Airflow resource utilization spikes

Question

We have 32 V-CPUs with 28 GB ram with Local Executor but still airflow is utilizing all the resources and this is resulting in over-utilization of resources which ultimately breaks the system execution.

Below is the output for ps -aux ordered by memory usage.

   PID %CPU %MEM    VSZ   RSS TTY      STAT START   TIME COMMAND
  1336  3.5  0.9 1600620 271644 ?      Ss   Feb18  23:41 /usr/bin/python /usr/local/bin/airflow webs
  9434 32.3  0.9 1835796 267844 ?      Sl   03:09   0:31 [ready] gunicorn: worker [airflow-webserver
 10043  9.1  0.9 1835796 267844 ?      Sl   03:05   0:33 [ready] gunicorn: worker [airflow-webserver
 25397 17.4  0.9 1835796 267844 ?      Sl   03:08   0:30 [ready] gunicorn: worker [airflow-webserver
 30680 13.0  0.9 1835796 267844 ?      Sl   03:06   0:36 [ready] gunicorn: worker [airflow-webserver
 28434 60.5  0.5 1720548 152380 ?      Rl   03:10   0:12 gunicorn: worker [airflow-webserver]       
 20202  2.2  0.3 1671280 111316 ?      Sl   03:07   0:04 /usr/bin/python /usr/local/bin/airflow run 
 14353  1.9  0.3 1671484 111208 ?      Sl   03:07   0:04 /usr/bin/python /usr/local/bin/airflow run 
 14497  1.8  0.3 1671480 111192 ?      Sl   03:07   0:03 /usr/bin/python /usr/local/bin/airflow run 
 25170  2.0  0.3 1671024 110964 ?      Sl   03:08   0:03 /usr/bin/python /usr/local/bin/airflow run 
 21887  1.8  0.3 1670692 110672 ?      Sl   03:07   0:03 /usr/bin/python /usr/local/bin/airflow run 
  5211  4.7  0.3 1670488 110456 ?      Sl   03:09   0:05 /usr/bin/python /usr/local/bin/airflow run 
  8819  4.9  0.3 1670140 110264 ?      Sl   03:09   0:04 /usr/bin/python /usr/local/bin/airflow run 
  6034  3.9  0.3 1670324 110080 ?      Sl   03:09   0:04 /usr/bin/python /usr/local/bin/airflow run 
  8817  4.6  0.3 1670136 110044 ?      Sl   03:09   0:04 /usr/bin/python /usr/local/bin/airflow run 
  8829  4.0  0.3 1670076 110012 ?      Sl   03:09   0:04 /usr/bin/python /usr/local/bin/airflow run 
 14349  1.6  0.3 1670360 109988 ?      Sl   03:07   0:03 /usr/bin/python /usr/local/bin/airflow run 
  8815  3.5  0.3 1670140 109984 ?      Sl   03:09   0:03 /usr/bin/python /usr/local/bin/airflow run 
  8917  4.2  0.3 1669980 109980 ?      Sl   03:09   0:04 /usr/bin/python /usr/local/bin/airflow run 

From the RSS field we can see that the RAM being utilized for web-server is more than 10 GB and per task an average of 1 GB is being used.

The tasks are just for monitoring an end point of a rest API.

Below is the Airflow Configuration file

[core]
# The home folder for airflow, default is ~/airflow
airflow_home = /airflow
# The folder where your airflow pipelines live, most likely a
# subfolder in a code repository
# This path must be absolute
dags_folder = /airflow/dags
# The folder where airflow should store its log files
# This path must be absolute
base_log_folder = /airflow/logs/
# Airflow can store logs remotely in AWS S3 or Google Cloud Storage. Users
# must supply an Airflow connection id that provides access to the storage
# location.
remote_logging = True
remote_log_conn_id = datalake_gcp_connection
encrypt_s3_logs = False
# Logging level
logging_level = INFO
# Logging class
# Specify the class that will specify the logging configuration
# This class has to be on the python classpath
logging_config_class = log_config.LOGGING_CONFIG
# Log format
log_format = [%%(asctime)s] {%%(filename)s:%%(lineno)d} %%(levelname)s - %%(message)s
simple_log_format = %%(asctime)s %%(levelname)s - %%(message)s
# The executor class that airflow should use. Choices include
# SequentialExecutor, LocalExecutor, CeleryExecutor, DaskExecutor
executor = LocalExecutor
# The SqlAlchemy connection string to the metadata database.
# SqlAlchemy supports many different database engine, more information
# their website
sql_alchemy_conn = mysql://user:pass@127.0.0.1/airflow_db
# The SqlAlchemy pool size is the maximum number of database connections
# in the pool.
sql_alchemy_pool_size = 400

