Ref: Add "fly app rollback 731" command · Issue #579 · superfly/flyctl · GitHub I imagine this feature won’t be easy for Fly to implement.
The way we do it is, deploy to two Fly apps. One runs last known good deployment but with only one or two VMs (I call this env, lgtm-box), typically a release behind. The other is a proper prod environment, which means, it is also our test environment 
If something goes wrong, I scale up the lgtm-box and flip the DNS entries to point to it, while I figure out just what went wrong when testing in prod…
This means, if you’ve a database layer or any external state, it must at all times work with code running in the lgtm-box.