GitHub user kenho811-b1 created a discussion: High CPU utilisation caused by frequent updates to dag_run table
### Apache Airflow version Other Airflow 2 version (please specify below) ### If "Other Airflow 2 version" selected, which one? 2.9.1 ### What happened? **Setup** We are self-hosting Airflow via Helm Chart. We use google's CloudSQL as our Airflow metadata store. **Issue** We observe that the below query has been called many times, causes database locks and is causing high CPU utilisation ``` UPDATE dag_run SET last_scheduling_decision=$1::timestamptz, updated_at=$2::timestamptz WHERE dag_run.id = $3 ``` For instance, referring to the below screenshot, in 1 day, 193,362 times has been called. <img width="868" alt="image" src="https://github.com/user-attachments/assets/188d902f-1952-4305-a1f6-c532a0705a87";> **Spec of the CLoudSQL instance** The CLoudSQL instance has 4 vCPUs and 15GB ram. ### What you think should happen instead? Updates to the dag_run on last_scheduling_decision should be reduced (or configurable?) So we can test if database CPU utilisation can be reduced and Database locking problem can be resolved. ### How to reproduce Not easily reproducible. ========== At the time of the UPDATE SQLs, I observe that only 5 instances of DAG runs are running. We have a similar setup in another environment where the other CloudSQL instance also has only 4 vCPUs. In the other environment, however, the number of times the UPDATE SQL is called is smaller. Also no huge database locks are observed in the other environment. ### Operating System linux ### Versions of Apache Airflow Providers _No response_ ### Deployment Official Apache Airflow Helm Chart ### Deployment details _No response_ ### Anything else? _No response_ ### Are you willing to submit PR? - [ ] Yes I am willing to submit a PR! ### Code of Conduct - [X] I agree to follow this project's [Code of Conduct](https://github.com/apache/airflow/blob/main/CODE_OF_CONDUCT.md) GitHub link: https://github.com/apache/airflow/discussions/61003 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
