* feat(nsjail): optional disk-backed /tmp via instance setting
* test(nsjail): unit-test tmp mount resolver and narrow visibility
* refactor(nsjail): switch tmp backing to select + conditional UI
* ui(nsjail): make tmpfs the visible default in /tmp backing select
* fix(nsjail): refuse preexisting jail_tmp to block symlink escape
* fix(nsjail): allow jail_tmp reuse on sequential nsjail calls
Codex flagged that python/ruby/rust executors invoke nsjail twice per
job_dir (install then run). The previous resolver treated any preexisting
jail_tmp as hostile and silently fell back to tmpfs on the second call,
so disk-backed mode never reached the main script run for those langs.
Use symlink_metadata().is_dir() to distinguish a real directory left by
an earlier call in the same job_dir (safe to reuse) from a symlink or
other entity (still refused, as the codebase-tar escape requires).
Also loosen the frontend visibility predicate: only hide nsjail settings
when job_isolation is explicitly 'none' or 'unshare', so deployments
that enable nsjail via DISABLE_NSJAIL=false with no DB setting can
still see the controls.
* feat(nsjail): make tmpfs size configurable via instance setting
Adds a new `nsjail_tmpfs_size_mb` instance setting that overrides the
size of the `/tmp` tmpfs mount inside the nsjail sandbox across all
languages. When unset, the existing per-language defaults (500MB or
800MB) continue to apply, so no behavior change for existing
deployments.
The setting is exposed under Settings → Jobs and is read at job
execution time, so changes take effect on the next job without a
restart.
Fixes WIN-1963
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(nsjail): unify default tmpfs size to 800MB
Previously each executor passed its own per-language default (500MB or
800MB) to resolve_nsjail_tmpfs_size. Unify on a single
DEFAULT_NSJAIL_TMPFS_SIZE_BYTES constant (800MB) so the placeholder
behavior is consistent across languages.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(nsjail): resolve tmpfs size outside ruby download closure
The download.ruby config render runs inside a sync closure passed to
par_install_language_dependencies_seq, so `.await` on
resolve_nsjail_tmpfs_size() was a compile error under the `ruby`
feature. Resolve the size once before the closure and capture the
string instead.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(nsjail): rename resolver to *_bytes and clarify fallback
Addresses CI review feedback:
- Rename `resolve_nsjail_tmpfs_size` to `resolve_nsjail_tmpfs_size_bytes`
so the returned unit is unambiguous at the call site (cubic P2).
- Fix the `NSJAIL_TMPFS_SIZE_MB` doc comment that still said "per-language
default" — there is no per-language fallback anymore, all unset
values resolve to the unified 800MB `DEFAULT_NSJAIL_TMPFS_SIZE_BYTES`
(codex/pi P2).
- Expand the resolver doc to call out that `Some(0)` and negative values
also fall back, since the match arm is `Some(mb) if mb > 0`.
No behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: bind /etc resolver files read-only in nsjail sandboxes
* docs(nsjail): explain why per-file /etc resolver binds are load-bearing
The explicit /etc/hosts, /etc/resolv.conf and /etc/hostname binds look
like removable duplication of the read-only /etc bind above them. They
are not: on Kubernetes those files are separate kubelet bind-mounts on
top of /etc and nsjail's read-only remount is non-recursive, so without
these shadow binds they stay writable and a job can persist cross-tenant
DNS poisoning for the pod lifetime. Comment guards against a future
"dedup cleanup" silently reintroducing the vulnerability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(nsjail): shorten the load-bearing-bind comment to 3 lines
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix(python): fix uv can't find ssl certificates
- Add `PY_NATIVE_CERT` flag, forces UV to use native tls
- Rename `PIP_INDEX_CERT` to `PY_INDEX_CERT`
- Rename `PIP_TRUSTED_HOST` to `PY_TRUSTED_HOST`
For backwards compatibility PIP* variables are still accessible
* feat(python): add `custom_wheels` directory to PYTHONPATH
Add global directory by path `<CACHE_DIR>/python_xyz/custom_wheels`
For example for scripts running python 3.11, in every execution
`<CACHE_DIR>/python_311/custom_wheels` will be accessible and all wheels placed there could be imported and used.
This is usefull for preinstalling wheels before runtime
* Make it work with Nsjail
* Rework and make custom_wheels optional
* Remove `create_dir_all` from imports
* Use sync version of metadata
* Rename `custom_wheels` to `global-site-packages`
* feat: Handle `pip install` by `uv`
Dirty and untested, but already something working
* Integrate with NSJAIL and prepare fallbacks
* Refactor fallback
no_uv disable compile and install
where no_uv_install and no_uv_compile are a bit more specific
* Remove `--disable-pip-version-check`
Reason:
warning: pip's `--disable-pip-version-check` has no effect
* Fix backend compilation error
* Pip fallback overwrite UV's cache
* Initially refactor cache (No S3)
* Support S3
* Remove unused import
* Handle flags for NSJAIL
* Return deleted flag
* Remove verbose mode and enable link-mode=copy
* Granural migration of lockfiles
Before i realized we dont need it :)
* Initial draft (not-working)
* Add fallback
* Fix bug preventing uv from installing deps
'\n' - Love it
* Add verbosity indicator
* Iterate on feature
- Added instance python version
- Rework logic
* Fix EE build error
error[E0599]: no method named `iter` found for tuple `(PyVersion, std::vec::Vec<std::string::String>)` in the current scope
* Support S3
* Support NSJAIL
* Refactor `get_python`
* Make NSJAIL work [Unsafe]
config file missed /proc mount causing install phase to fail
* Trigger CI
* Clean up
* Make Actions build it
* Trigger CI #2
* Update Dockerfile and clean up
* Change fallbacks
now there is only no_uv and NOUV
* Expose INSTANCE_PYTHON_VERSION through env variable
* Change namings
* Include py-version to requirements.in
Also add comments and make code much cleaner
* Use const for python installation dir
It was hardcoded before
* Pin preinstalled version
* Update python_executor.rs
* Up to date branch
* Create PYCACHE dirs
TODO: PY_TAR_DIRS
* Fix after merge
* Make it safer
* Implement USE_SYSTEM_PYTHON
* Implement latest_stable option
* Load INSTANCE_PYTHON_VERSION on startup
* Check for multiple annotations used
* Fix Latest Stable button not pressed if selected
* Proper error handling for conflict on multiple annotations
* Fix merge conflicts
* Preinstall 3.11 and Latest Stable
* Preinstall latest stable in non-blocking manner
* Fix Warning
* Gate preinstall logic behind "python" feature
* Handle raw_deps properly
* Make it work with nsjail
* Revert docker-image.yml
* Revert Dockerfile
* Cleanup + Fixing
* Add windows support
---------
Co-authored-by: Ruben Fiszel <ruben@windmill.dev>
Ruben Fiszel
hugocasa
Pyra
Ruben Fiszel
Kai Jellinghaus