# The SqlAlchemy pool recycle is the number of seconds a connection
# can be idle in the pool before it is invalidated. This config does
# not apply to sqlite.
sql_alchemy_pool_recycle = 3000

# The amount of parallelism = 32
# the max number of task instances that should run simultaneously
# on this airflow installation
parallelism = 64

# The number of task instances allowed to run concurrently by the scheduler
dag_concurrency = 32

# 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 = 400

# The maximum number of active DAG runs per DAG
max_active_runs_per_dag = 16

# Whether to load the examples that ship with Airflow. It's good to
# get started, but you probably want to set this to False in a production
# environment
load_examples = False

# Where your Airflow plugins are stored
plugins_folder = /airflow/plugins

# Secret key to save connection passwords in the db
fernet_key = <FERNET KEY>

# Whether to disable pickling dags
donot_pickle = False

# How long before timing out a python file import while filling the DagBag
dagbag_import_timeout = 120

# The class to use for running task instances in a subprocess
task_runner = BashTaskRunner

# If set, tasks without a `run_as_user` argument will be run with this user
# Can be used to de-elevate a sudo user running Airflow when executing tasks
default_impersonation =

# What security module to use (for example kerberos):
security =

# Turn unit test mode on (overwrites many configuration options with test
# values at runtime)
unit_test_mode = False

# Name of handler to read task instance logs.
# Default to use file task handler.
task_log_reader = gcs.task

# Whether to enable pickling for xcom (note that this is insecure and allows for
# RCE exploits). This will be deprecated in Airflow 2.0 (be forced to False).
enable_xcom_pickling = True

# When a task is killed forcefully, this is the amount of time in seconds that
# it has to cleanup after it is sent a SIGTERM, before it is SIGKILLED
killed_task_cleanup_time = 60

[cli]
# In what way should the cli access the API. The LocalClient will use the
# database directly, while the json_client will use the api running on the
# webserver
api_client = airflow.api.client.json_client
endpoint_url = http://0.0.0.0:8080

[api]
# How to authenticate users of the API
auth_backend = airflow.api.auth.backend.default

[operators]
# The default owner assigned to each new operator, unless
# provided explicitly or passed via `default_args`
default_owner = Airflow
default_cpus = 1
default_ram = 125
default_disk = 125
default_gpus = 0


[webserver]
# The base url of your website as airflow cannot guess what domain or
# cname you are using. This is used in automated emails that
# airflow sends to point links to the right web server
base_url = http://localhost:8080

authenticate = False
auth_backend = airflow.contrib.auth.backends.password_auth

# The ip specified when starting the web server
web_server_host = 0.0.0.0

# The port on which to run the web server
web_server_port = 8080

# Paths to the SSL certificate and key for the web server. When both are
# provided SSL will be enabled. This does not change the web server port.
web_server_ssl_cert =
web_server_ssl_key =

# Number of seconds the gunicorn webserver waits before timing out on a worker
web_server_worker_timeout = 120

# Number of workers to refresh at a time. When set to 0, worker refresh is
# disabled. When nonzero, airflow periodically refreshes webserver workers by
# bringing up new ones and killing old ones.
worker_refresh_batch_size = 1

