Confluent Cloud Passthrough Benchmark
This benchmark compares connecting to Confluent Cloud directly with connecting through Zilla Plus.
Summary
- Under 3 ms added at p99. Zilla Plus added 2.8 ms to publish latency and 4.1 ms to end-to-end latency at p99.
- Lower latency in the far tail. From p99.9, latency through Zilla Plus was 6 to 10% lower than connecting to Confluent Cloud directly.
- Each Zilla Plus instance carried about 200 MB/s, inbound and outbound, at about 65% CPU across 4 cores.
- Latency spikes on the direct path lined up with CPU spikes on the Confluent Cloud cluster.
Setup
| Setting | Value |
|---|---|
| Deployment | m5.xlarge: 3 Zilla Plus, 6 benchmark clients, 1 controller |
| Cluster | Confluent Cloud, 2 CKU |
| Security | SSL |
| Traffic | 100 MB/s, 45 topics with 200 partitions each, 1 producer and 3 consumer groups per topic |
| Messages | 2 KB |
| Producer | acks=all |
| Duration | 10-minute warm-up, 20-minute measurement |
| Measurement | Publish and end-to-end latency from the OpenMessaging Benchmark |
Results
Publish latency, in milliseconds:
| Setup | p50 | p95 | p99 | p99.9 | p99.99 |
|---|---|---|---|---|---|
| Confluent Cloud | 29.4 | 78.7 | 127.2 | 245.6 | 414.6 |
| Through Zilla Plus | 31.3 | 82.7 | 130.0 | 230.8 | 374.6 |
| Difference | +1.9 | +4.0 | +2.8 | −14.8 | −40.0 |
End-to-end latency, in milliseconds:
| Setup | p50 | p95 | p99 | p99.9 | p99.99 |
|---|---|---|---|---|---|
| Confluent Cloud | 29.7 | 80.8 | 133.8 | 262.2 | 444.3 |
| Through Zilla Plus | 32.1 | 86.0 | 137.9 | 247.2 | 404.5 |
| Difference | +2.4 | +5.2 | +4.1 | −15.0 | −39.8 |
Caveats
- The run uses a 10-minute warm-up and a 20-minute measurement. In a longer 30-minute warm-up and 60-minute run, direct connections to Confluent Cloud degraded significantly after about 39 minutes of sustained 100 MB/s load.
- Confluent Cloud is a managed, shared service, so cluster-side CPU spikes show up in both paths.
- Latency depends on cluster size, message size, partition count, and client settings. Measure your own workload before sizing a deployment.