Closed crioux-stripe closed 3 weeks ago
615 tests +1 605 ✅ +1 8m 7s ⏱️ +6s 142 suites ±0 10 💤 ±0 142 files ±0 0 ❌ ±0
Results for commit 5756e685. ± Comparison against base commit 7f996f85.
:recycle: This comment has been updated with latest results.
On the subject of the pushes for test values: It took me a few cycles to realize that multiple unit tests were sharing the same metrics so I needed to use a unique ID.
I believe the current publish failure is unrelated, it looks like an actor call flaked/timed out on an unrelated piece of code.
Confirmed by committing a no-op comment change that the CI failure was a flaky unrelated test.
We have a job internally that uses the same format as JobSource but for our GRPC client. This happened to drops after deploying this change to that source.
On the subject of the pushes for test values: It took me a few cycles to realize that multiple unit tests were sharing the same metrics so I needed to use a unique ID.
I believe the current publish failure is unrelated, it looks like an actor call flaked/timed out on an unrelated piece of code.
this is very cool! btw did you see any comparison on latency if the buffer size is set to something higher?
We didn't see a ton of latency impact but mostly because the use case is just a little bursty. We have enough capacity in any given 50ms
period but we just may not have the capacity in the 5ms
when the data comes in.
We did see some memory usage increase but the job is also handling a million RPS over three workers so that was to be expected! We also had memory to spare.
Context
The
SseWorkerConnection
uses aDropOperator
with a metric named suffixed withwithBuffer
set in the SseWorkerConnectionFunction class. The class takes abufferSize
parameter but does nothing with it! I've had the illusion that we've been buffering here for darn near a decade!This change is pretty straightforward implementation.
rebatchRequests
simply sendsn
requests into theDropOperator
and continues to stay ahead of the requests made by the downstream by up to n. There is a batching implementation inside of the operator that RxJava uses to implement this.Users could then set the buffer with the:
mantisClient.buffer.size
andworkerClient.buffer.size
properties.Checklist
./gradlew build
compiles code correctly./gradlew test
passes all tests