# Number of seconds to wait before refreshing a batch of workers.
worker_refresh_interval = 30

# Secret key used to run your flask app
secret_key = temporary_key

# Number of workers to run the Gunicorn web server
workers = 4

# The worker class gunicorn should use. Choices include
# sync (default), eventlet, gevent
worker_class = sync

# Log files for the gunicorn webserver. '-' means log to stderr.
access_logfile = -
error_logfile = -

# Expose the configuration file in the web server
expose_config = False

# Set to true to turn on authentication:
# http://pythonhosted.org/airflow/security.html#web-authentication
#authenticate = False

# Filter the list of dags by owner name (requires authentication to be enabled)
filter_by_owner = False

# Filtering mode. Choices include user (default) and ldapgroup.
# Ldap group filtering requires using the ldap backend
#
# Note that the ldap server needs the "memberOf" overlay to be set up
# in order to user the ldapgroup mode.
owner_mode = user

# Default DAG view.  Valid values are:
# tree, graph, duration, gantt, landing_times
dag_default_view = graph

# Default DAG orientation. Valid values are:
# LR (Left->Right), TB (Top->Bottom), RL (Right->Left), BT (Bottom->Top)
dag_orientation = LR

# Puts the webserver in demonstration mode; blurs the names of Operators for
# privacy.
demo_mode = False

# The amount of time (in secs) webserver will wait for initial handshake
# while fetching logs from other worker machine
log_fetch_timeout_sec = 5

# By default, the webserver shows paused DAGs. Flip this to hide paused
# DAGs by default
hide_paused_dags_by_default = True

# Consistent page size across all listing views in the UI
page_size = 40

[email]
email_backend = airflow.utils.email.send_email_smtp


[smtp]
# If you want airflow to send emails on retries, failure, and you want to use
# the airflow.utils.email.send_email_smtp function, you have to configure an
# smtp server here
smtp_host = smtp.gmail.com
smtp_starttls = True
smtp_ssl = False
# Uncomment and set the user/pass settings if you want to use SMTP AUTH
#smtp_user = airflow
#smtp_password = airflow
smtp_port = 25
smtp_mail_from = airflow@example.com


[celery]
# This section only applies if you are using the CeleryExecutor in
# [core] section above

# The app name that will be used by celery
celery_app_name = airflow.executors.celery_executor

# The concurrency that will be used when starting workers with the
# "airflow worker" command. This defines the number of task instances that
# a worker will take, so size up your workers based on the resources on
# your worker box and the nature of your tasks
celeryd_concurrency = 16

# When you start an airflow worker, airflow starts a tiny web server
# subprocess to serve the workers local log files to the airflow main
# web server, who then builds pages and sends them to users. This defines
# the port on which the logs are served. It needs to be unused, and open
# visible from the main web server to connect into the workers.
worker_log_server_port = 8793

# The Celery broker URL. Celery supports RabbitMQ, Redis and experimentally
# a sqlalchemy database. Refer to the Celery documentation for more
# information.
broker_url = sqla+mysql://user:pass@127.0.0.1/airflow_db

# Another key Celery setting
celery_result_backend = db+mysql://user:pass@127.0.0.1/airflow_db

# Celery Flower is a sweet UI for Celery. Airflow has a shortcut to start
# it `airflow flower`. This defines the IP that Celery Flower runs on
flower_host = 0.0.0.0

# This defines the port that Celery Flower runs on
flower_port = 5555

# Default queue that tasks get assigned to and that worker listen on.
default_queue = default

# Import path for celery configuration options
celery_config_options = airflow.config_templates.default_celery.DEFAULT_CELERY_CONFIG

[dask]
# This section only applies if you are using the DaskExecutor in
# [core] section above

# The IP address and port of the Dask cluster's scheduler.
cluster_address = 127.0.0.1:8786


[scheduler]
# Task instances listen for external kill signal (when you clear tasks
# from the CLI or the UI), this defines the frequency at which they should
# listen (in seconds).
job_heartbeat_sec = 20

# The scheduler constantly tries to trigger new tasks (look at the
# scheduler section in the docs for more information). This defines
# how often the scheduler should run (in seconds).
scheduler_heartbeat_sec = 60

# after how much time should the scheduler terminate in seconds
# -1 indicates to run continuously (see also num_runs)
run_duration = -1

