fishy commented on pull request #2190:
URL: https://github.com/apache/thrift/pull/2190#issuecomment-652550901


   Now bcae3bb (old) vs. this PR (new) has been running for 2hr+, here are the 
2hr charts (still, cyan is new and green is old).
   
   Service A goroutines:
   ![Screenshot 2020-07-01 at 10 15 
01](https://user-images.githubusercontent.com/78183/86273178-b5e54480-bb84-11ea-9daa-268931f1dfb6.png)
   
   Service B goroutines:
   ![Screenshot 2020-07-01 at 10 15 
29](https://user-images.githubusercontent.com/78183/86273217-c3023380-bb84-11ea-9452-41579c317ef2.png)
   
   Service A server side latency p50:
   ![Screenshot 2020-07-01 at 10 16 
37](https://user-images.githubusercontent.com/78183/86273255-ceedf580-bb84-11ea-8846-c89bf6e3e910.png)
   
   Service A server side latency p99:
   ![Screenshot 2020-07-01 at 10 17 
25](https://user-images.githubusercontent.com/78183/86273281-d7dec700-bb84-11ea-9f72-6b94b103d0ee.png)
   
   Service B server side latency p50:
   ![Screenshot 2020-07-01 at 10 18 
09](https://user-images.githubusercontent.com/78183/86273313-e75e1000-bb84-11ea-8494-7daeb8941962.png)
   
   Service B server side latency p99 (this one is weird, the old one actually 
hits the timeout/40ms a lot, probably just bad luck?):
   ![Screenshot 2020-07-01 at 10 19 
09](https://user-images.githubusercontent.com/78183/86273355-fb097680-bb84-11ea-91d9-33725e8c4bc7.png)
   
   A->B client side latency p50:
   ![Screenshot 2020-07-01 at 10 20 
02](https://user-images.githubusercontent.com/78183/86273372-0361b180-bb85-11ea-989a-1730ad8b3dfd.png)
   
   A->B client side latency p99 (note that it's not always old A hitting old B 
and new A hitting new B. service A and B are both in their pool so any A->B 
request can hit any B service with a roughly equal chance):
   ![Screenshot 2020-07-01 at 10 20 
23](https://user-images.githubusercontent.com/78183/86273393-0a88bf80-bb85-11ea-95c7-c8b325b6c013.png)
   
   The most notable gap is on B server side p50 latency, which could be as 
large as 1ms (it could be calculated as ~50% increase in latency, depending on 
how you calculate that). I think we can just document that (probably in 
`lib/go/README.md`?), and also document how to disable this feature if that 1ms 
is important to people there?


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to