roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Prepare for Apache Airflow interview questions on how pipelines are scheduled, run and recovered. Covers DAG design, data intervals, operators and sensors, executors, and safe reruns and backfills.
73 of 73 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 73 new. 0 cut, outside your plan.
Airflow questions come attached to a pipeline. Data engineering, analytics engineering and MLOps loops raise them inside a design discussion or a walk through the stack on your CV, and sometimes hand over a take-home DAG to review. The questions are about time and failure more than syntax: why a daily run on an interval-based schedule starts only after the day it covers has closed, why a database call at the top of a DAG file runs every time that file is parsed, and why a sensor waiting all night in poke mode holds a worker slot that a deferrable operator would give back.
Answers separate on reruns. After a bug fix, someone has to clear a month of runs: does each task overwrite the partition its data interval names, or append a second copy? A task that reads the wall clock instead of its interval cannot be rerun safely, and a backfill without max_active_runs or a pool can flood the warehouse and crowd out tonight's run. Senior candidates are also asked to pick an executor, a standing pool of Celery workers or a pod per task on Kubernetes, and defend it on startup latency, isolation and cost.
The first module leaves Airflow aside. Orchestration concepts are the questions an interviewer can put about any scheduler, and Airflow's answers make sense only against them. It opens on dependency semantics, what an edge between two tasks promises, then turns to time: data intervals fix which slice of data a run owns, and idempotency and backfills depend on that slice, because a rerun is safe only when it rewrites the same one. Retries and alerting and lineage cover pipelines failing in production, and the module closes on where transformation runs, which decides how much work the orchestrator should do itself.
The second module carries those concerns into the tool, in the order a pipeline is written and then operated. DAGs and tasks and operators and sensors are authoring: how a Python file becomes a graph, and whether a task does work, hands it to another system or waits. Scheduling restates the first module's intervals in Airflow's terms, timetables and catchup among them. XComs follow because passing data between tasks is where the ETL question returns: hand on a reference to the output, not the output. The module ends in operation, with executors deciding where tasks physically run and monitoring and retries answering for Airflow what the first module asked in general.
If you already write DAGs at work, the gaps are usually in the first module, not the second. Of the page's controls, study order is the one to move: by topics delivers the concepts whole, ahead of any operator question, and the senior follow-ups on reruns are built from them. Coming from backend or analytics work with no scheduler behind you, take the same sequence for another reason: Airflow's settings stick once the ideas they implement have names. If Airflow is one line on a broader posting and the date is close, trimming importance to must know is enough of a cut; spend the hours on scheduling and reruns. The first module carries over to Dagster or Prefect; the second does not. Nor does the plan cover what the orchestrator runs: the warehouse, Spark, the SQL models a task kicks off. Those, with the modeling and SQL rounds Airflow usually sits beside, are on the [Data Engineer roadmap](/roadmaps/data-engineer); for the transformation layer alone, [dbt](/roadmaps/dbt) goes further.
If you already write DAGs at work, the gaps are usually in the first module, not the second. Of the page's controls, study order is the one to move: by topics delivers the concepts whole, ahead of any operator question, and the senior follow-ups on reruns are built from them. Coming from backend or analytics work with no scheduler behind you, take the same sequence for another reason: Airflow's settings stick once the ideas they implement have names. If Airflow is one line on a broader posting and the date is close, trimming importance to must know is enough of a cut; spend the hours on scheduling and reruns.
The first module carries over to Dagster or Prefect; the second does not. Nor does the plan cover what the orchestrator runs: the warehouse, Spark, the SQL models a task kicks off. Those, with the modeling and SQL rounds Airflow usually sits beside, are on the Data Engineer roadmap; for the transformation layer alone, dbt goes further.