I notice that LDT are enabled on this namespace. Is there LDT traffic going on at the same time as the regular write traffic?
Could you record iostat and capture how it evolves during those latency spikes? (to check if it could be caused by the underlying device).
You could also turn on microbenchmarks and analyze as described in this post about write performance analysis, especially to see if those latency spikes are caused by the network.