It seems like the checkpointFrequencyPerPartition determines how many events get processed before a checkpoint is written (-> bigger numbers result in slower checkpointing)

I'll confirm this in about a month when the next billing information is available.