FASTER vs RocksDB

I hope to start a discussion on the upcoming implementation of RocksDB which I understand is tried and tested, vs something I just stumbled across called FASTER GitHub - microsoft/FASTER: Fast persistent recoverable log and key-value store + cache, in C# and C++.

I have very little to contribute to the discussion but this is how I stumbled across it in trying to learn why NF chose RocksDB https://youtu.be/xWNbbkQMtfI

So, can FASTER be compared and contrasted to RocksDB?

3 Likes

It does look interesting.

FASTER: How Does Microsoft KV Store Achieve 160 Million OPS

Test

The paper compares FASTER with two high-performance pure in-memory systems (Masstree and Intel TBB hash map) and two leading KV stores (RocksDB and Redis).

For a single thread, according to the uniform and Zipf distribution, FASTER performs best, TBB performs second, and RocksDB performs last.

When 256 threads are used, the uniform distribution of FASTER reaches 110 million, and that of TBB is also close. But for Zipf, FASTER gathers hot data in the mutable region and achieves a throughput of 160 million, which significantly outperforms other systems.

In the scalability test, FASTER scales very well on both one CPU and multiple CPUs. Masstree also scales well but has much lower performance. TBB scales well on one CPU, but falls over when running on multiple CPUs. For more details about the test, see the paper.

4 Likes

The database backend is designed to be swapped out by a pluggable backend; RockDB was the first such plugin.

Feel free to run some comparisons and report data.

5 Likes