Stop writing, then think

Database Down or Damaged? The First Hour Before Anyone Presses Restore

The orders table is empty, or the app cannot connect, or a migration ran at 10:40 and half the customer records now show the wrong prices. Someone already has the hosting dashboard open, one finger over the restore button.

November 18, 2026 · 10 min read

acmeoutfitters.com · orders database incident
One bad import, worked in order
Restored into a copy, nothing written over
app paused
  1. 10:41 nightly import overwrote prices on 2,118 products
  2. 10:52 maintenance mode on, import job and queue paused
  3. 10:58 manual snapshot of the damaged database taken
  4. 11:06 retention read: 7 days, restore target 10:39
  5. 11:31 restored into new instance acme-orders-1039
  6. 11:48 prices, orders and customers compared: match
  7. 12:05 switched over, tested checkout, forms, login
Illustrative: sample emergency job on a fictional outdoor retailer's orders database

The short answer

Stop new writes to the database, take a copy of it exactly as it is, and find out how far back your backups or point-in-time window actually reach. Then restore into a new instance or branch, never over the original, compare the two, and switch the app over only after the tables that hold money check out.

On this page
  1. The restore button is the most dangerous control on the screen
  2. Sixty minutes, worked in this order
  3. What your provider actually keeps, as of November 2026
  4. Unreachable, full, wrong or stolen: four failures that look alike
  5. Seven ways a recoverable database becomes a lost one
  6. Handle it alone, or call before the next click

The restore button is the most dangerous control on the screen

A restore is often the right move. Done first and done in place, it is also the most reliable way to turn a bad morning into permanent loss. Restoring over the live database erases the evidence of what went wrong, and it erases every good row written since the backup was taken: the orders that came in at 9:15, the customer who updated her address, the invoice someone marked paid.

In a database emergency, the order of work matters more than the speed of work. Stop new writes. Copy what exists. Learn exactly what your backups can reach, in hours and days, from the settings rather than from anyone's memory. Only then put data back, and put it into a fresh copy you can check before the application ever sees it.

A restore you cannot undo is not a fix. It is a second incident with better intentions.

This page is about databases behind custom software and web apps: PostgreSQL, MySQL and the managed versions sold by AWS, Neon and others. If a WordPress site shows a single line about a database connection and nothing else, our database connection fix is the faster read. If the whole application is down and nobody is sure the database is to blame, start with the web app triage page and come back here once the trail leads to the data.

Sixty minutes, worked in this order

None of these steps needs a developer to begin. Each one keeps an option open that the next hurried click would close. Do them in order, and write down what you did and when.

  1. 1

    Stop the writes

    Put the application in maintenance mode, pause scheduled jobs, imports and queue workers, and switch off any sync that pushes records in from another system. Every write after a bad change either piles onto the damage or becomes good data that a restore would wipe out. An hour with a closed sign costs less than a week of reconciling orders by hand.

  2. 2

    Write down the times

    Note when things last looked right and when they first looked wrong, with the time zone. A point-in-time restore asks for a target moment. A guess of this morning, sometime, is how people restore to the wrong side of the problem.

  3. 3

    Copy what exists now

    Take a manual snapshot or a full copy of the database as it stands, damage and all, before anyone runs a repair. Amazon's documentation notes that manual snapshots are not deleted when an instance is deleted. That copy is your evidence and your undo button.

  4. 4

    Read how far back you can reach

    Open the backup settings and read the retention period or history window configured today. That number, not the one anyone remembers choosing, decides whether last Tuesday is still recoverable.

  5. 5

    Restore into a new copy

    Restore to a new instance or branch, never over the original. Compare the two: row counts on the tables that hold money, a handful of recent orders and customers, and the record that first tipped you off.

  6. 6

    Switch over and test the money paths

    Point the application at the verified copy. Place a test order, submit a form, log in as a customer, and run the report the owner checks every morning. Reopen only after all of them work.

  7. 7

    Keep the damaged original

    Do not delete the broken database the moment the new one works. Keep it until someone knows what happened and nothing else turns up missing a week later.

What your provider actually keeps, as of November 2026

Backups are a setting, and settings have defaults that nobody revisits. Here is what three common setups keep, from each vendor's own documentation as read in November 2026. Plans and limits change, so check your own console before you plan a recovery around these numbers.

Point-in-time recovery by provider, from vendor documentation as of November 2026
How far back you can goThe catch
Amazon RDS automated backupsAny point in the retention period, which can be set from 0 to 35 days; the default is one day through the API or CLI and seven days in the consoleA retention period of 0 turns automated backups off, and backups do not run while the instance is in a state such as storage_full
Self-managed PostgreSQLAny time since the base backup, provided write-ahead log files were archived alongside itpg_dump files are not file-system backups and cannot take part in continuous archiving
NeonFree plan 6 hours, capped at 1 GB of history; Launch defaults to 1 day, up to 7 days; Scale defaults to 1 day, up to 30 daysOnly the window set before the problem counts

Sources: Amazon's backup retention page, the PostgreSQL project's continuous archiving chapter for PostgreSQL 18, and Neon's history window page, all read in November 2026. The PostgreSQL chapter explains why point-in-time recovery works at all: the write-ahead log "records every change made to the database's data files," so replaying it on top of a base backup can return the database to any moment since that backup was taken.

Collage of a stethoscope and a magnifying glass resting on a server panel of glowing green and amber lights
Most of the first hour in a database emergency is copying and reading, not typing commands.

