0

Backups, PITR, read replicas & cost traps

Opsused in ~55% of apps

Someone on the team runs a migration script against production with a WHERE clause that doesn't match what they thought, and half of DevLog's posts are gone. The question that decides whether this is a bad afternoon or a resume-updating event is one you should have answered BEFORE it happened: what's your actual recovery point, how destructive is getting there, and does the read replica everyone's been pointing reads at help even slightly? (It doesn't.)

Core idea. Backups, PITR, and read replicas solve three different problems and none of them substitutes for another. Daily backups and PITR both answer "how much data can I lose" (RPO) — a day, versus seconds — but restoring either is destructive and takes the whole project down; there is no per-table undo. A read replica answers "can I survive load and serve reads near users," and because it's a live streaming copy, it faithfully replicates your mistakes too — a DELETE reaches the replica just as surely as it reaches the primary. And the cost model has its own trap: Spend Cap being ON does not mean you're protected from a big bill, it means overage-based scaling is silently disabled while add-ons you explicitly turned on keep billing regardless.
Playground, this isn't a screenshot
click a point in the week — that's when someone runs a bad DELETEday 3.4
now-7d
now
disaster

Blue dots = daily backups. Green band = a PITR window covering the whole week.

Daily backup only

Nearest restore point is the backup taken at the start of day 3. You recover the DELETE — and also lose every legitimate write made since then: ~9.6h of real data, gone too.

PITR ($100/mo per 7-day window)

Restore to the second before the DELETE ran. ~0 data lost. The restore itself is still destructive and takes the whole project down while it runs — there is no per-table undo.

Someone suggests: "just fail over to the read replica"
↑ play a little first, then dig in