Open-source streaming · Apache 2.0
Keep your Kafka clients.
Measure a different broker.
krabka implements supported Kafka APIs in safe Rust, with its own KRaft quorum. Inspect its throughput, tail latency and CPU use before you decide where it fits.
Latest measured release: 1.1.5. Published OMB results, with source files and test conditions. Try krabka in your browser
Latest full OMB run · krabka 1.1.5
What the measurements show
Shared page cache
krabka and Kafka use the Linux page cache in this suite. Compare all max-rate workloads, then inspect the differences.
14 workload / replication-factor pairs. 1 measured round per workload. 0 omitted trial across the full suite.
Throughput = the lower of a workload’s publish and consume rates, so a consumer that falls behind does not raise it. CPU efficiency index = throughput ÷ total broker CPU time, relative to Kafka. CPU capture includes startup, warm-up and shutdown. This is a proxy, not CPU cost per delivered message.
Across published releases
See where the gap changed
Compare the replicated 1 KiB workloads with the full max-rate set. Check the full set alongside this group.
Disk preparation can change between runs; each row records SSD trim. These results do not isolate the effect of code changes.
| Version | 1 KiB, RF3 | All max-rate | CPU index |
|---|---|---|---|
| 1.0.1Uncontrolled trim | 75% | 98% | 1.24× |
| 1.0.3Uncontrolled trim | 97% | 99% | 1.12× |
| 1.0.5Trimmed SSD | 103% | 103% | 1.17× |
| 1.1.4Trimmed SSD | 105% | 131% | 1.99× |
| 1.1.5Trimmed SSD | 104% | 131% | 2.01× |
100% = Kafka’s rate. 1× = Kafka’s CPU index. Values are medians of workload ratios, not pooled throughput.
Direct I/O: median 153% of Redpanda’s throughput on max-rate workloads
A separate full OMB run used Krabka 1.1.5 with preallocate=true, so Krabka wrote its segments with O_DIRECT, as Redpanda does. Redpanda ran as Redpanda 26.2.2.
14 max-rate workload / replication-factor pairs; workload range 116% to 748%. One measured round per workload.
Throughput is the lower of each workload’s publish and consume rates. It does not show tail latency. The benchmarks page lists end-to-end p99 for each workload.
Compare Krabka's direct I/O results with Redpanda| Replication factor | Median throughput |
|---|---|
| RF1 | 151%7 max-rate workloads |
| RF3 | 156%7 max-rate workloads |
Results apply to this host and these workloads. Similar throughput does not establish equal latency or crash durability.
Use your existing Kafka clients
Trying your clients takes one line. Moving a production cluster takes more: read what changes when you migrate a cluster.
Each example runs against Krabka nightly in CI.
Memory-safe, proved, and model-checked
-
forbid(unsafe_code)
the lint every workspace crate builds under
No unsafe Rust. The compiler enforces it across the workspace.
-
259
kernel functions proved
Creusot proofs. Vote admission, quorum size, ISR and leader failover, and other decision kernels are proved against their contracts.
-
13M
states checked across 35 models
Stateright models. Consensus, failover, transactions, and group coordination are checked for every reachable state within each model’s bounds.