It's only after I wrote this commit that I ran the whole test suite and then I realized that it breaks backup, because the backup master relies on invalidations to know which partition needs to be synced from upstream.

Even if this commit would reduce the impact of a memory leak, it does not avoid it completely, so it's just optimization. Is it worth, for example:

• adding an RPC for the backup master to tell the upstream one that it's interested in invalidations for both modified and new oids ?
• or rather stop using invalidations for that and send directly the list of partitions ?

The second option is actually an extra optimization.