The schema registry as metadata

Why can schema lookups be cached forever, and what general principle does that show?

A schema registry stores the schema (field names and types) of messages, giving each version an ID. Unlike most catalogs, it is load-bearing at read time: a consumer cannot decode a message without the schema, so the registry sits in the data path, and its availability is a schema-evolution concern. What belongs here is why the usual fix is safe. Consumers cache schemas aggressively, and the cache is correct for free, because schema IDs are immutable: an ID never maps to a different schema. So the cache never needs invalidation, and its time-to-live (TTL) can be infinite.

This is a general and underused principle:

That is the whole reason schema IDs are not schema names. A name can be re-pointed to a new schema; an ID cannot.

Case, resolvedConstructed case · Hana at Trailmix

A terrible plan on the fastest-growing table

The planner picked a terrible plan on the fastest-growing table, pipeline version 14 wrote wrong values for two days unnoticed, and a 20-minute schema-service outage stopped every consumer. Where it started: The description layer →

Bad plan on page_views. The statistics came from a nightly scan of a table that changes every ten minutes. She moves the table to write-time statistics stored in manifests (min, max, counts and distinct-count sketches), so every new file brings its own exact numbers. For queries with many joins she turns on adaptive re-planning, because no statistic would fix an 81× compound error.
Version 14. She adds _pipeline_version and _ingested_at columns. Because batches share the values, the extra storage is negligible. The next incident is repaired by reprocessing only the rows from the bad run, not 30 days of data. She also makes pipeline outputs go to a branch first, and publishes only after checks pass.
Schema outage. The consumers now cache each schema by ID with no expiry. In the next outage, running consumers keep going; only brand-new schema versions wait.