Understand histogram percentile data generated with TFB #23
kusumachalasani
started this conversation in
General
Replies: 1 comment 1 reply
-
@kusumachalasani do you have a link to the '/db' endpoint src? |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Benchmark : TFB
How the benchmark runs? :
Benchmark is run for 3 iterations. Each iteration is run for 6 minutes in total in which first 3 mins is considered as warmup and the next 3 mins for measurement. Before the start of each iteration, the pod is deployed again.
TFB benchmark is run with 1 instance of application server (Quarkus) in openshift and standalone postgres DB on RHEL.
Queries used:
Below queries are used to collect the percentile data of application server are below. Because it is a single instance, I collected the data using histogram_quantile and the basic query to understand how the data looks like.
percentile data:
Below is the percentile data collected using query1 and query2 and the max server_request time using query3 for 3 iterations.
All the data mentioned below is in seconds.
I picked the max_value during the run for the max_server_request ( query3 ) as this value resets for every 2 mins.
From the above data, 95p and 99.999p from query1 and query2 are almost at the same level. But 100th percentile in histogram gives a very high data point.
From the multiple runs I did, atleast one iteration shows 30 secs as 100th percentile with query2.
The max server_request time is < 1 sec in all these cases.
How can histogram 100th percentile data (query2) gives 30sec if the max_server_request time is < 1 sec ?
Any thoughts of what would have caused this ?
@dinogun @johnaohara @ebullient
Beta Was this translation helpful? Give feedback.
All reactions