Slow running Airflow 1.10.2 ETL when using ExternalTaskSensor for DAG task dependency? - celery

Question

I have two DAGs that I need to run with Airflow 1.10.2 + the CeleryExecutor. The first DAG (DAG1) is a long-running data load from s3 into Redshift (3+ hours). My second DAG (DAG2) performs computations on data loaded by DAG1. I want to include an ExternalTaskSensor in DAG2 so that the computations are reliably performed after the data loads. Theoretically so simple!
I can successfully get DAG2 to wait for DAG1 to complete by ensuring both DAGs are scheduled to start at the same time (schedule="0 8 * * *" for both DAGs) and DAG2 is dependent on the final task in DAG1. But I'm seeing a massive delay in our ETL on DAG1 when I introduce the sensor. I at first though it was because my original implementation used mode="poke" which I understand locks a worker. However, even when I changed this to mode="reschedule" as I read in the docs https://airflow.readthedocs.io/en/stable/_modules/airflow/sensors/base_sensor_operator.html I still see a massive ETL delay.
I'm using the ExternalTaskSensor code below in DAG2:
wait_for_data_load = ExternalTaskSensor(
dag=dag,
task_id="wait_for_data_load",
external_dag_id="dag1",
external_task_id="dag1_final_task_id",
mode="reschedule",
poke_interval=1800, # check every 30 min
timeout=43200, # timeout after 12 hours (catch delayed data load runs)
soft_fail=False # if the task fails, we assume a failure
)
If the code were working properly, I'd expect the sensor to perform a quick check whether DAG1 had finished and, if not, reschedule for 30 min time as defined by the poke_interval, causing no delay to DAG1 ETL. If DAG1 fails to complete after 12 hours, then DAG2 would stop poking and fail.
Instead, I'm getting frequent errors for each of the tasks in DAG1 saying (for example) Executor reports task instance <TaskInstance: dag1.data_table_temp_redshift_load 2019-05-20 08:00:00+00:00 [queued]> finished (failed) although the task says its queued. Was the task killed externally? even though the tasks are completing successfully (with some delay). Just before this error is sent, I see a line in our Sentry logs saying Executor reports dag1.data_table_temp_redshift_load execution_date=2019-05-20 08:00:00+00:00 as failed for try_number 1 though (again) I can see the task succeeded.
The logs on DAG2 are also looking a bit strange. I'm seeing repeated attempts logged at the same time intervals like the excerpt below:
--------------------------------------------------------------------------------
Starting attempt 1 of 4
--------------------------------------------------------------------------------
[2019-05-21 08:01:48,417] {{models.py:1593}} INFO - Executing <Task(ExternalTaskSensor): wait_for_data_load> on 2019-05-20T08:00:00+00:00
[2019-05-21 08:01:48,419] {{base_task_runner.py:118}} INFO - Running: ['bash', '-c', 'airflow run dag2 wait_for_data_load 2019-05-20T08:00:00+00:00 --job_id 572075 --raw -sd DAGS_FOLDER/dag2.py --cfg_path /tmp/tmp4g2_27c7']
[2019-05-21 08:02:02,543] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:02,542] {{settings.py:174}} INFO - settings.configure_orm(): Using pool settings. pool_size=5, pool_recycle=1800, pid=28219
[2019-05-21 08:02:12,000] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:11,996] {{__init__.py:51}} INFO - Using executor CeleryExecutor
[2019-05-21 08:02:15,840] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:15,827] {{models.py:273}} INFO - Filling up the DagBag from /usr/local/airflow/dags/dag2.py
[2019-05-21 08:02:16,746] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:16,745] {{dag2.py:40}} INFO - Waiting for the dag1_final_task_id operator to complete in the dag1 DAG
[2019-05-21 08:02:17,199] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:17,198] {{cli.py:520}} INFO - Running <TaskInstance: dag1. wait_for_data_load 2019-05-20T08:00:00+00:00 [running]> on host 11d93b0b0c2d
[2019-05-21 08:02:17,708] {{external_task_sensor.py:91}} INFO - Poking for dag1. dag1_final_task_id on 2019-05-20T08:00:00+00:00 ...
[2019-05-21 08:02:17,890] {{models.py:1784}} INFO - Rescheduling task, marking task as UP_FOR_RESCHEDULE
[2019-05-21 08:02:17,892] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load /usr/local/lib/python3.6/site-packages/requests/__init__.py:91: RequestsDependencyWarning: urllib3 (1.25.2) or chardet (3.0.4) doesn't match a supported version!
[2019-05-21 08:02:17,893] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load RequestsDependencyWarning)
[2019-05-21 08:02:17,893] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load /usr/local/lib/python3.6/site-packages/psycopg2/__init__.py:144: UserWarning: The psycopg2 wheel package will be renamed from release 2.8; in order to keep installing from binary please use "pip install psycopg2-binary" instead. For details see: <http://initd.org/psycopg/docs/install.html#binary-install-from-pypi>.
[2019-05-21 08:02:17,894] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load """)
[2019-05-21 08:02:22,597] {{logging_mixin.py:95}} INFO - [2019-05-21 08:02:22,589] {{jobs.py:2527}} INFO - Task exited with return code 0
[2019-05-21 08:01:48,125] {{models.py:1359}} INFO - Dependencies all met for <TaskInstance: dag2. wait_for_data_load 2019-05-20T08:00:00+00:00 [queued]>
