Small Datum

Last October someone asked on the RocksDB email list whether RocksDB can use all of the IO capacity of a fast storage device. They provided more context -- they want to use it to cache objects that are ~4kb and give RocksDB as little memory as possible and the workload is mostly point queries (RocksDB Get). That person provided a shell script to show how they used db_bench to test this. I started with that script, we had some discussion on that thread and then I went quiet. At last I have some results to share. To be clear, I am trying to answer two questions. First, can the QPS rate for RocksDB approach the IO capacity of the storage system. Second, can the IO rate for a RocksDB benchmark approach the IO capacity of the storage system. And the context is a for a read-mostly/large-object workload that gives RocksDB as little memory as possible. The RocksDB workload should do ~1 IO/query for the answer to the first question to be true. If it does much more than 1 per query then we are