I didn't catch the part where "Parent wants storage at the source node.".

So if the goal is to eventually have the timeseries data merged back to a timeseries DB, and latency isn't too much of a concern then wouldn't batch writing to Kafka (Kinesis, etc) be tolerable?