dbt at Scale — Week 1 CheckWeek 1 of dbt & Analytics Engineering: what breaks past a few thousand models, and the materialization decisions that dominate cost.

The four core materializations — view, table, incremental, ephemeral — trade off what?

Incremental models are the biggest documented cost lever in the corpus (100-200× BigQuery savings in one case). What's the correctness price?

Past a few thousand models (Discord-scale), what breaks FIRST in a stock dbt monorepo?

`dbt build` on every merge rebuilds the whole project and takes 3 hours. The standard fix:

ref() is more than syntactic sugar for a table name. What does it buy structurally?

A model rebuilds 90 days of history every run 'to be safe' — the analytics version of a non-idempotent backfill problem. The disciplined alternative: