Slightly higher latency is in the eye of the beholder. Initially I was also bit surprised by the increase in latency. A single node standalone DB will shit all over any distributed database. Read scalability mostly easy if you can get by with eventual consistency, solved problem 15 years ago. Where the single node and naive replication story plateaus is write scalability and equally important DDL.
"Instead of vibe coding it all, let’s use an ancient technique called learning. First of all, what is a web server? There are a few essential syscalls."
https://t.co/ioHgWxDUyt
Ran a perf test on 3 TiKV nodes and accidentally put one on a QLC NVMe. Write latency started climbing, and PD quietly moved the region leaders to the two healthy hosts. With most databases this is the kind of thing that keeps you up at night. With TiKV I forgot it was even a feature, you take it for granted. I noticed it while I was puzzling over a latency spike. Too good!
This server cost me about 300USD with RAM IIRC, the ConnectX-4 NIC ~75USD, PLX controller ~100 USD for the NVMEs. Drives are consumer level NVME and SSDs. The last time I checked I will have to pay a lot more per month for all of this and won't get anywhere near the latency and throughput.
@rene_cannao I think the MySQL protocol needs to be improved, databases like TiDB can do a lot of parallel processing. I remember Meta has been asking for proper async support forever.
A good library of basic building blocks pays off. Read every line of it, and make it as correct and performant as you can. Normally you’d just reuse it in other projects. I point the agent at the library and have it pull code and idioms from there to build something new. So far it’s working really well for me.
Current performance is 80% of raw/local disk performance. My target is to keep it above 95%.
I'm going to combine the two, the backend will be the same, an optimized KV block store with optional direct pass through (TBD), different front end protocols.
The iSCSI is already doing real work in my homelab. In hindsight good decision to start using it because I encountered a lot of bugs. I'm sure there are plenty more but this is probably the best way to find and fix them.