The short answer
The dbt (Directed Acyclic Graph) is the graph of models, sources, seeds, snapshots, and tests connected by dependencies. Nodes are resources; edges come mainly from ref(), source(), and test relationships.
Mental model
source(raw.orders) source(raw.customers)
| |
v v
stg_orders stg_customers
\ /
\ /
v v
fct_orders
|
v
dim_customer_orders (example mart)"Directed" means dependencies have a direction (upstream → downstream). "Acyclic" means no cycles: A cannot ref B if B already depends on A.
How dbt builds it
- 1Parse all models and YAML.
- 2Resolve every
ref('x')/source('a','b'). - 3Order execution so upstream models run before downstream ones.
- 4Power selectors like
dbt run --select +fct_orders+(parents, itself, children).
Why it matters
- Correct run order without manual scripts
- Impact analysis: "if I change
stg_orders, what breaks?" - Selective runs and CI on changed subgraphs
- Lineage in
dbt docs
Interview tip: Contrast with Airflow DAGs: Airflow orchestrates tasks/jobs; the dbt DAG describes data dependencies among SQL models inside the warehouse transform layer.
Reading selectors on the graph
The + sign walks the graph. dbt run --select +fct_orders runs fct_orders and everything it depends on. --select fct_orders+ runs it and everything that depends on it. Before running anything, dbt ls --select stg_orders+ prints which nodes a selection would touch, which is a quick impact check.
If 2 models ref each other, dbt refuses to run and reports a cycle at parse time. That is the "acyclic" rule being enforced: the fix is to move the shared logic into a third, earlier model that both can ref.