You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Create a gRPC client that maintains 10 concurrently active unary RPCs to a service that immediately responds with a payload of 1MB, for 60 seconds.
What did you expect to see?
No errors and a normal distribution of response latencies.
What did you see instead?
~1% of requests fail with either DEADLINE_EXCEEDED or CANCELLED
p99 latency of 3 seconds compared to p50 of only 90ms
I believe the long tail latency / timeouts is caused by some call ids always hashing to the end of BinderTransport's ongoingCalls container. Every time space in the flow control window opens up, call ids that appear early in the iteration order gobble it all up. By the time we get to the end of ongoingCalls, flowController.isTransmitWindowFull() is returning true again and Outbound.send() just returns without making any progress.
The text was updated successfully, but these errors were encountered:
What version of gRPC-Java are you using?
head
What is your environment?
Android/Linux
Steps to reproduce the bug
Create a gRPC client that maintains 10 concurrently active unary RPCs to a service that immediately responds with a payload of 1MB, for 60 seconds.
What did you expect to see?
No errors and a normal distribution of response latencies.
What did you see instead?
~1% of requests fail with either DEADLINE_EXCEEDED or CANCELLED
p99 latency of 3 seconds compared to p50 of only 90ms
I believe the long tail latency / timeouts is caused by some call ids always hashing to the end of BinderTransport's ongoingCalls container. Every time space in the flow control window opens up, call ids that appear early in the iteration order gobble it all up. By the time we get to the end of ongoingCalls, flowController.isTransmitWindowFull() is returning true again and Outbound.send() just returns without making any progress.
The text was updated successfully, but these errors were encountered: