Stop vs destroy
This is the most important distinction in the product, and the one that most often produces an unexpected bill. Read it once, properly.
The short version
| Stop | Destroy | |
|---|---|---|
| Container | shut down | removed |
| Your disk and its contents | kept | deleted, permanently |
| The GPU slot | held for you (unless your balance runs out) | released to the marketplace |
| GPU charges | stop | stop |
| Storage charges | continue | stop |
| Can you resume it? | yes, in place | no |
Stopping is a pause that still costs money. Destroying is final and costs nothing further.
Why stopping still bills you
When you stop an instance, your disk stays on that physical machine, and that machine cannot give the space away to anyone else. You are charged for that reservation at the storage rate — quoted per GB per hour — for as long as the instance exists in a stopped state.
The GPU itself is no longer charged; while the instance is stopped you pay storage only. An instance you stopped yourself keeps its slice: those GPUs do not go back on the marketplace (unless your balance runs out, below), and there is no queue to hold a place in.
This means a large disk you forgot about is a slow, quiet leak. A 500 GB volume left stopped for a month costs real money even though nothing ran.
If you are finished with the work, destroy it. Stop is for "I'll be back this afternoon", not "I'm done".
Why the GPU slot matters
When you stop an instance yourself, it holds your slot on that specific machine. When you start again, you land back on the same hardware with the same disk — no re-download, no re-setup, and no risk that the offer was rented out from under you while you were away.
Destroying releases the slot immediately. The offer goes back on the marketplace, and someone else may take it within seconds. There is no way to reclaim it.
An empty balance is different. Where the machine supports it, the platform then releases the GPUs back to the marketplace, both those of the instances it auto-stops and those of instances you had stopped yourself, while each disk stays on its machine. A later start has to take back the same GPUs on the same machine: if they are free, the instance starts as usual; if they are in use by someone else or out of service on that machine, the start fails, and the instance stays stopped with its disk intact. Nothing can reach that disk until a start succeeds, so try again later, and destroy the instance only if you no longer need the data: destroying deletes the disk. See After the GPUs are released.
Which should you use?
Stop when all of these are true:
- You'll return within hours or a day or two.
- There's state on the disk you'd rather not rebuild — a dataset, a conda environment, checkpoints.
- The disk is small enough that the storage charge is cheaper than the time it would take to rebuild.
Destroy when any of these are true:
- The work is finished.
- Everything valuable is already pushed somewhere else — object storage, a git remote, a model registry.
- The disk is large, and re-downloading is cheaper than parking it.
Before you destroy
Destruction deletes the disk. There is no snapshot, no trash, and no undo. Copy anything you care about off the instance first:
# from your own machine — port, user and host are the ones in the
# connect command on the instance's Connect tab
scp -P <port> -r <user>@<host>:/workspace/checkpoints ./checkpoints
Or push it to wherever it belongs — object storage such as Alibaba Cloud OSS, a git remote. Services outside mainland China can be slow or unreachable from the machine.
What the system does on its own
Two automatic actions apply the same distinction, to stop a forgotten instance from draining an account to nothing:
- When your balance reaches ¥0, running instances are stopped. Your data survives; storage charges continue against a balance that is already empty. Where the machine supports it, their GPUs are released, and so are the GPUs of instances you had stopped yourself (see above).
- If the balance then stays at or below ¥0 for 24 hours and reaches −¥35 or lower, the stopped instances are destroyed. At that point the data is gone.
Top up before then. See How billing works.