Collisions are likely in headAppender.Add's ref logic about tsdb HOT 4 CLOSED

Yes, this should be fixed before the next alpha. A counter within the transaction and a signaling bit for clients to not store the ref beyond the transaction would be one approach.

That means that all clients that want to use AddFast have to maintain two separate caches – for within the current transaction and one for beyond. As it is, after the transaction initially creating a series, the next transaction has to to a regular Add for the same series regardless.
Given that Prometheus itself does not currently append multiple samples for the same series in a transaction, I wonder if it is worth it at this point.

Even easier would be to reserve the 0 or max ID for "not found" and just always return it if a series was newly created.

from tsdb.

How bad is it just to do the Add every time?

from tsdb.

Every time as in never do the AddFast? It's one of the major reason for saving CPU aside from series maintenance. Without it, about 80% of CPU cycles are spent on allocating/GCing byte slices and hashing them.

from tsdb.

This can be closed.

from tsdb.