Jmeter is causing 4x latency for high volume of 429s due to backend microservice rate limiting compared to latency for 200 OK response

[mkaranth-tech]

Expected behavior

Expected behaviour should be equal or less latency than 200 OK response.

Actual behavior

We are trying to simulate 3 million requests/min from jmeter to background service. When we don't have rate limiting, jmeter instances are able to generate 3 million request per min. But when we enable rate limiting, throughout is reduced by 1/4th of what we are using for all success (200 OK) responses. When we checked the latency at background service/ingress, latency seems to be fine. But jmeter reports 4 times higher response time. Is there any limitation of 429s on jmeter load generation? Why 429s are reducing throughput at jmeter and increasing the latency at Jmeter?

Steps to reproduce the problem

1. Simulate very volume of test with all sucess response from a service
2. Now repeat the test with enabling rate limiting for that service and configure rate limit to reject 50% of the requests.
3. Now you should be able to see latency of 429s are 4x slower than 200 OK responses.

JMeter Version

latest

Java Version

latest

OS Version

Macos

[vlsi]

I don't think JMeter has 429 handling, however, it looks it might be a reasonable feature to add: retry certain HTTP codes, and support Retry-After header.