databricks-botops-course

Course for doing AgentOps (e.g. ML/LLM/AIOps/BotOps) on Databricks, based on a data mesh monorepo structure.

What will you learn in this course?

• How to build LLM Bots on Databricks
• Move LLM Bots (chat bots, agents etc) from dev to prod
• Using git branches and commits
• We will not do Github Actions here, but the processes and code needed are shown
• Evaluation-driven development
• Testing Bots with chat interfaces
• Structure your environments to allow for dev deployments of Bots

What is a Bot?

A Bot is a an app which uses LLM components, e.g. an agent (if it does autonomous routing/decisions), a chatbot, or an LLM-based endpoint which replies to requests using LLM (Large Language Models=).

Focus

How can we deploy Databricks bots in a way that is:

• (git-)versioned
• usable
• ordered and sustainable
• enabling each developer / data scientist to do deploy
• way of working for exploration, development, staging and production of pipelines

Repository structure

We will do our tasks in the context of the folder representing the tripbot ml flow:

orgs/acme/domains/transport/projects/taxinyc/flows/ml/tripbot/

The structure is a proposal, which might have to be adapted in a real world organization.

The structure is:

• org: acme
  • domain: transport
    • project: taxinyc
      • flowtype: ml (meaning Machine Learning / AI, the alternative is prep, for ETL/data engineering)
        • flow: revenue

The structure will be applied to:

• Agent / model naming
• Data code, i.e. the pyspark code herein git
• The database tables produced by that code
• The data pipelines being deployed

The purpose of this structure is to have sufficient granularity to enable each department/org, team/domain, project and pipeline, to be kept apart.

You can explore the structure here in Databricks, or more easily in the repo with a browser.

Longer explanation of the repo structure

A longer explanation of the ideas behind the repo structure can be found in the article Data Platform Urbanism - Sustainable Plans for your Data Work.

Dataops libs

For the dataops code, we use the brickops package from Pypi, to enable a versioned pipeline deployment and way of working. The main logic is under dataops/deploy.

Reusing the structure and libs

The structure and brickops libs can be used in your own projects, by installing the pypi package brickops, forking the repo and copying the content and adapting it.

Course

1. Go to course/
   • DO NOT run anything under course/00-Workshop-Admin-Prep
2. Got to course/01-Student-Prep/01-General
   • Go through the instructions under that folder
3. Start with the tasks under 02-DeployTasks
   • Some sections are just for reading or running, others you need to solve.