Self-Assessment Quiz: Apache Airflow

Twenty questions. Aim for 16 or more. Questions 5, 6, and 17 are the three that decide whether your pipelines survive a retry.

Question 1

Cron stops being adequate at roughly:

• A. Ten tasks
• B. Three dependent steps, or when a step's duration varies enough that a fixed offset is a guess
• C. Any data pipeline
• D. When you need retries

Question 2

A DAG file is:

• A. The running pipeline
• B. A Python file parsed repeatedly by the scheduler — every 30 seconds by default
• C. Compiled once at deploy
• D. Executed by a worker

Question 3

Therefore, module-level code in a DAG file:

• A. Runs once
• B. Runs on the scheduler, every parse, forever
• C. Is ignored
• D. Runs on a worker

Question 4

A DAG with schedule="@daily" and a data interval of 2026-03-17 actually runs:

• A. At the start of 2026-03-17
• B. Just after midnight on 2026-03-18 — at the END of the interval
• C. Whenever a worker is free
• D. At noon on 2026-03-17

Question 5

datetime.now() inside a task means:

• A. Nothing, it works fine
• B. A retry processes a different window, and a rerun writes today's data into a historical partition
• C. The task runs slower
• D. Timezones are wrong

Question 6

data_interval_start and data_interval_end are:

• A. Recomputed on each attempt
• B. Stable across every retry and every rerun, forever
• C. Only available in Airflow 3
• D. The same as datetime.now()

Question 7

catchup=True with a start_date two years old:

• A. Runs once
• B. Schedules 730 runs the moment you deploy it
• C. Is required
• D. Only affects the UI

Question 8

Which failure makes non-idempotency a certainty rather than a risk?

• A. A syntax error
• B. A write that succeeded and lost its acknowledgment, followed by a retry
• C. A slow query
• D. A full disk

Question 9

The right size for a task is:

• A. As small as possible
• B. One operator
• C. The smallest unit you would want to retry independently
• D. Under 100 lines

Question 10

A sensor with deferrable=False:

• A. Is faster
• B. Holds a worker slot for its entire wait
• C. Cannot time out
• D. Requires a triggerer

Question 11

An ExternalTaskSensor with no execution_delta:

• A. Waits for any run of the other DAG
• B. Looks for a run at the same logical date — and if the schedules differ, there is none
• C. Waits forever by design
• D. Is deprecated

Question 12

Datasets invert the dependency because:

• A. They are faster
• B. The producer declares what it produces and the consumer declares what it consumes; the schedule is derived
• C. They avoid the metadata database
• D. They only work cross-DAG

Question 13

Which limit is Chapter 20's lock against a backfill racing the nightly run?

• A. parallelism
• B. max_active_tasks
• C. A pool, plus max_active_runs=1
• D. retries

Question 14

In the §24.8 starvation incident, what failed?

• A. The backfill
• B. The nightly DAG
• C. Nothing — every task succeeded and the 6am SLA was missed by 41 minutes
• D. The scheduler

Question 15

Why would an Airflow task-level SLA miss not have fired there?

• A. SLAs were disabled
• B. No individual task was slow — the DAG was slow because it started late
• C. The SLA was too generous
• D. SLAs only apply to sensors

Question 16

XCom should carry:

• A. Whatever fits
• B. References — a path, an ID, a count. Anything you would not put in a log line does not go in it
• C. DataFrames, via a custom backend
• D. Nothing

Question 17

A task returning an 8.4 MB DataFrame twenty-four times a day for fourteen months produces:

• A. Slow tasks
• B. 83.9 GB in the shared metadata database, presenting as Airflow being broken
• C. A disk warning on the worker
• D. Nothing, XCom is compressed

Question 18

The test for whether logic belongs in the orchestrator:

• A. Is it Python?
• B. Could this run correctly outside Airflow?
• C. Is it under 50 lines?
• D. Does it need a connection?

Question 19

A scheduler that is not running produces:

• A. A failure alert
• B. No events at all — nothing runs and nothing fails
• C. Queued tasks
• D. An SLA miss

Question 20

The fix for that class of failure is:

• A. More alerts
• B. A heartbeat — something that must arrive, so its absence is an event — monitored from outside Airflow
• C. A second scheduler
• D. Longer retries

Answer Key

1. B — §24.1. And be honest about the cost: Kestrel's managed Airflow is $310 a month before any task runs.

2. B — §24.2.

3. B — §24.2. time python your_dag.py should be under a second.

4. B — §24.3. A run that processes a day's data cannot start until the day is over. The design is confusing and correct.

5. B — §24.3, and Case Study 1. 71,012 August rows in March partitions, $1,994,727 overstated.

6. B — §24.3. They are properties of the run, not of the moment — which is the whole point of the date model.

7. B — §24.4. Right for independent time windows; wrong for anything maintaining current state.

8. B — §24.5. It is the normal failure mode of a network, not an edge case.

9. C — §24.6. The unit of failure is the unit of design — and splitting costs a scheduling round trip per boundary.

10. B — §24.7. Sixteen sensors × 41 minutes = 656 slot-minutes a night out of 32 slots.

11. B — §24.7. It then times out reporting an upstream failure that did not happen. An offset between two cron expressions is a dependency waiting to be wrong.

12. B — §24.7. The same idea as ref() and TaskFlow's call structure, a third time.

13. C — §24.8. Both halves: the pool serializes across DAGs, max_active_runs=1 stops a DAG joining itself.

14. C — §24.8. Chapter 23 §23.3's distinction, appearing in the orchestrator.

15. B — §24.8. A different measurement — put the SLA on the DAG's completion time.

16. B — §24.9.

17. B — Case Study 2. And airflow db clean had never run, because it is not on by default and is in no getting-started guide.

18. B — §24.10. A @task should be four lines calling a library function.

19. B — §24.12 and Case Study 2. The absence of activity looks like an absence of work.

20. B — Case Study 2. Four lines of DAG, and the monitor must be outside Airflow — an alert Airflow sends you when Airflow is down is not an alert.

Topic map