Abhibhaw Asthana

Work / Devtron · AI assistant

Atlas: from hours to a minute

When a deployment failed, users read raw logs for hours or opened a support ticket. I shipped Atlas, an AI assistant that finds the root cause and suggests a fix.

was hours now ~1 minto find the root cause
was hours now minutesto recover
User
Developers and on-call engineers whose deployment just failed on Kubernetes.
Problem
A failed deployment showed a status, not a reason. Users read raw logs and events for hours, or opened a support ticket and waited.
Decision
An AI assistant inside the product. It reads the failure events, logs and service context, then explains the cause in plain words and suggests a fix.
Not built
Automatic fixes in the first version. Atlas explains and suggests; a person decides. Automated, human-approved runbooks came later. We also did not train our own model: teams bring the LLM they already use.
Result
Root-cause time fell from hours to about a minute. Time to recovery fell from hours to minutes.

How it works

Kubernetes says that something failed. It rarely says why.

Before
  1. Deploy fails
  2. user digs
  3. removed: Events, logs, configsread by hand, hours
  4. or gives up
  5. Support ticket
After
  1. Deploy fails
  2. user clicks Explain
  3. Atlasreads events, logs and service context
  4. about a minute
  5. Cause and fixa person decides
Atlas starts where the failure is shown, with the context already attached.
Application details page with a Degraded status, a failed deployment and an Explain button

The entry point is the failure itself. Public image from Devtron docs.

AI Analysis panel explaining that a pod failed due to DeadlineExceeded, with container details and possible causes

The answer names the cause and likely fixes. Users rate each answer. Public image from Devtron docs.

What I left out

  • Automatic fixes. Version one explained and suggested. Approved runbooks with an audit log came later.
  • A model of our own. Teams bring the LLM provider they already approved.

What happened

  • Root-cause time fell from hours to about a minute.
  • Recovery time fell from hours to minutes.