Backend.AI Kernel Runner Package for glibc-based Kernels
This package contains a statically built Python distribution and 3rd-party libraries required to run our krunner module inside containers whose images are provided by users.
The krunner wheel itself can be installed into any Python environments (the content of this plugin package is agnostic to platforms and architectures as it's a mere declaration of the plugin interface), but we still apply the binary wheel platform tags so that setuptools could distinguish the target CPU architecture as follows:
| Where is Backend.AI running? | What is the user container's base image? | The krunner wheel used |
|---|---|---|
| manylinux (x86-64) | manylinux (x86-64) | backend.ai_krunner_static_gnu-X.X.X-py3-none-manylinux2014_x86_64.musllinux_1_1_x86_64.macosx_11_0_x86_64.whl |
| manylinux (x86-64) | musllinux (x86-64) | backend.ai_krunner_alpine-X.X.X-py3-none-manylinux2014_x86_64.musllinux_1_1_x86_64.macosx_11_0_x86_64.whl |
| manylinux (aarch64) | manylinux (aarch64) | backend.ai_krunner_static_gnu-X.X.X-py3-none-manylinux2014_aarch64.musllinux_1_1_aarch64.macosx_11_0_arm64.whl |
| manylinux (aarch64) | musllinux (aarch64) | backend.ai_krunner_alpine-X.X.X-py3-none-manylinux2014_aarch64.musllinux_1_1_aarch64.macosx_11_0_arm64.whl |
| musllinux (x86-64) | manylinux (x86-64) | backend.ai_krunner_static_gnu-X.X.X-py3-none-manylinux2014_x86_64.musllinux_1_1_x86_64.macosx_11_0_x86_64.whl |
| musllinux (x86-64) | musllinux (x86-64) | backend.ai_krunner_alpine-X.X.X-py3-none-manylinux2014_x86_64.musllinux_1_1_x86_64.macosx_11_0_x86_64.whl |
| musllinux (aarch64) | manylinux (aarch64) | backend.ai_krunner_static_gnu-X.X.X-py3-none-manylinux2014_aarch64.musllinux_1_1_aarch64.macosx_11_0_arm64.whl |
| musllinux (aarch64) | musllinux (aarch64) | backend.ai_krunner_alpine-X.X.X-py3-none-manylinux2014_aarch64.musllinux_1_1_aarch64.macosx_11_0_arm64.whl |
| macOS (x86-64) | manylinux (x86-64) | backend.ai_krunner_static_gnu-X.X.X-py3-none-manylinux2014_x86_64.musllinux_1_1_x86_64.macosx_11_0_x86_64.whl |
| macOS (x86-64) | musllinux (x86-64) | backend.ai_krunner_alpine-X.X.X-py3-none-manylinux2014_x86_64.musllinux_1_1_x86_64.macosx_11_0_x86_64.whl |
| macOS (aarch64) | manylinux (aarch64) | backend.ai_krunner_static_gnu-X.X.X-py3-none-manylinux2014_aarch64.musllinux_1_1_aarch64.macosx_11_0_arm64.whl |
| macOS (aarch64) | musllinux (aarch64) | backend.ai_krunner_alpine-X.X.X-py3-none-manylinux2014_aarch64.musllinux_1_1_aarch64.macosx_11_0_arm64.whl |
We named the krunner package based on musl 1.2 as "alpine" because practically Alpine linux is the only distribution which actively uses musl and it is currently not possible to build static CPython which can import 3rd-party dynamic modules on top of the musl ecosystem.
This package is to distribute prebuilt binaries, so the source distribution does not have prebuilt binaries and does not work as intended. Just refer this repository on how we build stuffs.
{distro}is a string likestatic-gnu,alpine, etc. depending on which repository you are in.{distro_}is a string same to{distro}but with hyphens replaced with underscores for Python package names and paths. (e.g.,static_gnu,alpine)
$ git clone https://github.com/lablup/backend.ai-krunner-{distro} krunner-{distro}
$ cd krunner-{distro}
$ pyenv virtualenv 3.12.2 venv-krunner # you may share the same venv with other krunner projects
$ pyenv local venv-krunner
$ pip install -U pip setuptools
$ pip install -U -r build/requirements.txt
$ pip install -U -e .- Modify Dockerfile and/or other contents.
- To update the Python version, update
src/ai/backend/krunner/{distro_}/krunner-python.{distro}.txtand the dockerfiles (both python and wheels) accordingly, including thePYTHON_VERSIONenvironment variable and the download URL of the statically built Python distribution.
- Increment the volume version number specified as a label
ai.backend.krunner.versioninsrc/ai/backend/krunner/{distro_}/krunner-env.{distro}.dockerfile - Run
scripts/build.py. - Repeat the above steps for each distro version. (For static builds, there is only one.)
- Increment the package version number in
src/ai/backend/krunner/{distro_}/__init__.py rm -r dist/* build/*(skip if these directories do not exist and or are empty)- Commit.
- Create a signed annotated tag and push the tag to let GitHub Action build and publish wheels.
Note that src/ai/backend/krunner/{distro_}/krunner-version.{distro}.txt files are
overwritten by the build script from the label.
WARNING: We should choose x86_64_v2 binaries from the indygreg repository
when updating the Python runtime version for CPU compatibility with some of our
test setups and customer sites.
Use CentOS 7 or later and install this list of packages. Also refer the test script.
- ttyd uses
muslas their C stdlib, notglibc. - The
musltoochain used by the build script is x86_64 binaries.
libwebsockets>=4.0.0 features auto ping/pong with 5 min default interval.
(https://github.com/warmcat/libwebsockets#connection-validity-tracking) And,
ws_ping_pong_interval of ttyd is not effective in libwebsockets>=4.0.0.
This seems to be the reason why ttyd>=1.6.1 does not set
ws_ping_pong_interval for libwebsockets>=4.0.0.
(https://github.com/tsl0922/ttyd/blob/master/src/server.c#L456)
To fix this issue, we modify and build the latest version of libwebsockets used by the ttyd build script manually.
# Prepare Ubuntu environment (possibly, through container) and dependencies.
sudo apt-get update
sudo apt-get install -y autoconf automake build-essential cmake curl file libtool
# Download ttyd source.
git clone https://github.com/tsl0922/ttyd.git
cd ttydNow let's modify ./scripts/cross-build.sh.
Add these two lines under pushd "${BUILD_DIR}/libwebsockets-${LIBWEBSOCKETS_VERSION}":
sed -i 's/context->default_retry.secs_since_valid_ping = 300/context->default_retry.secs_since_valid_ping = 20/g' lib/core/context.c
sed -i 's/context->default_retry.secs_since_valid_hangup = 310/context->default_retry.secs_since_valid_hangup = 30/g' lib/core/context.c Finally, build the ttyd binary.
# Run build script.
./scripts/cross-build.sh
# Check ttyd binary version.
./build/ttyd --version