The first issue I faced was about maintenance_work_mem because I set it to 16GB and the server silently crashed during a VACUUM because I didn’t consider that it could take up to autovacuum_max_workers * maintenance_work_mem (roughly 48GB).
I don't think that this is the true cause of the problem. In current versions of PostgreSQL, VACUUM cannot make use of more than 1GB of process-local memory, even if maintenance_work_mem is set to a far greater value.