In-Process Python Functions

This section describes the required steps to enable and use In-Process Python functions on a production SQreamDB machine.

Overview 

In order to use In-Process Python functions on a production machine, several system-level dependencies must be installed and configured manually.

Before proceeding, make sure you have completed all steps described in: Pre-Installation Configuration

This section describes how to verify, install, and configure the CUDA Toolkit required for In-Process Python functions.

The CUDA Toolkit version must match the installed CUDA driver version.

First, verify whether CUDA is already present on the system.

Check if a CUDA repository is installed:

rpm -qa | grep cuda

If a CUDA repository is installed, you should see output similar to:

cuda-repo-rhel8-12-3-local-12.3.2_545.23.08-1.x86_64

If no CUDA repository is found (cuda-repo does not appear):

Check repository files:
```
ls -l /etc/yum.repos.d/ | grep cuda
```
If CUDA was installed manually using a runfile, the repository will not appear.

Verify where CUDA is installed:
```
which nvcc
```
If CUDA binaries are not in PATH, check the default installation location:
```
cd /usr/local
ls -l | grep cuda
```

You should see output similar to:

cuda -> /usr/local/cuda-13.0/
cuda-13.0/

If installed cuda toolkit does not match the driver version, it must be removed before installing the correct version.

Verify installed toolkit packages:

rpm -qa | grep cuda-toolkit

Remove all CUDA Toolkit packages:

sudo dnf remove "cuda-toolkit*"

If the toolkit was installed using a runfile:

cd /usr/local
sudo rm -rf /usr/local/cuda-13.0/
sudo rm cuda

Install the CUDA Toolkit that matches your driver version.

sudo dnf install cuda-toolkit-12-3

Warning

If your CUDA driver version is different (for example 12.9), you must install the matching toolkit:

sudo dnf install cuda-toolkit-12-9

Download the runfile:

wget https://developer.download.nvidia.com/compute/cuda/12.3.2/local_installers/cuda_12.3.2_545.23.08_linux.run

Install the toolkit only:

sudo sh cuda_12.3.2_545.23.08_linux.run \
  --silent \
  --toolkit \
  --no-man-page \
  --override

Add CUDA library paths to the dynamic linker configuration:

sudo tee /etc/ld.so.conf.d/cuda-12-3.conf > /dev/null << 'EOF'
/usr/local/cuda-12.3/lib64
/usr/local/cuda-12.3/targets/x86_64-linux/lib
EOF

Reload the dynamic linker configuration:

sudo ldconfig

Ensure Python 3.11 is installed and set as the default Python version (as required by the pre-installation configuration).

python3 --version

Expected output:

Python 3.11.7

Install CuPy using pip:

pip3 install cupy

Alternatively, install CuPy using one of the supported methods described in: https://docs.cupy.dev/en/stable/install.html#installing-cupy-from-pypi

After completing all steps:

The system is now ready to run In-Process Python functions within SQreamDB.