-
Type:
Task
-
Resolution: Done
-
Priority:
Unknown
-
None
-
Affects Version/s: None
-
Component/s: None
-
None
One of the goals of new client backpressure overload retry behavior is to withstand temporary load spikes on the server. The operation burst workload (PERF-7190) shows that 3 retries is insufficient at reducing the error rate during brief overload periods. This can be attributed to the short effective retry time window of 3 retries and initial delay 100ms which is only 100ms + 200ms + 400ms = 700ms. With 5 retries the time window increases to 100ms + 200ms + 400ms + 800ms + 1600ms = 3100ms.
Without any retries the workload encounters ~8000 overload errors. With 3 max retries the workload encounters ~500 errors. With 5 max retries the workload encounters ~5 errors.
The design has been updated to propose increasing the retry limit to 5.