Fifteen services, one install

A data platform assembled from best-of-breed open-source systems has a dirty secret: it is miserable to deploy. Fifteen services, eight databases, an identity provider, vector stores, analytics engines, and orchestration, each with its own installation story, is not a product. It is a consulting engagement.

I own the work that turned ours into a product: the entire platform deploys into a customer’s Kubernetes environment in one step.

Three commitments made it real. First, single sign-on across every component with one identity flowing through every layer, because a platform with fifteen login screens is fifteen products. Second, isolation by default: each service owns its own database, every resource is scoped to its release, and multiple instances coexist on one cluster without contention. Third, operator-managed state: the databases, vector stores, and compute clusters are run by Kubernetes operators rather than by hand, so day-two operations do not depend on the person who did day one.

The discipline that protects all of it is a release-gating integration suite: dozens of cross-service scenarios that run against a live deployment before any release ships. When the test suite is the contract, “it installs” stops being a claim and becomes a checkable fact.

The deeper lesson: in platform engineering, installability is a feature with a higher return than almost any capability you could build instead, because it is the feature every customer experiences first.