cosmos/up.sh restarts the stack on the Cosmos image with docker-compose.cosmos.yaml, which mounts dags_cosmos/ into the DAG folder; the plain stack and the tests of Unit Testing DAGs with pytest never import Cosmos. cosmos/c2.sh then ran booknest_daily_cosmos for 30 June:
task_id | state | try | started | secs --------------------------------+---------+-----+----------+------ ingest.extract | success | 1 | 00:46:32 | 2.0 ingest.load | success | 1 | 00:46:34 | 1.1 sales_mart.stg_orders_run | success | 1 | 00:46:37 | 10.5 ... sales_mart.genre_groups_seed | success | 1 | 00:46:37 | 10.1 sales_mart.fct_sales.run | success | 1 | 00:46:48 | 12.6 sales_mart.dim_book.run | success | 1 | 00:46:49 | 12.4 sales_mart.fct_sales.test | success | 1 | 00:47:01 | 13.6 sales_mart.dim_book.test | success | 1 | 00:47:02 | 12.6 check | success | 1 | 00:47:15 | 1.5 publish | success | 1 | 00:47:17 | 1.8 run success in 48 s
The run took 48 seconds where the BashOperator version took 26-31; the graph now shows what dbt 37,942 did. The first attempt had failed in ingest.load before dbt started: l2-pg had been stopped underneath the stack, and the two retries could not help. Had a test failed instead, only that model's group would need clearing, and publish would never have written a partial day. For BookNest the trade is worth it at four mart models; with hundreds, use TestBehavior.BUILD or the watcher mode, and keep the per-model view.