Dependency Verification

Diagnose and resolve package version mismatches between local and distributed worker environments.

When running Geneva UDFs on distributed workers, your code is serialized locally and executed on remote workers. If the worker environment differs from your local environment, you may encounter subtle and difficult-to-debug errors.

Example environment mismatch errors

Symptom	Likely Cause
TypeError: Enum.__new__() missing 1 required positional argument	attrs version mismatch
TypeError: Can't instantiate abstract class	Package structure differences
ArrowInvalid: cannot cast / serialization errors	NumPy 1.x vs 2.x mismatch
ModuleNotFoundError on workers	Package only installed locally
Model loading failures	PyTorch version mismatch
Permission denied errors	Missing API keys in envrionment variables

These issues are notoriously difficult to debug because the error messages often don’t indicate the root cause.

The compare_ray_environments Tool

Geneva provides a diagnostic tool to compare your local environment against Ray workers.

If you are encountering a hang or exception you can use the following diagnosis worklflow to resolve the problem.

1. Run the diagnostic tool programatically or via the CLI.

2. Check PACKAGES and ENV VARS output sections for mismatches.

3. Identify critical packages: numpy, torch, pyarrow, attrs, pydantic.

4. Identify inconsistent environment variables: AWS_*, GOOGLE_APPLICATION_CREDENTIALS

5. Fix with manifest for quick testing:

6. OPTIONAL: Build custom image for production (if using KubeRay).

Programmatic Usage

Python

from geneva.runners.ray.compare_env import compare_ray_environments

# Compare and print (requires Geneva context to be initialize via `with db.context(..)`)
result = compare_ray_environments()

# Compare environments, filtering environment variables with specified prefix
result = compare_ray_environments(env_prefix="PY")

CLI Usage

CLI

# Connect to existing Ray cluster
python -m geneva.runners.ray.compare_env

# Start new local Ray cluster
python -m geneva.runners.ray.compare_env --address local

# Filter env vars by prefix
python -m geneva.runners.ray.compare_env --env-prefix RAY

# Show full JSON snapshots
python -m geneva.runners.ray.compare_env --show-all

# Skip sys.path comparison
python -m geneva.runners.ray.compare_env --no-sys-path

Understanding the Output

The tool outputs several sections to help you identify mismatches.

PYTHON / PLATFORM

Shows Python version and OS information for both environments:

=== PYTHON / PLATFORM ===
Local:
  Python: 3.11.9 (main, Apr 19 2024, 16:48:06) [GCC 11.2.0]
  Impl  : CPython
  Exec  : /home/user/.venv/bin/python
  OS    : Linux 5.15.0-generic (x86_64)

Remote:
  Python: 3.11.9 (main, Apr 19 2024, 16:48:06) [GCC 11.2.0]
  Impl  : CPython
  Exec  : /home/ray/anaconda3/bin/python
  OS    : Linux 5.4.0-aws (x86_64)

Watch for different Python versions or different OS types (macOS local vs Linux remote).

Architecture Mismatch (macOS to Linux)

If you see different OS types (e.g., Darwin locally vs Linux remotely), compiled extensions may fail with ModuleNotFoundError or segfaults.

Solution: Run Geneva from the same OS/architecture as your cluster (typically Linux x86_64). Use a Linux VM, container, or remote development environment.

Environment Variables

Environment variables present in only one environment:

=== ENV VARS: keys only in LOCAL ===
  + CONDA_PREFIX
  + VIRTUAL_ENV

=== ENV VARS: keys only in REMOTE ===
  + RAY_ADDRESS
  + KUBERNETES_SERVICE_HOST

Missing AWS_* or GOOGLE_APPLICATION_CREDENTIALS can cause storage authentication failures.

Passing Environment Variables to Workers

If critical environment variables are missing on workers, you can pass them via the manifest or cluster configuration.

Option 1: Via Manifest

Python

from geneva.manifest.builder import PipManifestBuilder
import os

manifest = (
    PipManifestBuilder.create("my-manifest")
    .env_vars({
        "AWS_ACCESS_KEY_ID": os.environ["AWS_ACCESS_KEY_ID"],
        "AWS_SECRET_ACCESS_KEY": os.environ["AWS_SECRET_ACCESS_KEY"],
        "MY_API_KEY": os.environ["MY_API_KEY"],
    })
    .build()
)

Option 2: Via Cluster Configuration

Avoid hardcoding secrets. Use os.environ to pass values from your local environment, or use a secrets manager in production.

Packages

The tool shows version mismatches and packages only present in one environment:

=== PACKAGES: version mismatches ===
  * numpy: local=1.26.4  remote=2.2.6
  * torch: local=2.0.1  remote=2.8.0+cpu
  * attrs: local=23.2.0  remote=24.2.0
  * pyarrow: local=14.0.1  remote=17.0.0

=== PACKAGES: only in LOCAL ===
  + my-custom-package
  + dev-tools

=== PACKAGES: only in REMOTE ===
  + kuberay-client

Watch for major version differences (NumPy 1.x vs 2.x) and PyTorch version mismatches.

Common Package Issues

Issue	Symptoms	Fix
NumPy 1.x vs 2.x	ArrowInvalid, ValueError: cannot convert, serialization failures	Pin numpy==1.26.4
PyTorch mismatch	Model loading failures, CUDA errors, unexpected inference results	Pin to matching torch version
attrs mismatch	TypeError: Enum.__new__() missing 1 required positional argument	Pin attrs to local version
Missing package	ModuleNotFoundError: No module named 'xyz'	Add package to manifest

Fixing Package Mismatches

Option 1: Manifest pip Dependencies

Specify packages in a Geneva manifest for a quick fix:

Pros: Quick, reusable across sessions, stored with your database.

Cons: Slower startup (downloads packages), may not work for complex dependencies.

Option 2: Custom Ray Worker Image

For KubeRay deployments, build a custom worker image:

Dockerfile

# Dockerfile.ray-worker
FROM rayproject/ray:2.30.0-py311

# Install exact versions
RUN pip install \
    numpy==1.26.4 \
    torch==2.0.1 \
    attrs==23.2.0 \
    geneva==0.8.0

# Copy any custom packages
COPY ./my_udfs /app/my_udfs

Then reference in RayCluster spec:

Kubernetes

spec:
  workerGroupSpecs:
    - template:
        spec:
          containers:
            - image: your-registry/ray-worker:latest

Pros: Fastest startup, reproducible.

Cons: Requires image build/push workflow.

Option 3: Conda Environment

Use a conda environment on workers via the cluster builder:

Or specify conda channels and dependencies inline:

Pros: Best for complex dependencies with native libraries (ffmpeg, CUDA).

Cons: Slower environment creation, requires conda on workers.

API Reference

For detailed API documentation on the environment comparison functions, see the Geneva Diagnostics API Reference.