWS SageMaker
- Microsoft Azure ML Studio
- Google AI Platform
- Spark
- AWS
- Microsoft Azure
- Google Cloud
- Google App Engine
- BigQuery
- Google Data Studio
- AWS Lambda
- AWS Step Functions
- GitHub
- GitHub Actions
- Microsoft Teams
- LinkedIn
- Splunk
- Erlang
- Go
- Julia
- AutoML
- Udacity
- O'Reilly
- Alfredo Deza
- Paul Graham
- Airbnb
- Tesla
- S&P 500
- Alex Jones
- QAnon

## Who should watch
- You are a data scientist who can build models but depends on other engineers to deploy or operate them.
- You are choosing a cloud platform or an MLOps startup and need to think about support, documentation, hiring, and long-term risk.
- You are designing machine learning training or hiring a team and want practical guidance on software delivery, teamwork, and externalities.

## Editor's note

Noah Gift argues that students should deploy an existing model in an application, build its feedback loop, and set up continuous integration and delivery before studying model internals deeply. ZenML turns Python steps into pipelines whose runs record inputs, outputs, and code versions, so the workflow can be traced and repeated. Unchanged steps are cached instead of recomputed.

Written by the MLOps Talks editors (the ZenML team), not by the speaker.

## Related talks

- [Practical MLOps Part 2](https://mlopstalks.com/talks/practical-mlops-part-2) (Alfredo Deza, Author and Speaker, 1:01:38)
- [The Godfather Of MLOps](https://mlopstalks.com/talks/the-godfather-of-mlops) (D. Sculley, Google, 51:25)
- [MLOps Insights](https://mlopstalks.com/talks/mlops-insights) (David Aponte-Demetrios Brinkmann-Vishnu Rachakonda, 37:47)
- [Doing MLOps](https://mlopstalks.com/talks/doing-mlops) (Noah Gift, Pragmatic AI Labs, 1:01:22)
- [MLOps and DevOps, Parallels and Deviations](https://mlopstalks.com/talks/mlops-and-devops-parallels-and-deviations) (Damian Brady, Microsoft, 55:32)