Docs

Files & compute

A workbook keeps project state separate from a running machine, so you can keep the useful record of work without leaving CPU or GPU compute running.

What persists after a stop#

Saved workbook files, code, execution history, and retained results remain available after compute stops. Each execution keeps its own record of inputs, sources, outputs, logs, status, and initiator information where available, as described in Workbooks. You can inspect and download saved results without starting compute.

What does not persist#

Stopping a runtime releases its active processes and memory, including GPU memory. Consoles does not restore a running server, an in-memory model, installed system state, or an open terminal session. Use setup commands and saved workspace files to reproduce the environment on the next start.

The CPU/Linux badge describes the selected capability. It does not promise a specific vendor image, operating-system distribution, or package set.

Storage limits#

For hosted workbook storage, the current limits are 256 MiB per file, 2 GiB per workbook, and 1,000 files, including retained run archives.

Some compute providers have smaller snapshot or restore limits. Consoles surfaces a retention problem with the run instead of treating an incomplete save as a successful retained result, so check the run record when an expected output is missing.

Keep inputs and outputs explicit#

Upload a file or save it into the workbook before an execution when it matters to the result. Save generated outputs into workbook storage or as declared artifacts. A file reference in a document or session is not a stored asset.

You can also explicitly import a public HTTPS file into a new workbook path. Import saves the bytes and the source provenance; mentioning a URL in a document does not import it. Resumable uploads are not available yet, so complete an upload or import before depending on that input.

Work with files from instant TypeScript#

An instant TypeScript cell can use await files.list(), await files.read(path), and await files.write(path, content) to work with persistent UTF-8 workbook files. Paths must be relative to the workbook, and each text read or write is limited to 1 MiB. The isolate cannot use the network, a shell, npm, or the operating-system filesystem.

Start, run, stop#

  1. Create a workbook and choose or attach a suitable runtime.
  2. Review its setup and file inputs.
  3. Explicitly start CPU or GPU compute.
  4. Run a code cell or recipe control, then inspect its status, logs, and saved outputs.
  5. Stop compute when you are finished. The workbook and its retained files remain available.

Starting a CPU or GPU runtime can incur charges. Creating and configuring a workbook alone does not start paid compute.

Apps that run in a workbook#

Some workloads run a web application inside a compute job. Start the server first, keep its job running, then add or open its authenticated app link from the workbook. Opening an app link does not install software or start its server.

  • Workbooks for what a workbook and each run retain.
  • Recipes for prepared starting points on CPU and GPU.
  • Troubleshooting for missing results, failed uploads, and app links.