Python sandbox runners for executing code in isolation aka snekbox.
Supports a memory read/write file system within the sandbox, allowing text or binary files to be sent and returned.
A client sends Python code to a snekbox, the snekbox executes the code, and finally the results of the execution are returned to the client.
%%{init: { 'sequence': {'mirrorActors': false, 'messageFontWeight': 300, 'actorFontFamily': '-apple-system,BlinkMacSystemFont,Segoe UI,Helvetica,Arial,sans-serif' } } }%%
sequenceDiagram
actor Client
participant Snekbox
participant NsJail
participant Python as Python Subprocess
Client ->>+ Snekbox: HTTP POST
Snekbox ->>+ NsJail: Python code
NsJail ->>+ Python: Python code
Python -->>- NsJail: Execution result
NsJail -->>- Snekbox: Execution result
Snekbox -->>- Client: JSON response
The code is executed in a Python process that is launched through NsJail, which is responsible for sandboxing the Python process.
The output returned by snekbox is truncated at around 1 MB by default, but this can be configured.
Communication with snekbox is done over a HTTP REST API. The framework for the HTTP REST API is Falcon and the WSGI being used is Gunicorn. By default, the server is hosted on 0.0.0.0:8060
with two workers.
See snekapi.py
and resources
for API documentation.
A Docker image is available in the GitHub Container Registry. A container can be started with the following command, which will also pull the image if it doesn't currently exist locally:
docker run --ipc=none --privileged -p 8060:8060 ghcr.io/python-discord/snekbox
To run it in the background, use the -d
option. See the documentation on docker run
for more information.
The above command will make the API accessible on the host via http://localhost:8060/
. Currently, there's only one endpoint: http://localhost:8060/eval
.
By default, the executable that runs within nsjail is defined by DEFAULT_EXECUTABLE_PATH
at the top of nsjail.py
. This can be overridden by specifying executable_path
in the request body of calls to POST /eval
or by setting the executable_path
kwarg if calling NSJail.python3()
directly.
Any executable that exists within the container is a valid value for executable_path
. The main use case of this feature is currently to specify the version of Python to use.
Python versions currently available can be found in the Dockerfile
by looking for build stages that match builder-py-*
. These binaries are then copied into the base
build stage further down.
Configuration files can be edited directly. However, this requires rebuilding the image. Alternatively, a Docker volume or bind mounts can be used to override the configuration files at their default locations.
The main features of the default configuration are:
/home
NsJail is configured through snekbox.cfg
. It contains the exact values for the items listed above. The configuration format is defined by a protobuf file which can be referred to for documentation. The command-line options of NsJail can also serve as documentation since they closely follow the config file format.
On each execution, the host will mount an instance-specific tmpfs
drive, this is used as a limited read-write folder for the sandboxed code. There is no access to other files or directories on the host container beyond the other read-only mounted system folders. Instance file systems are isolated; it is not possible for sandboxed code to access another instance's writeable directory.
The following options for the memory file system are configurable as options in gunicorn.conf.py
memfs_instance_size
Size in bytes for the capacity of each instance file system.memfs_home
Path to the home directory within the instance file system.memfs_output
Path to the output directory within the instance file system.files_limit
Maximum number of valid output files to parse.files_timeout
Maximum time in seconds for output file parsing and encoding.files_pattern
Glob pattern to match files within output
.The sandboxed code execution will start with a writeable working directory of home
. By default, the output folder is also home
. New files, and uploaded files with a newer last modified time, will be uploaded on completion.
Gunicorn settings can be found in gunicorn.conf.py
. In the default configuration, the worker count, the bind address, and the WSGI app URI are likely the only things of any interest. Since it uses the default synchronous workers, the worker count effectively determines how many concurrent code evaluations can be performed.
wsgi_app
can be given arguments which are forwarded to the NsJail
object. For example, wsgi_app = "snekbox:SnekAPI(max_output_size=2_000_000, read_chunk_size=20_000)"
.
All environment variables have defaults and are therefore not required to be set.
Name | Description |
---|---|
SNEKBOX_DEBUG |
Enable debug logging if set to a non-empty value. |
SNEKBOX_SENTRY_DSN |
Data Source Name for Sentry. Sentry is disabled if left unset. |
By default, the Python interpreter has no access to any packages besides the standard library. Even snekbox's own dependencies like Falcon and Gunicorn are not exposed.
To expose third-party Python packages during evaluation, install them to a custom user site:
docker exec snekbox /bin/sh -c \
'PYTHONUSERBASE=/snekbox/user_base /snekbin/python/default/bin/python -m pip install --user numpy'
In the above command, snekbox
is the name of the running container. The name may be different and can be checked with docker ps
.
The packages will be installed to the user site within /snekbox/user_base
. To persist the installed packages, a volume for the directory can be created with Docker. For an example, see docker-compose.yml
.
See CONTRIBUTING.md.