A benchmark of the Java Kafka client (kafka-clients 4.3.1, Temurin 25) against the Rust client (rdkafka 0.39.0, librdkafka 2.12.1): same host, same single Kafka 4.3.1 broker, byte-identical payloads, everything in Docker.
- The story (what the first run showed, how each gap was diagnosed, what fixed it): the blog post
- The numbers and caveats:
docs/ANALYSIS.md, with the diagnoses indocs/diagnosis.md(Rust) anddocs/java-diagnosis.md(Java) - The fairness rules both harnesses implement:
CONTRACT.md
Each client tuned, 1 KB messages, 6 partitions, one client instance, medians of 3 reps on an EC2 c6i.8xlarge.
| Java | Rust | |
|---|---|---|
Producer, each at its best batch.size |
1.05M msg/s (1.12M with send.buffer.bytes=-1) |
1.00M msg/s |
| Consumer, 1 KB / 10 KB / 100 B | 1.21M / 137k / 5.75M msg/s | 1.35M / 305k / 2.10M msg/s |
End-to-end p99 at 50k msg/s, linger.ms=0 |
0.93 ms | 1.93 ms (max.in.flight=5) |
| Peak RSS / startup | about 2.4 GB / 260 ms | 9 MB to 1.2 GB / 0.8 ms |
Most of the first run's gaps were defaults, not languages:
Needs only a docker CLI whose active context points at the benchmark host (docker context use <your-context>). Builds, runs and analysis happen in containers on that host; runs are detached, so the local machine can disconnect.
./kbench.sh build # build kbench-java, kbench-rust, kbench-runner
./kbench.sh run --dry-run # plan, run count and ETA
./kbench.sh run --reps 3 --out full-1 # run everything; the report is generated at the end
./kbench.sh status # or: logs, ls
./kbench.sh fetch full-1 # copy results to ./results/full-1
./kbench.sh stop # stop the runner, remove its clients and topics
bash infra/down.sh # remove the broker and kbench-net
Useful options: --reps N, --scale F, --only REGEX, --clients LIST, --rust-config-profile matched|native, --skip-scaling, --resume. Regenerate a report with ./kbench.sh report <dir>.
--clients variant |
What it is |
|---|---|
java |
Java client, contract configuration |
rust |
Rust client, BaseConsumer poll loop |
rust-tuned |
Rust consumer with the batch API and fetch.queue.backoff.ms=10 |
rust-lowlat |
Rust producer with max.in.flight=5; e2e only, not in the default list |
--rust-config-profile matched (default) forces Java-like librdkafka producer settings; native leaves librdkafka defaults. The published runs used native.
| Path | Contents |
|---|---|
java-bench/, rust-bench/ |
The two harnesses |
bench/ |
Orchestrator (run.py), scenario matrix (matrix.py), report (report.py), charts (story_plots.py, hero_svg.py) |
infra/ |
Broker compose file and helper scripts |
runner/, kbench.sh |
Runner image and the local launcher |
docs/ |
Analysis and diagnosis write-ups |
Each result directory holds report.md (tables, charts, findings, caveats), summary.csv (one row per scenario, client and role), raw.jsonl (every run with its exact command and effective config), and host and environment details.
