A Dagster deployment has a few moving parts. Code locations load your definitions in their own processes and environments, so teams pin their own dependencies and deploy independently. The webserver serves the UI. The daemon evaluates schedules, sensors and automation, and launches queued runs. A run launcher starts each run (a process, a container, a Kubernetes job), and an executor runs its steps.
The run queue limits how many runs execute at once; concurrency pools cap how many steps use a shared resource across all runs: @dg.asset(pool="warehouse") with a limit on the warehouse pool keeps a backfill from opening a hundred connections.
You can run open-source Dagster yourself (Docker, or Kubernetes with the Helm chart), or use Dagster+, the hosted product, in serverless or hybrid mode, with branch deployments for pull requests. This simulator runs everything in one process, one run at a time; the questions here are about the real architecture.
Steps in the warehouse pool share one concurrency limit across runs.