At Comper we have a very hot key-value store for annotating git data. We maintain a parallel git-blame data structure so we can do incremental "git blame -w -M -C -C". Typically a very expensive operation, but if you make it incremental, you can make it very cheap when new commits need to be analyzed. However, building the git blame tree is still pretty intensive for large repos.

We currently use rocksdb with storage on the same node, and hit rocksdb 1000s of times per second during our analysis. About 20% writes, 80% reads. The issue is that we need to start scaling horizontally, for burstable workers and zero-downtime deployment. So we're thinking to offload to an external kv service instead of a local rocksdb.

TiKV seems a good replacement, about 3-4x slower, but very scalable. Reading this article, I think a separate postgres cluster with unlogged tables might be a good idea. If anyone has some experience to share, let me know!