pg-recover/README.md

pg-recover

⚠️ DO NOT DO THIS… well ever really, but especially on a server with failing disks. This is done on a server with perfectly fine disks, but corrupted Postgres blocks.

A dirty, terrible, dangerous Postgres recovery tool.

This is designed to recover as much data as possible from a Postgres table with bad disk blocks BUT good physical disks (say after an unclean exit, or you've snapshotted bad disks and mounted them on a good server). It does this by creating a clean "recovery" table with the same schema and reading rows from the bad table one at a time, skipping rows and blocks with bad disk errors.

For more specifics on how this works technically, see my blog post.

Requirements:

• Bash + tools (cat, tac, cut, sed, &c.)
• Postgres client (psql and pg_dump)
• Table must have a SERIAL primary key and at least one other non-NULL column

Usage:

USAGE: pg-recover.sh <user> <host> <database> <table> <primary key> <nonnull col> [<commit size>=500] [<start at>]

user - the Postgres user to connect with
host - the Postgres server host
database - the Postgres database
table - the Postgres table
primary key - the name of the SERIAL primary key column
nonnull col - the name of a DIFFERENT non-null column on the table
commit size - how many rows to recover before committing the transaction (default: 500)
start at - start at the specific primary key (descending)

Copyright (c) 2025 Garrett Mills <shout@garrettmills.dev>
https://code.garrettmills.dev/garrettmills/pg-recover

Once the script finishes, you can import the recovered data like so:

psql [...] < pqr-final-attempt.sql

This will create a new table <original table>_recovery with the recovered data.

License: See the LICENSE file.