Have you tidied away the unwanted rrd files? I don’t know if will still carefully update/compress the archives as time marches on, it might keep populating the per-minute pigeonholes even if it is with NaN, Meaning it still processes 600 archives every minute.
The random nature of failure does not surprise me.
There was recent post about similar jdbc workload-related errors. It would be nice if the persistence service logged WARNs for timeouts, if that is what is happening.