Pipeline Requirements Management and Dependency Conflicts

[Maleya]

The current pipeline requirements installation system installs all dependencies globally as root without isolation, leading to potential dependency conflicts, security vulnerabilities, and overall system instability.

Current Behavior

As far as I can see pipeline requirements are handled through two mechanisms:

1. Frontmatter requirements - specified in pipeline docstrings (e.g., requirements: langfuse<3.0.0)
2. Global requirements file - via PIPELINES_REQUIREMENTS_PATH environment variable

Both methods install packages globally using:

subprocess.check_call([sys.executable, "-m", "pip", "install", req])

Via:

• Pipeline loading: main.py:load_module_from_path()
• Frontmatter parsing: main.py:install_frontmatter_requirements()
• Setup script: start.sh:install_frontmatter_requirements()

Issues

This leads to a number of potentially serious issues.

• Multiple pipelines can specify conflicting version requirements
  • Example: Pipeline A requires requests==2.25.0, Pipeline B requires requests>=2.30.0
  • No conflict resolution mechanism exists
  • Last installed version "wins", potentially breaking other pipelines
• All packages installed with root privileges
  • No sandboxing or permission restrictions
• All dependencies installed in global Python environment
  • Potential conflicts with base system requirements
  • Installation order affects final environment state
  • Same pipeline configuration can result in different environments

Reproduction Steps

This should be fairly easy to reproduce. For instance:

1. Create two pipelines with conflicting requirements
2. Install both pipelines
3. Observe that only the last installed version is available

Expected Behavior

• Pipelines should have isolated dependency environments
• Dependency conflicts should be detected and reported
• System should remain stable regardless of pipeline requirements
• Security boundaries should prevent unauthorized system modifications

Suggested changes

At the very least each pipeline should have its own virtual environment (1) to avoid conflicts. But further changes could be:

1. Virtual Environment Isolation: Each pipeline gets its own virtual environment
2. Container-based Isolation: Use containers for pipeline execution
3. Dependency Caching: Shared cache for common dependencies
4. Package Verification: Verify package integrity and sources
5. Resource Limits: Implement installation timeouts and size limits

Related Github issues

• https://github.com/open-webui/pipelines/issues/477

[winstonallo]

+1 on this, this would be a huge improvement