Differential Backup

Incremental backups are small but fragile. Full backups are reliable but expensive. Differential backups sit between the two, giving you faster restores than incrementals and smaller files than fulls.

Stripe discovered this during a 2019 architecture review. They had been running incremental backups for five years because the storage cost was roughly half that of differentials. When they measured actual restore time under simulated disaster load, rebuilding a Tier-0 database from a six-day incremental chain took 90 minutes longer than restoring from a full plus one differential. At their cost of downtime, that 90 minutes was worth more than a year of the storage savings. They switched.

The simplicity advantage is real. A differential restore always needs exactly two files: the last full and the latest differential. No chain walking, no dependency tracking, no ordering bugs at 3am. That predictability is its own form of reliability.

A differential backup captures all blocks that changed since the last full backup, not since the last backup of any type. This means Monday's differential is small, but by Friday it contains the accumulated changes of the entire week.

The critical difference from incrementals: Monday's differential captures Monday's changes. Tuesday's differential captures Monday's changes plus Tuesday's changes. Wednesday's captures Monday through Wednesday. Each differential is self-contained relative to the full. There is no chain to break.