Can you quantify this in terms of throughput / latency. Do you see any errors when doing rapid writes?
flush-max-ms and fsync-max-sec are only relevant to how fast write buffers are flushed to the disk. And this is not generally required to be set when you have very fast write rate, in which case write buffers are flushed at rapid rate already.
The initial config chooses to persist the data in the storage engine to file, which uses the filesystem, rather than a raw device. Indeed, writing to the block device gives better write performance in the event of a high write rate.
Read the SSD setup section in the operations manual.
Why not directly - historical… The problem is not relevant anymore
Microbench showed problems with the network. We just upgrade the cluster to really good hardware.