The last twenty releases are still running, so rollback is a routing change

Rollback on most stacks means redeploying an older artifact: pull the image, start it, wait for health checks, hope the migration was reversible. Keeping the old version warm was the obvious answer, and the per-instance bill was the obvious reason nobody did.

On Hoody each release is a container copied from the last one, copy-on-write, holding only the blocks that changed. Release 47 serves the alias; releases 28 through 46 sit behind it, idle, burning no CPU, costing the bare metal box nothing measurable. When 47 misbehaves at 2am, you delete the alias and recreate it on 46. Two requests. 46 was never stopped; it was just not being asked.

Twenty warm releases on per-instance pricing is twenty bills. On a flat-rate box, it's a deploy history you can route to.

Rollback stopped being a redeploy and became a pointer.