# after how much time a new DAGs should be picked up from the filesystem
min_file_process_interval = 5

dag_dir_list_interval = 300

# How often should stats be printed to the logs
print_stats_interval = 30

child_process_log_directory = /airflow/logs/scheduler

# Local task jobs periodically heartbeat to the DB. If the job has
# not heartbeat in this many seconds, the scheduler will mark the
# associated task instance as failed and will re-schedule the task.
scheduler_zombie_task_threshold = 300

# Turn off scheduler catchup by setting this to False.
# Default behavior is unchanged and
# Command Line Backfills still work, but the scheduler
# will not do scheduler catchup if this is False,
# however it can be set on a per DAG basis in the
# DAG definition (catchup)
catchup_by_default = False

# This changes the batch size of queries in the scheduling main loop.
# This depends on query length limits and how long you are willing to hold locks.
# 0 for no limit
max_tis_per_query = 256

# Statsd (https://github.com/etsy/statsd) integration settings
statsd_on = False
statsd_host = localhost
statsd_port = 8125
statsd_prefix = airflow

# The scheduler can run multiple threads in parallel to schedule dags.
# This defines how many threads will run.
max_threads = 12

authenticate = False

[ldap]
# set this to ldaps://<your.ldap.server>:<port>
uri =
user_filter = objectClass=*
user_name_attr = uid
group_member_attr = memberOf
superuser_filter =
data_profiler_filter =
bind_user = cn=Manager,dc=example,dc=com
bind_password = insecure
basedn = dc=example,dc=com
cacert = /etc/ca/ldap_ca.crt
search_scope = LEVEL

[mesos]
# Mesos master address which MesosExecutor will connect to.
master = localhost:5050

# The framework name which Airflow scheduler will register itself as on mesos
framework_name = Airflow

# Number of cpu cores required for running one task instance using
# 'airflow run <dag_id> <task_id> <execution_date> --local -p <pickle_id>'
# command on a mesos slave
task_cpu = 1

# Memory in MB required for running one task instance using
# 'airflow run <dag_id> <task_id> <execution_date> --local -p <pickle_id>'
# command on a mesos slave
task_memory = 256

# Enable framework checkpointing for mesos
# See http://mesos.apache.org/documentation/latest/slave-recovery/
checkpoint = False

# Failover timeout in milliseconds.
# When checkpointing is enabled and this option is set, Mesos waits
# until the configured timeout for
# the MesosExecutor framework to re-register after a failover. Mesos
# shuts down running tasks if the
# MesosExecutor framework fails to re-register within this timeframe.
# failover_timeout = 604800

# Enable framework authentication for mesos
# See http://mesos.apache.org/documentation/latest/configuration/
authenticate = False

# Mesos credentials, if authentication is enabled
# default_principal = admin
# default_secret = admin


[kerberos]
ccache = /tmp/airflow_krb5_ccache
# gets augmented with fqdn
principal = airflow
reinit_frequency = 3600
kinit_path = kinit
keytab = airflow.keytab


[github_enterprise]
api_rev = v3


[admin]
# UI to hide sensitive variable fields when set to True
hide_sensitive_variable_fields = True

What are we doing wrong here?

Answer 1

A recommended method is to set the CPUQuota of Airflow to max 80%. This will ensure that Airflow process does not eat up all the CPU resources which sometimes cause the system to hang.

You can use a ready-made AMI (namely, LightningFLow) from AWS Marketplace which is pre-configured with the recommended configurations.

Note: LightningFlow also comes pre-integrated with all required libraries, Livy, custom operators, and local Spark cluster.

Link for AWS Marketplace: https://aws.amazon.com/marketplace/pp/Lightning-Analytics-Inc-LightningFlow-Integrated-o/B084BSD66V

Answer 2

The size shown in RSS field is in KB. The first process is using about 265 MB, not something over 10 GB.

The MEM field shows the memory usage in percentage, not GB. 0.9% of 28 GB is 252 MB. You can see stats about memory with the free command.

See http://man7.org/linux/man-pages/man1/ps.1.html. In short, it's not airflow over utilising resources that's breaking your system.