PlanetScale Makes Postgres Backup Time Independent of Size

PlanetScale detailed a sharded Postgres backup architecture where duration depends on shard count rather than data volume. Each shard gets a temporary node that restores the previous backup from object storage, replays archived WAL with wal-g, catches the final minutes from the primary, then encrypts and uploads before being torn down. The primary is never read for the bulk transfer, which is why production does not feel it. A 32TB database takes about 2.8 hours on 8 shards and 42 minutes on 32, at transfer rates above 50 GB/s. Backup duration stops being an argument against sharding.