Why Is Airflow Draining CPU? A Step‑by‑Step Diagnosis and Fix

1. Problem Phenomenon

Machine A runs Spark Master, Airflow, Hive, Sqoop and other heavy workloads, resulting in high memory and CPU usage. Over the past three days the CPU stayed above 95% for most of the day, especially after 18:00 when Spark tasks are few.

2. Investigation Process

2.1 Check Applications

At around 09:30 the CPU was high while five SparkSubmit tasks were running; no abnormal applications were found and no single app showed excessive CPU or memory consumption.

2.2 Check Network Connections

netstat revealed many TIME_WAIT connections, mainly to MySQL on hadoop11, exceeding 3,700 connections. The kernel parameters were adjusted in

/etc/sysctl.conf

:

<code>net.ipv4.tcp_tw_reuse = 1
net.ipv4.tcp_tw_recycle = 1</code>

After applying with

/sbin/sysctl -p

, TCP connections normalized but CPU remained high, indicating the issue was not caused by network sockets.

2.3 Check Airflow

Machine A (hadoop16) connects to MySQL on hadoop11 only via Airflow. Airflow runs webserver, scheduler, master, and worker processes, using CeleryExecutor with a parallelism of 16.

(1) Confirm Airflow as the cause

Restarting Airflow temporarily drops CPU usage, which spikes again once Airflow starts, confirming a correlation but not solving the root problem.

(2) Research similar issues

References include a StackOverflow discussion and the Airflow documentation on

min_file_process_interval

.

(3) Apply fix

The Airflow configuration

airflow.cfg

was updated:

<code>min_file_process_interval = 10</code>

After restarting Airflow, CPU usage returned to normal and matched the new file‑scan interval setting.