[2019-05-21 08:01:48,311] {{models.py:1359}} INFO - Dependencies all met for <TaskInstance: dag2. wait_for_data_load 2019-05-20T08:00:00+00:00 [queued]>
[2019-05-21 08:01:48,311] {{models.py:1571}} INFO -
--------------------------------------------------------------------------------
Starting attempt 1 of 4
--------------------------------------------------------------------------------
[2019-05-21 08:01:48,417] {{models.py:1593}} INFO - Executing <Task(ExternalTaskSensor): wait_for_data_load> on 2019-05-20T08:00:00+00:00
[2019-05-21 08:01:48,419] {{base_task_runner.py:118}} INFO - Running: ['bash', '-c', 'airflow run dag2 wait_for_data_load 2019-05-20T08:00:00+00:00 --job_id 572075 --raw -sd DAGS_FOLDER/dag2.py --cfg_path /tmp/tmp4g2_27c7']
[2019-05-21 08:02:02,543] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:02,542] {{settings.py:174}} INFO - settings.configure_orm(): Using pool settings. pool_size=5, pool_recycle=1800, pid=28219
[2019-05-21 08:02:12,000] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:11,996] {{__init__.py:51}} INFO - Using executor CeleryExecutor
[2019-05-21 08:02:15,840] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:15,827] {{models.py:273}} INFO - Filling up the DagBag from /usr/local/airflow/dags/dag2.py
[2019-05-21 08:02:16,746] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:16,745] {{dag2.py:40}} INFO - Waiting for the dag1_final_task_id operator to complete in the dag1 DAG
[2019-05-21 08:02:17,199] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load [2019-05-21 08:02:17,198] {{cli.py:520}} INFO - Running <TaskInstance: dag2.wait_for_data_load 2019-05-20T08:00:00+00:00 [running]> on host 11d93b0b0c2d
[2019-05-21 08:02:17,708] {{external_task_sensor.py:91}} INFO - Poking for dag1.dag1_final_task_id on 2019-05-20T08:00:00+00:00 ...
[2019-05-21 08:02:17,890] {{models.py:1784}} INFO - Rescheduling task, marking task as UP_FOR_RESCHEDULE
[2019-05-21 08:02:17,892] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load /usr/local/lib/python3.6/site-packages/requests/__init__.py:91: RequestsDependencyWarning: urllib3 (1.25.2) or chardet (3.0.4) doesn't match a supported version!
[2019-05-21 08:02:17,893] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load RequestsDependencyWarning)
[2019-05-21 08:02:17,893] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load /usr/local/lib/python3.6/site-packages/psycopg2/__init__.py:144: UserWarning: The psycopg2 wheel package will be renamed from release 2.8; in order to keep installing from binary please use "pip install psycopg2-binary" instead. For details see: <http://initd.org/psycopg/docs/install.html#binary-install-from-pypi>.
[2019-05-21 08:02:17,894] {{base_task_runner.py:101}} INFO - Job 572075: Subtask wait_for_data_load """)
[2019-05-21 08:02:22,597] {{logging_mixin.py:95}} INFO - [2019-05-21 08:02:22,589] {{jobs.py:2527}} INFO - Task exited with return code 0
[2019-05-21 08:33:31,875] {{models.py:1359}} INFO - Dependencies all met for <TaskInstance: dag2.wait_for_data_load 2019-05-20T08:00:00+00:00 [queued]>
[2019-05-21 08:33:31,903] {{models.py:1359}} INFO - Dependencies all met for <TaskInstance: dag2.wait_for_data_load 2019-05-20T08:00:00+00:00 [queued]>
[2019-05-21 08:33:31,903] {{models.py:1571}} INFO -
--------------------------------------------------------------------------------
Starting attempt 1 of 4
--------------------------------------------------------------------------------
Though all logs say Starting attempt 1 of 4, I do see attempts records about every 30 min, but I see multiple logs for each time interval (10+ of the same logs printed for each 30 min interval).
From searching around I see other people are using sensors in production flows https://eng.lyft.com/running-apache-airflow-at-lyft-6e53bb8fccff, which makes me think there's a way around this or I'm implementing something wrong. But I'm also seeing open issues in the airflow project related to this issue, so perhaps there's a deeper issue in the project? I also found a related, but unanswered post here Apache Airflow 1.10.3: Executor reports task instance ??? finished (failed) although the task says its queued. Was the task killed externally?
Also, we are using the following config settings:
# The amount of parallelism as a setting to the executor. This defines
# the max number of task instances that should run simultaneously
# on this airflow installation
parallelism = 32
# The number of task instances allowed to run concurrently by the scheduler
dag_concurrency = 16
# Are DAGs paused by default at creation
dags_are_paused_at_creation = True
# When not using pools, tasks are run in the "default pool",
# whose size is guided by this config element
non_pooled_task_slot_count = 128
# The maximum number of active DAG runs per DAG
max_active_runs_per_dag = 16

Answer 1

These symptoms were actually caused by a call to Variable.set() in the body of DAG1 that DAG2 then used to retrieve DAG1s dynamically generated dag_id. The Variable.set() all was causing an error (discovered in the worker logs). As described here, the scheduler polls the DAG definitions with every heartbeat to update keep DAGs up-to-date. That meant an error with every heartbeat, which caused a large ETL delay.