Feature Store Isolation at Scale

One of the key decisions when adopting a feature store is the tenancy model. For most organizations starting out, a single shared instance is the right call: one platform team manages it, one bill, simple governance.

As ML adoption grows, that model starts to show cracks. A team running high-volume writes affects query latency for everyone else. Schema changes need coordination across teams. Peak traffic sizing becomes a negotiation. The platform team sits in the critical path of every change.

At Zalando, we moved to dedicated isolated Hopsworks clusters per team. Thirteen are in production today, each independently sized and owned by the team that uses it.

This talk explores how we made that decision and what operating that model looks like in practice. It covers:

- When a shared setup stops working and what the signals are - What isolation buys teams in autonomy and independent scaling - What it costs: the infrastructure bill, the operational overhead, the things that get harder - How we manage a fleet of clusters - Where we think the model breaks