The win is in two layers: a key-name denylist for fields you know are volatile (createdAt, requestId, traceId) and a value-shape sweep for things that look like dates, uuids, or hashes regardless of where they appear. I have learned to keep them separate because the shape sweep catches surprises: an upstream API adds a metadata.runId: '7c1ad9...' and your snapshot does not break. The output uses angle-bracket tokens (<ISO_DATE>) rather than dropping the field, so a missing field still fails the snapshot but a churning timestamp does not.

Jest's serializer contract is two functions: test(value) returns true when this serializer should handle the value, and serialize(value, ...) returns the string. Narrowing test to plain objects and arrays is essential: if you return true for everything, you steal serialization from Jest's React plugin and your <App /> snapshots turn into JSON garbage. Deferring to Jest's printer argument is the right move in production; the example uses JSON.stringify because we are simulating outside a real test runner. Once registered, every expect(x).toMatchSnapshot() automatically gets the redaction.

The override layer is what keeps the serializer useful past the first dozen tests. Most tests want the global rules, a few want extra ones (Stripe ids in a billing test, IPs in a logging test), and you do not want to register a brand-new serializer every time. A try-finally restore is the simplest scope-aware override mechanism and survives Jest's parallel-test model because each worker is its own process. In practice I expose withRedaction from the shared test helper file and call it inside the specific describe block that needs custom rules.