Blizzard
fix: Apply timeout on batch basis
What
Applies timeout to the entire batch instead of on a per message basis.
How
The calls to the individual consume share the batch timeout and each call to consume can only use the timeout that the batch has remaining.
Why
The existing behavior is described in detail in this issue.
In short, for d millisecond delay, c count, and b blocking time: b = d * c, given a constant topic rpm 60000 / d.
Given any rpm > 60000 / d: b = c * (60 / rpm) seconds
For example, a 1000ms delay with a batch count of 100 will cause the consumeNum loop to block for up to 100 seconds given a constant topic rpm of 60.
References
Issue: https://github.com/confluentinc/confluent-kafka-javascript/issues/262 PR Introduced: https://github.com/Blizzard/node-rdkafka/pull/34
Test & Review
With c = 100, d = 1000, rpm = 700, expectation is that pre-change we'd block for about 8.6 seconds (it seems like it takes longer in the example, though) before returning the 100 messages.
My testing environment has a 50ms cooldown on calling consume, so after the change we expect a batch of messages to be returned every 1050ms.
Pre-Change
Post-Change
