DevConf.IN 2026

Feature engineering is eating your training time. Data loading is your bottleneck. Sound familiar?
If your training jobs crawl, your features take forever to compute, or your pipeline breaks every time you scale, this talk is for you.

In this session, we’ll show how to turn a slow, file-based ML pipeline into a distributed, production-ready architecture using modern open-source tooling:
- Feast for feature management
- Ray for distributed data processing
- Kubeflow Training Operator for orchestrating distributed training on Kubernetes

We’ll demonstrate an end-to-end pipeline, powering a Temporal Fusion Transformer trained on 421K rows of Walmart sales data. Using PyTorch DDP across multiple GPUs, how we can cut training time, while hitting 10.5% MAPE (compared to the typical 15–20% industry baseline).

You’ll see:
- Faster feature loading using Ray + Feast
- Raw data flowing through a fully managed feature platform
- Distributed PyTorch jobs launched and scaled with Kubeflow Training Operator
- Production inference path powered by Feast’s hybrid storage & compute
- How Ray transforms feature engineering performance at scale
- How Feast standardizes feature computation across training & inference

You’ll leave with a repeatable blueprint for building ML pipelines that scale as your models, data, and teams grow, along with the confidence to adopt these tools in your own production environment.