Given this scenario, I believe the behavior you are seeing is by design. The solution is for us to provide server-side buffering (which is still on our long-term roadmap). In the meantime, your options are likely the ones you listed - turn off PI SDK buffering or use PI AF SDK (PI API is not recommended) to delete data from each collective member. The downside is that you will need to connect to each member individually to delete the data (and write the data if you decide not to use buffering). You may also need to verify the data is properly written/deleted on each member. This will still "preserve" HA in terms of synchronizing the point configuration, but you will be responsible for ensuring the data is in sync between the collective members.