Rollback

A rollback is the act of reverting a running system back to a previous, known-good version — most commonly used to quickly undo a deployment that turned out to cause errors, performance problems, or incorrect behavior in production. Rollbacks are a core piece of incident response: when something goes visibly wrong shortly after a deploy, the fastest and safest fix is very often "undo the change that just went out," rather than trying to hot-fix forward under pressure while an incident is actively ongoing. Why it matters for AI/SaaS builders: the ability to roll back quickly and confidently is what makes frequent deployment safe — teams that deploy multiple times a day rely on a fast rollback path as their safety net, since even a well-tested change can behave unexpectedly under real production traffic and data in ways staging never revealed. A slow or risky rollback process (one that itself might fail or take 30 minutes) undermines the entire premise of shipping fast, since every deploy effectively becomes higher-stakes. How it works: the specific mechanics depend on the deployment strategy. With immutable container deployments, a rollback is often as simple as re-pointing the load balancer or orchestrator (Kubernetes, ECS) at the previous container image tag, which is already built and tested — no new build required, just a traffic switch, often completing in seconds to a couple of minutes. With database migrations involved, rollbacks are trickier, since reverting application code while leaving a schema change in place (or vice versa) can itself cause errors — well-designed migrations are written to be backward-compatible for at least one release, specifically so a code rollback doesn't require an immediate, risky database rollback too. Worked example: a team deploys a new version of their API at 2pm. By 2:15pm, their observability dashboard shows the error rate on the `/api/checkout` endpoint has jumped from a baseline 0.1% to 8%. Rather than trying to diagnose and hot-fix the root cause under pressure with real customers actively failing to check out, the on-call engineer runs a rollback command that re-points the production load balancer at the previous, known-good container image (which was still running in the cluster, just receiving no traffic). Within 90 seconds, the error rate drops back to baseline; the team then calmly investigates the root cause offline, with no ongoing customer impact, before attempting the fix and a fresh deploy later.