search

NVIDIA / spark-rapids-jni

RAPIDS Accelerator JNI For Apache Spark
Apache License 2.0
32 stars 60 forks source link

Improved detection of bad dependencies from native build #2164

Closed jlowe closed 1 week ago

jlowe commented 2 weeks ago

Is your feature request related to a problem? Please describe.

2162 should be something caught during the build process.

Describe the solution you'd like The build should analyze the dynamic dependencies of the shared libraries being produced and verify nothing unexpected appears as a dependency.

mattahrens commented 2 weeks ago

Scope: add new test in a different environment such that NVIDIA driver is the only NVIDIA package available.

pxLi commented 2 weeks ago

Let me try figure out some good solutions to detect the issue during submodule-syncup, pre-merge or nightly-build

gerashegalov commented 2 weeks ago

Different environment could be just another Dockerfile (different from the one used for build-in-docker) for testing that does not contain (nor mounts from the host) the paths containing cudart.

pxLi commented 2 weeks ago

Different environment could be just another Dockerfile (different from the one used for build-in-docker) for testing that does not contain (nor mounts from the host) the paths containing cudart.

yep, this could simply be within a non-cuda image, and then check ldd status.

decide to use the same ENV but check cuda-related dynamical links directly to save resources https://github.com/NVIDIA/spark-rapids-jni/pull/2173