Extended workMicros bench sweep: 10–100ms (the production regime)

[eschizoid]

The bench sweep #148 asked for is on main. We used to stop at workMicros=1000 (1ms), which is below where most production Kafka consumers actually sit. HTTP/DB hops are 10–100ms. Chuck Larrieu Casias (Conduktor) flagged the gap on LinkedIn; this sweep covers it.

Full writeup: benchmarks/results/2026-06-07.md. Raw JSON next to it.

100ms per record (≈ a normal HTTP call)

Runtime	ops/s	Ordering
Raw KafkaConsumer + virtual threads	166,553	none
KPipe PARALLEL	45,174	none
KPipe KEY_ORDERED	5,210	per-key
Kafka Share Consumer (KIP-932)	2,595	none
Confluent PC KEY	968	per-key
Confluent PC UNORDERED	960	none
Confluent PC PARTITION	— (DNF)	per-partition
Reactor Kafka	— (DNF)	none

kpipe leads the next framework (CPC UNORDERED) at this cell by ~47×. Reactor and CPC PARTITION couldn't drain 25k records inside the wait window. Read those DNFs as architecture, not flakiness: CPC PARTITION caps in-flight at partitions / workMicros ≈ 80 ops/s with 8 partitions and 100ms work, and Reactor collapses under Schedulers.parallel() blocking work.

The cross-library number I find most interesting is KPipe KEY_ORDERED vs Confluent PC KEY. Same per-key FIFO guarantee, different implementations (VT per key queue vs platform-thread pool with key gating). 5,210 vs 968 at 100ms, ~5.4× lead. Bigger at the lighter cells.

The honest caveat

Raw KafkaConsumer + a virtual-thread executor beats kpipe by 3-4× at every cell. 561k vs 372k at 0µs, 166k vs 45k at 100ms. That's the Loom floor without a framework, and it's the real ceiling for everything else in this table. The kpipe pitch isn't "fastest". It's framework conveniences (lowest-pending-offset commits, lag/in-flight backpressure, DLQ, batching, tracing) for ~3× of the bare-Loom number. I keep that disclaimer next to every kpipe number I quote, and you should too.

A few things to know before you quote a number

• singleThread was excluded. 25k records × 100ms sequentially is ~42 minutes per iteration; the full sweep would have been ~11 hours just for that one cell. The 0/100/1000 numbers in 2026-05-30.md establish the per-partition floor and the rest is trivially worse.
• CPC is still the deprecated 0.5.3.0 artifact. I'll rerun once astubbs/parallel-consumer publishes to Maven Central. That fork is the live one per Confluent's CSID team.
• Intel laptop, Docker-loopback broker. Absolute records/s will move on Apple Silicon and real server hardware. The within-table orderings are the part that's meaningful.

Refs: #148, PR #153.

Thanks @chuck-alt-delete for the original push.
(cc @igfasouza, @astubbs)