However, I wanted to share some issues I found with your benchmarking approach that I believe are worth addressing:
1. Testing on a MacBook laptop is not a good idea due to thermal throttling. At some point, the numbers become meaningless.
2. I am not very familiar with Graviton CPUs, and after checking the AWS website, it is not clear to me whether they are virtualized. Since they are labeled as "vCPUs," I assume they are virtualized. Virtualized CPUs are not ideal for benchmarking because they can suffer from work-stealing and noisy neighbor effects.
3. The replication factor in Kafka's "Getting Started" guide is set to 1, which is also the case for Fluvio. However, in real-world scenarios, RF=3 is typically used. A more representative benchmark should include RF=3.
4. You mentioned: "Given that Apache Kafka is the standard in distributed streaming, and it’s possible for intelligent builders to extrapolate the comparable RedPanda performance." However, this is not accurate. RedPanda uses a one-thread-per-core model with Direct I/O, which results in significantly better performance.
How to Address These Issues:
1. It would be preferable to test on a bare-metal server-grade CPU rather than virtualized hardware, such as i3.metal instances on AWS. 2. Run the benchmark with RF=3 to reflect real-world usage more accurately. 3. It would be more insightful to compare against RedPanda, as both Fluvio and RedPanda use non-garbage-collected programming languages. The goal should be to evaluate how well Fluvio scales with increasing CPU counts.
Cheers.
We will do a full setup and benchmarks comparing Kafka, Pulsar, RedPanda using a real dataset on barmetal servers soon.
