API schema design, drift detection internals, CI/CD patterns, and what we've learned building validation infrastructure at scale.
Schema diff is a deceptively hard problem. A naïve implementation that resolves every $ref and does a full AST comparison would be accurate but unacceptably slow on large specs. Here's how we built Homogenic's core diff engine to stay under 2 seconds on specs with 300+ endpoints — including the structural hashing trick that made it possible.
When we were designing Homogenic's contract format, we considered Protobuf, Thrift, Avro, and JSON Schema. This is the decision document we wish we'd had — the trade-offs around tooling, readability, ecosystem support, and what "schema-first" actually means in practice. Spoiler: we don't think Protobuf is wrong, just wrong for this problem.
GraphQL SDL support has been on our roadmap since v1.2. We shipped it in v2.4.0 — over a year after we first planned to. This post is an honest account of why it took so long: the fundamental differences between type-checking a REST contract and a GraphQL schema, the nullable handling rabbit hole, and why schema stitching turned out to be a harder problem than we expected.
A nullable field that starts returning null is the most dangerous kind of API change — because it's technically valid, it passes most validation checks, and it surfaces as a NullPointerException in production at 2am. This post walks through a real incident (anonymized) where this exact scenario caused a 45-minute outage, and how Homogenic would have caught it at the PR stage.
After 14,000 developer teams have connected their repos to Homogenic, we've started seeing patterns in how the successful ones structure schema governance across microservices. This post covers the three models we see most often — centralized schema registry, federated ownership with shared conventions, and the hybrid — and when each one is the right call.
Schema validation CI checks fail in a predictable way: they work fine, then one PR generates a false positive, a frustrated engineer turns off the check, and it never comes back on. We've spent two years studying our activation and deactivation data to understand what makes developers keep a CI integration enabled — and redesigned our GitHub Action based on what we found.