KeyDB

I run a fork called KeyDB with multithreading as well as a few optimizations directly around large values (redis does unnecessary copies). It may work well for your use case. https://github.com/JohnSully/KeyDB

I'm not sure if you're referring to KeyDB or something else as "closed source commercial software", but KeyDB is BSD licensed like Redis: https://github.com/JohnSully/KeyDB/blob/unstable/COPYING

If you ever have the time I’d appreciate any help you van give with KeyDB. It’s an attempt to accelerate progress in the Redis universe. https://github.com/JohnSully/KeyDB

KeyDB: a multi-threaded fork of Redis. https://github.com/JohnSully/KeyDB It seemed crazy to me Redis was still single threaded in 2019. Sometimes the rationale was one thread was enough to saturate a machine (it is not). Other times it was justified by the fact that you could create a local cluster - but I thought that was needlessly complex. Eventually I went ahead and just did it. The project has since expanded to include Active Replication and a bunch of other new features.

Network I/O is a major bottleneck for Redis but they leave a lot on the table by being single threaded. I can speak from experience because I maintain a Multithreaded Fork: https://github.com/JohnSully/KeyDB KeyDB can easily get 2-3x the QPS with half the latency.

I just implemented S3 loading in this change: https://github.com/JohnSully/KeyDB/commit/a45f212693956a6fb1... Its currently in the unstable branch, once it gets shaken out a bit I'll release 0.9.3 with it included.

Redis actually has a surprising number of performance issues with larger data sizes. But most people use it with small value sizes and TCP performance dominates. I could write for hours on how bad Linux TCP/IP perf is with small packets but I'll save that for another time. This is the bottleneck with the benchmarks op has done here (the value size is only 3 bytes in these benchmarks). Since we're sharing Redis clones, here's my multithreaded one: https://github.com/JohnSully/KeyDB

Assembly is still quite useful in a number of circumstances. The compiler was generating terrible code for a ticket (spin) lock, and rewriting in assembly helped considerably [1]. Its also still the best way to get the most out of AVX and SSE. I find the register allocation to be extremely poor when using intrinsics. 1. https://github.com/JohnSully/KeyDB/blob/unstable/src/fastloc...