Proposal

While we plan to implement the content negotiation feature (#13365) e.g to turn of certain type of transactionality/atomicity for classic histograms, I would like to create design on how such atomicity could be accomplished for remote write senders.

Firstly, this will allow more backends to implement remote write, which was not possible previously.

Secondly, even for existing backends this will mean improved reliability (scrape transactionality for histograms). Currently if one remote write shard will fail (or is slower) and second succeed you might have partial histograms in the end storage/query time.

This is related to the full "scrape transactionality" for all series as one user might e.g. query for both http_total{status=failed} and http_total{status=success} and those could be sent at different times. However focusing on histogram atomicity might be the first step.