Your files

/kaggle/working behaves like a disk that is always there.

Kaggle saves it. When a “Save & Run All” session ends, Kaggle keeps /kaggle/working as that version’s output — and it does so however the session ends. A clean exit, a crash and a cancel were all measured to save every file up to the moment the container stopped. Older versions stay readable (v1, v2, …), and they belong to the notebook, not to an account.

kdev puts it back. Kaggle starts every batch session with an empty /kaggle/working; its “Persistence” setting only covers the web editor. So each kdev up works out which saved version you left off at, and the new box downloads it before you connect.

how the last session ended	what the next kdev up gives you
kdev down	everything
kdev down, then kdev up straight away	everything: up waits for the save first (about 2 min per GB)
Stop session in the Kaggle UI	everything (a cancel still saves)
the session’s time ran out	everything (the box stops itself 2 min early)
the tunnel process died	everything
the box crashed	everything, up to the crash
your laptop or network dropped	the box is still running: kdev up connects to it
Kaggle’s machine itself failed	not testable; a git remote covers it

What kdev adds on top of Kaggle

Kaggle’s saved output keeps every regular file — thousands of them, dotfiles, .git, large binaries. It drops symlinks and empty directories, and loses executable bits, which would break a Python virtualenv. The box records those in .kdev/meta.json every 30 seconds, and the restore puts them back.

Picking the version to restore

A version with no files is a source-only save (group setup, a Quick Save), so kdev walks past it. A session that was stopped before its restore finished — or that you worked in before restoring — is not skipped: it gets stacked on the versions it was meant to be built on, so nothing done in it is lost. The layers underneath fill in only what that session’s restore never delivered (the box logs each file as it lands, in .kdev/fetched), so a file you deleted there stays deleted.

A restore never overwrites a file written in the current session, so it is safe to run again at any time:

kdev restore               # finish or repeat the restore on the running box
kdev restore --from v12    # add files v12 has that the box doesn't (not a rollback)
kdev workspace files       # what `kdev up` would restore (--version v12 for another)
kdev backup                # optional: an exact copy of the box on this machine

Your own cells in the notebook

kdev’s cell sits on top, and nothing below it runs. When a session ends, later cells are blanked for that run, so a training cell can’t start after kdev down and use quota. Your cells stay in the notebook, untouched.

Files made in the Kaggle web editor

The editor’s interactive session is the one place kdev can’t reach: its files only reach the API when a version is saved. Kaggle describes Quick Save as saving the notebook “as displayed”, which isn’t a clear promise to save the working directory, so check with kdev workspace files. Everything made through kdev is covered.