Skip to main content
When a release goes bad in production, you put the last good release back with an ordinary deploy. Forge has no rollback command and nothing rolls back automatically: you name the earlier release, the plan labels the move BEHIND, and rollback floors refuse any release too old to run on today’s data. This guide recovers a hosted environment from a release that deployed cleanly but broke the page, then sets a floor. Release gates explains why going back is a deploy, not an undo.

Before you begin

1

Find what changed

obs changes lists promotions and failed deploys, newest first:
terminal
Activity rows are trimmed. Here v2.10.0 is live, and v2.9.1 was the last release that served the page.
2

Check the floors admit the earlier release

terminal
floor check only reads. Exit 1 means a floor refuses that release; see Troubleshooting.
3

Redeploy the earlier release

Name the version, and say why with --note, which is recorded on the ledger:
terminal
Nothing is rebuilt: the deploy uses the digests v2.9.1 recorded when it was cut. To review first, use --plan-only, then --approve the digest it prints.
4

Verify the recovery

The page served v2.9.1 again (curl returned <title>guides-sandbox-promote v2.9.1</title>), and the timeline marks the move:
terminal
reliant forge env status prod compares what the platform serves with the release and reports 1 match, 0 drifted, 0 missing.
5

Set a floor so nobody deploys behind this release

When the release you are on changed something older code cannot run on, such as a migration, set an order floor at it:
terminal
A later attempt to deploy v2.8.1 was refused with exit 1 before anything was written:
A migration marked -- forge:breaks-rollback raises a declared floor by itself when its release is promoted, so you need this step only for changes forge cannot see.
6

Roll forward with the fix

The earlier release is a stopgap. Fix the bug, cut a new release, and promote it the usual way, through staging if you have one.

What this does not do

  • No automatic rollback. A rollout that never becomes healthy leaves the previous release serving, and a failed migration stops the deploy before any workload changes. A release that rolls out and then misbehaves, like v2.10.0 here, whose rollout SUCCEEDED, stays live until you deploy something else.
  • No undo. Going back does not reverse the newer release’s migrations or the data it wrote, which is why floors exist.
  • No break-glass for hosted environments. --offline-ledger redeploys from a cached ledger only where forge applies the release from your machine; it is refused for hosted environments. See Releases and promotion.

Troubleshooting

A floor refuses the release you named, and no deploy flag skips it. Roll forward instead, or, if the floor no longer applies, lower or lift it with a reason: reliant forge env floor show prod lists the floors, and Rollback floors covers lower, clear and lift.
Someone else deployed to the environment after your plan was computed. Re-run with --plan-only, read what landed, and decide again.
The bad release is still rolling out. Wait for it, or pass --supersede, which is recorded on the new promotion, to replace it now.

Next steps

Promote to production

Move a release staging runs to prod by digest.

Gate prod on release health

Stop a bad release in staging before prod takes it.

Release gates

obs check, health gates and every floor verb.

How release promotion works

Releases, the ledger, and why there is no rollback command.