Ten microseconds, 200 million messages a second: why streaming latency compounds
A calculator takes maybe 100 milliseconds to multiply two numbers. Nobody minds, because you press the button once and read the answer.
A computer that multiplied at that speed would run at 10 hertz, and nobody would use it. We wanted 8 megahertz, then 16, then a gigahertz, then 5 gigahertz times 128 cores. Most people could still get by multiplying two numbers at a time. Somehow we found a use for every one of those operations, and we keep asking for more.
Streaming is at the same point. The usual question about a fast streaming platform is who actually needs microseconds. This article gives our first benchmark numbers, and then our answer to that question.
The early numbers
These are preliminary figures from benchmarks still in progress, not a final result.
On a four-node system, AlgoX2 delivers a message in as little as 10 microseconds edge to edge, from publish to subscribe. The same system carries roughly 200 million messages a second: about 20 million published and about 180 million subscribed.
Why the test fans out
The split between published and subscribed is deliberate. The benchmark reads every message nine or ten times, because that is how streaming is used.
If you only needed to get data from point A to point B, you would just send it. People reach for a streaming platform because they need to send it from A to every point from A to Z. Many pipelines want the same data. One passes it to an LLM. Another backs it up, and a third backs it up somewhere else. A fourth runs analytics, and a fifth shows it in a user interface.
A large fan-out is the normal case, not the stress case, so that is the case we measure.
Build it and they’ll come
When a team says it has enough performance, it means the team is not doing what it could do with more.
History shows this again and again. People asked who needed more powerful graphics cards. Nvidia kept building them anyway, putting more and more pixels on the screen, until GPUs reached escape velocity. The same thing happened with CPUs, and with everything else that processes information. Demand for speed you could not predict appears once the speed exists.
We are building X2 on the same bet.
Latency compounds
Single-digit microseconds rarely matter for a single action. Few use cases need one operation done that fast.
But a modern system never does one operation. Something is published, processed at once and republished as a new data point. That point is processed and republished again, and so on. A real application is tens or hundreds of these round-trips, and every hop adds its own latency to the total.
If you have some sort of bottleneck in your system that’s occasionally causing 5,000 millisecond delays, you’re not going to be running certain apps that you could be running. If everything is in microseconds, sometimes even nanoseconds, that’s when it becomes a tool you can get leverage with.
When the floor is too high, infrastructure becomes hard to build on, because every extra step costs too much. When the floor drops low enough, chains of steps that were never practical become ordinary.
How AlgoX2 keeps each hop short is the subject of Everything is a Stream, and How to Build an Exchange shows a system that depends on it.
Speed is room to build
A fast platform does not matter for the one message you send. It matters for the thousandth step in a pipeline nobody has designed yet. Ten microseconds per hop is what leaves room for that pipeline.
Watch the video
Plays from YouTube, which may set cookies. See the Privacy Policy.
Want more context on the platform? Read about the product, or get in touch with us on the contact page.