Alright, let's talk CI/CD for a Python monorepo. We just went through a six-month evaluation of GitHub Actions, CircleCI, and GitLab CI. Our setup: a single repo with multiple Python services and shared libraries, Docker builds, integration tests, and the usual linting/pylint/mypy steps.
We started on GitHub Actions because it was the path of least resistance. The integration is seamless, and the configuration syntax felt straightforward. But as we scaled, the monorepo complexity bit us. The lack of built-in dependency caching between workflow runs for Python packages became a huge time sink. We tried caching `pip` installs, but matrix builds across Python versions and services made it messy. Our monthly bill started creeping up as we added more jobs—the per-minute cost adds up faster than you think when you're running heavy integration suites.
We then tried CircleCI, lured by its powerful caching and optimized Docker layer caching. The performance *was* better. However, the configuration complexity skyrocketed. Our `.circleci/config.yml` became a maze of orbs and conditional logic to handle monorepo path filtering. Debugging why a job didn't trigger for a specific service change was a headache. Also, their credit-based pricing felt opaque. We'd burn through credits on a few heavy build days and then hit a slowdown. The vendor said we could optimize with their new "runner fleet," but the setup and management overhead was a non-starter for our small platform team.
Finally, we gave GitLab CI a serious shot (using it with GitHub repos via mirroring). The built-in monorepo features were a game-changer. Using `rules:changes:` for path filtering is so much cleaner. Here's a snippet from our config:
```yaml
deploy-service-a:
stage: deploy
script:
- echo "Deploying Service A"
rules:
- if: $CI_COMMIT_BRANCH == "main"
changes:
- "services/service-a/**/*"
```
Plus, the built-in container registry and dependency proxying for PyPI cut our external pulls significantly. The cost? We're on their Premium tier, but it's predictable. The observability into pipeline performance is also better out-of-the-box compared to the others.
So, what actually happened? We're migrating to GitLab CI. GitHub Actions felt too basic and costly at scale. CircleCI was powerful but complex and financially unpredictable for our bursty workload. GitLab's tooling felt built for the monorepo reality.
Would we renew? We're sticking with GitLab CI for now. The setup isn't perfect—mirroring adds a slight delay, and their documentation can be sprawling—but the overall fit for a complex Python monorepo is just better.
cost first, then scale