The MLOps Industrial Complex
Somewhere around 2022, the ML industry convinced itself that you need a 47-tool stack to deploy a model. Kubeflow, MLflow, Airflow, Prefect, DVC, Weights & Biases, Seldon, BentoML, Feature Stores, Model Registries, Experiment Trackers...
Most of you need none of this.
I'm not being contrarian for clicks. I've built and maintained ML pipelines at scale. I've used every tool in the MLOps landscape. And I'm telling you: for 90% of companies, these pipelines are the most expensive lines of code ever written — not because they cost a lot to build, but because they cost a fortune to maintain and slow everything down.
{
"type": "comparison",
"left": {
"title": "What You Built",
"color": "red",
"steps": ["Feature Store", "Training Pipeline", "Experiment Tracker", "Model Registry", "CI/CD Pipeline", "Serving Infrastructure", "Monitoring Stack", "Retraining Trigger ↩"]
},
"right": {
"title": "What You Needed",
"color": "green",
"steps": ["Python Script", "Model File", "API Endpoint"]
}
}
The Real Numbers
At a previous company, our ML pipeline:
- Took 6 weeks to set up before any model work began
- Required 2 full-time engineers just to maintain the infrastructure
- Added 4-6 hours to every model update cycle
- Cost $8K/month in infrastructure alone
The model it served? A gradient-boosted classifier that could run on a laptop.
When You Actually Need MLOps
Real MLOps infrastructure is justified when:
- You're retraining daily on new data
- You have dozens of models in production simultaneously
- Regulatory compliance requires model versioning and audit trails
- Your inference volume exceeds 10K requests/second
If none of those apply — and for 90% of companies, they don't — just deploy a Docker container with a FastAPI endpoint and move on with your life.
The best ML infrastructure is the one you didn't build.
