Read the clock fewer times during message routing

[dferstay]

Previously, we read the system clock twice for each event unless we
switched partitions early due to reaching maxBatchingMessages or
maxBatchingSize.

Now, we read the clock once every maxBatchingMessages / 100 messages.

This improves performance (especially for larger batch sizes) as the
router function is called for every message produced.

A bench test of the default router was added; results are below:

name             old time/op    new time/op    delta
DefaultRouter       106ns ± 0%      16ns ± 1%  -84.94%  (p=0.016 n=4+5)
DefaultRouter-4     106ns ± 0%      16ns ± 0%     ~     (p=0.079 n=4+5)

Signed-off-by: Daniel Ferstay dferstay@splunk.com

Motivation

The default_router reads the system clock twice for every message routed in the production code path. When testing at scale with larger batch sizes the amount of time spent reading the system clock during message routing shows up in pprof profiles.

Modifications

This change modifies the default router to read the system clock once for every maxBatchingMessages / 100 messages routed.
A bench test of the default router was added to verify that the change yields a speedup.

Verifying this change

This change is a trivial rework / code cleanup without any test coverage.

Does this pull request potentially affect one of the following parts:

• Dependencies (does it add or upgrade a dependency): no
• The public API: no
• The schema: no
• The default values of configurations: no
• The wire protocol: no

Documentation

• Does this pull request introduce a new feature? no

[wolfstudy]

@dferstay Can you fix the ci error?