The sandbox you don't kill between turns

Your product gives every user a code-running assistant. The user asks, the assistant writes a script, runs it, shows the result. Then the user goes to lunch.

On platforms that meter the sandbox-minute, lunch is the expensive part. So you tear the sandbox down after each turn, serialize what state you can, and rebuild it when the user returns. Dependencies reinstall. The dataframe reloads. The user notices.

On Hoody the user's sandbox is a container on a flat-rate server. It stays. Between turns it burns no CPU and shares its disk blocks with every container cloned from the same base. The packages are still installed. The 2GB CSV is still on disk.

You delete the teardown-and-rehydrate code. The sandbox doesn't end when the turn ends. Neither does the bill begin.