8 days ago I wrote about forty-nine queries a second -- the first honest million-vector number, and the mutex of mine that explained it. Today: same collection, same dataset, same VectorDBBench harness, measured through the public API like any customer would hit it. 1,620 QPS at 0.9652 recall@100, serial p99 under 50ms. Numbers are on the benchmarks page with the raw output attached.
No single fix. 8 days of finding the next owner of the next millisecond and evicting it: the dispatch mutex from the last post, a billing writer that had no business on the query path, a response-router holding a lock across an await and convoying everything behind it, admission caps that were really latency knobs cosplaying as throughput knobs. None of it clever. All of it measured before and after, which is the only reason I trust the sum.
My favorite failure of the week: the first "clean" full-path run came back at 1,434 QPS and declining above 50 concurrent clients, which looks exactly like a saturation knee. It was 7 gateway workers times a 200-requests-per-second-per-key token bucket. I benchmarked my own rate limiter, appx to the digit. Raised the limit, reran: 1,620, dead flat from c100 to c200 while latency grows linearly -- a system doing what it can and queueing the rest. The limiter run is published too. Best lesson in benchmark skepticism I own, and I didn't even have to write it.
What's left is physics, and I can show my work. At saturation every worker core walks a graph where each hop is a random read of a 3KB vector that isn't in cache. The kernels are vectorized -- I checked the deployed binary's disassembly, because I didn't believe it either. Coordinator under 1% busy. Queues empty below the caps. The remaining levers are fewer bytes per hop and more cores. Real work, not found money.
The part I actually care about: the collection serving those 1,620 QPS holds 1M 768d vectors and costs $17.14 a month at list. The published VDBBench rows I compare against run on $832-to-$1,035-a-month configurations. They beat us on absolute throughput -- Zilliz by 8x, and I say so right on the benchmarks page -- but per dollar it isn't close, and the recall band is the same neighborhood: 0.9652 against their best row's 0.9588. Rerun all of it yourself. The harness is Zilliz's own, our client for it is public, and the API you'd hit is the one this page is served behind.
$0.02, at $17.14 a month.