KINTO Tech Blog

Building a culture of MLOps by holding a SageMaker Study Session (4/4)

Ikki Ikazaki
Ikki Ikazaki
Cover Image for Building a culture of MLOps by holding a SageMaker Study Session (4/4)

By Ikki Ikazaki, MLOps/Data Engineer at Analysis Group

This is the last part in a multi-part series on how KINTO Technologies Corporation(KTC) developed a system and culture of Machine Learning Operations(MLOps). Please have a look at Part1(How We Define MLOps in KTC), Part2(Training and Prediction as Batch Pattern), and Part3(Metadata Management or Experiment Tracking Using SageMaker Experiments).


The previous post (Part1) discussed the goal and scope of MLOps and I tried to gain understanding from my colleagues. However, just talking about the concept is not enough for engineers and ML practitioners. People establish their skills once they understand both the idea and how. For software engineers, it seemed to be easy to understand the concept and techniques of the B part of Reliable System Integration while for data scientists A part of Project Management With Speed. This difference in understanding and interests makes it difficult to have better communication and thus a bridge over the gap between them was needed.



The bridge is an opportunity to let them understand each other's interests and methodologies by making an effort together. Conducting a study session on the common ML platform, SageMaker, was expected to be such an opportunity. The goals of study session was set up as follows:

  • To build a good relationship between software engineers and data scientists.
  • To form a common understanding on MLOps and SageMaker.


I didn't make a presentation but just focused on organizing the study session by planning how we proceed. The basic flow was as follows.

  1. Define a scope of contents.
  2. Assigned who takes charge of the contents. Give the material to be studied.
  3. Prepare for the session by creating slides and notebook in two weeks.
  4. Hold the study session.
  5. Back to 2) and continue until all the topics have been discussed.

First of all, I outlined the study contents. SageMaker is a huge platform and we don't need to catch up on everything about it. The outlined contents are the below seven.

  • The basics of SageMaker and SageMaker Studio
  • SageMaker's Training Job
  • Batch prediction and Endpoints
  • SageMaker Processing
  • SageMaker Experiments
  • SageMaker Pipelines
  • SageMaker Projects (CI/CD)

Then, I proposed to colleagues at both my team (Analysis Group) and another team (Platform Group) about the study session. There were four presenters who come together — one is a data scientist and the other three are infrastructure or DevOps Engineers — and I assigned them the above contents. Because of the difference in interests of each role, I took care of who will be in charge of the contents so that all attendees understand the contents without a lot of preparation. Especially, some contents are targeted at engineers such as SageMaker Processing and Projects while some contents are for data scientists such as SageMaker's Training Job and Experiments.

About the study contents, since I've read through the entire developer's guide on SageMaker before and this guide is really well-written, I let the presenter read as well by specifying a scope and executing the relevant notebooks. I believe this way of preparation saves a lot of time for presenters to summarize the information and ensures the quality of the sessions because it is already offered by AWS. To be honest, I just wanted to let them know how well the developer guide is prepared.



All seven sessions were held bi-weekly from November to January. I hope this study session contributed to building common knowledge about SageMaker and understanding about each other's roles. Through the sessions, it was good that one of the DevOps engineers understood how to deploy SageMaker Pipelines and actually he incorporated it into the internal CI/CD program which is based on GitHub Actions. I hope we can talk about how we developed the CI/CD pipeline of SageMaker with GitHub Actions someday.
The study session was a really good way to build a shared culture for both software engineering and data science.

How about the four-part series of blog posts on how KINTO Technologies Corporation developed the system and culture of MLOps? Follow our Tech Blog for future posts to stay up to date.