One conclusion follows from all three, and it is our reasoning rather than a vendor's promise: a restore can only reach back as far as the window that was configured before the trouble began. If a bad import ran eight days ago and the retention period is seven, no button brings that data back. Neon's instant restore page offers its customers one safety net worth knowing about: "Restore backups are created automatically in case you make a mistake."

Unreachable, full, wrong or stolen: four failures that look alike

"The database is down" covers several different failures, and each wants a different first move. Telling them apart takes minutes. Treating them all as a restore job costs hours, and sometimes the data.

It will not answer

The app cannot connect at all. The data is usually intact; a password, a host name, a connection limit or the database service itself has changed. A restore solves nothing here and risks a great deal. Check credentials and service status first, as our connection error guide walks through.

The disk is full

Writes fail and strange errors spread through the app. On Amazon RDS this is urgent for a second reason: automated backups do not occur while an instance sits in a state such as storage_full, per Amazon's documentation as of November 2026. Add storage before anything else.

The data is wrong

A bad migration, an import that overwrote a column, a script that deleted the wrong customers. This is the point-in-time restore case, done into a new copy, with the time of the bad change written down first.

Someone else was in it

Admin accounts nobody created, your customer list turning up somewhere it should not, a ransom note sitting in a table. Do not wipe or restore yet. Read our page on the first hours after a data breach and call your attorney.

If a release went out shortly before the trouble began, the database may be the victim rather than the cause. Our page on a failed software update covers that case, including the part that catches people out: rolling back the code does not roll back the data. And if your cloud provider is having its own bad day, read about cloud and server outages before you change anything on your side of the line.

Seven ways a recoverable database becomes a lost one

  • Restoring over the original, which overwrites the evidence and every good row written after the backup in a single click.
  • Deleting the instance to start clean and taking the automated backups with it, exactly as Amazon's documentation warns.
  • Letting the app keep writing after a bad migration, so damaged records mix with good new orders that a restore would then erase.
  • Treating last night's pg_dump file as the whole plan. It is a useful copy, but on its own it cannot replay the hours since it ran.
  • Running a repair or cleanup command on the only copy, before anyone took a snapshot of it.
  • Testing a backup for the first time during the emergency. A backup nobody has restored is a rumor.
  • Changing three things at once, then not knowing which one worked, or which one did the damage.

The satire writes itself: the overnight backup job that has failed every night since spring, faithfully emailing its error to a former employee's inbox. The cure is dull and it works. Restore a backup into a scratch copy once a quarter, on a calendar, time how long it takes, and send the alerts to an address the business owns. Our website health check is a free place to start looking at the rest of the stack.

Handle it alone, or call before the next click

You can likely handle it if

  • The database is unreachable and your provider's status page shows an incident in progress
  • You have a fresh snapshot and a backup that falls inside the window
  • The damage is limited to one table, the app is paused, and you know the exact time
  • Your developer is reachable and has restored this system before

Call us if

  • Nobody knows whether backups exist or how far back they reach
  • A restore was already tried and things look worse
  • Writes are still running and you cannot stop them safely
  • You suspect someone got in, or payment or health data is involved

When you call, we respond the same business day; our hours are on the contact page. Database work falls under our emergency software repair service, from $575, depending on complexity, with a written estimate before work starts. The emergency pricing page explains what moves that number. We will tell you plainly what your backups can and cannot reach. We do not promise to recover data that no backup or log still holds, and nobody honest can.

Afterward, the work that prevents the next emergency is unglamorous: a retention window long enough to notice a problem, a restore that has been rehearsed, and alerts that reach someone who still works there. Our hosting and maintenance plans start at $55 a month. For systems we build or rebuild, our custom software and web app development work includes the backup plan, and if the database lives inside an old application nobody understands, read our page on legacy software rescue. DataCram is a San Antonio web studio that has been building for the web since 1999 (about us); the emergency repair hub covers every other kind of outage.

Questions owners ask in an emergency

Not over the live database. First stop new writes and take a snapshot of the database as it is, damage included. Then restore the backup into a new instance or branch, compare it with the original, and point the app at it only after the important tables check out. Restoring in place erases the evidence of what happened and any good data written since the backup.

To any point within the backup retention period, which Amazon's documentation, as of November 2026, says can be set between 0 and 35 days. The default is one day when an instance is created through the API or CLI and seven days in the console. A setting of 0 disables automated backups entirely, so read the number on your own instance before planning anything.

It is a copy you can load, and it is worth having. But the PostgreSQL documentation says pg_dump and pg_dumpall do not produce file-system-level backups and cannot be used as part of continuous archiving. A dump alone cannot replay changes made after it ran. Point-in-time recovery needs a base backup plus archived write-ahead log files.

Neon calls it instant restore, also known as point-in-time restore. Its documentation says you can revert a root branch to any time within the history window, down to the millisecond. As of November 2026 the window is 6 hours on the Free plan, and it defaults to 1 day on Launch and Scale, with maximums of 7 and 30 days.

Usually not, and we will say so before you spend anything. A restore reaches back only as far as the retention or history window configured before the problem. Sometimes a copy survives elsewhere, such as an export, a replica, or another system that received the same records, and we will look for those. We never promise data recovery.

We respond the same business day, during the hours on our contact page. Work starts once we agree on a written estimate; emergency software repair starts from $575, depending on complexity. While you wait, stop the writes, take a snapshot and write down the times. Those three steps keep every recovery option open.

Keep reading

Free · No obligation

Get a second pair of eyes before you restore

Tell us which database you run, what changed and what your backup settings say. We reply the same business day with what we would check first and a written estimate.

Which of these sounds like you? (pick any)

Rather talk now? Call (210) 346-0848.

CallGet a quote