Software triage

Web App Down: A Triage Order for Custom Software When the Error Page Is All Your Users Can See

A custom web app fails in more ways than a brochure site. There is a database behind it, a deploy pipeline, a login system and a handful of outside services, each with its own bad days. When it goes down, the first question is not how to fix it. It is which of those parts broke.

November 18, 2026 · 10 min read

portal.acmefreightlines.com · incident log
A customer portal returning 504s, one morning
Rolled back at 10:31, the slow query found by noon
  1. 10:02 portal returning 504 on shipment pages, all users
  2. 10:06 deploys frozen; host and cloud status pages clear
  3. 10:14 last release 09:40: new shipment search filter
  4. 10:19 logs copied; first error is a database timeout
  5. 10:31 rolled back to the 09:12 release, pages load
  6. 11:05 data checked: no migration ran, records intact
  7. 11:52 cause: filter query missing an index, fix on a copy
Illustrative: a sample incident log for a fictional freight company's customer portal

The short answer

When a custom web app goes down, triage in this order: write down the exact error users see, check the status pages of your host and outside services, look at what deployed most recently, read the logs for the first error, then check the database. If a deploy caused it, roll back first and diagnose second. Leave logs and data untouched until the cause is known.

On this page
  1. The status code on the error page is your first witness
  2. Sixty minutes on a broken web app
  3. Six layers that fail, and the deeper guide for each
  4. Roll back first, then investigate on a copy
  5. How a two-hour outage becomes a two-day outage
  6. One honest update beats forty worried emails
  7. Calling in DataCram: response, price and the access we need

The status code on the error page is your first witness

Before anyone opens a terminal, write down exactly what users see, including the number. The 5xx codes are defined in the IETF's RFC 9110, and each one points at a different layer of the stack. A screenshot from a customer that says only "it's broken" costs ten minutes. A screenshot that says 504 saves them.

What each common error code says, per RFC 9110, and where to look first
CodeWhat the standard saysWhere we look first
500The server hit an unexpected condition; MDN calls it a generic errorApplication logs, the most recent deploy
502A gateway or proxy got an invalid response from the server behind itThe app server or function behind the proxy
503The server is temporarily overloaded or down for maintenanceTraffic spikes, capacity, a maintenance switch left on
504A gateway or proxy did not get a timely response from upstreamSlow database queries, a stalled outside service
401 or 403Credentials missing or invalid, or known and refusedLogin provider, expired keys, changed permissions

Our existing pages on the 500 error, the 502 bad gateway, the 503 service unavailable and the 504 gateway timeout cover each one in depth, mostly for websites. This page is the emergency order of work for software: the app your staff or customers log into, built for your business, where nobody can simply reinstall a theme.

Sixty minutes on a broken web app

  1. 1

    Minute 0: one person in charge, one channel, no deploys

    Name who makes decisions and where updates go. Stop all deploys and configuration changes. Write the time the problem started and the time it was noticed; they are rarely the same.

  2. 2

    Minutes 0 to 10: capture the failure exactly

    Which pages, which users, which actions, which error code, since when. Does it fail for everyone or only for one account, one browser, one region? Save screenshots with timestamps.

  3. 3

    Minutes 10 to 15: read the status pages you depend on

    Your host, your cloud provider, your payment processor, your login provider, your email service. Vercel's status page, for example, lists Builds, Functions, CDN, DNS and TLS Certificates separately. A matching incident changes the whole plan.

  4. 4

    Minutes 15 to 25: check what shipped last

    If a deploy, a dependency update or a configuration change went out before the failure, roll back to the last version that worked. Diagnose the bad release afterward, on a copy, without users waiting on you.

  5. 5

    Minutes 25 to 40: save the logs, then read them

    Copy the application and server logs from before the failure before anyone restarts anything. Look for the first error, not the loudest one. The loud errors are often a downstream echo of a quiet one.

  6. 6

    Minutes 40 to 60: check the database without writing to it

    Is it available, is storage full, are connections maxed out, does recent data look right? If data looks wrong, stop the writes that are changing it. Do not restore a backup over it yet.

Warm shallow-focus desk at night, a laptop with a blank glowing screen, a phone face down beside a mug of coffee, one hand resting near the keyboard
The first hour on a broken app is mostly reading: error codes, status pages, the deploy history and the logs.

Six layers that fail, and the deeper guide for each

Once the first hour narrows the cause to one layer, the work becomes specific. Each row points to the page in this section that covers that layer in full.

Matching a web app failure to its layer
LayerTypical symptomFirst checkDeeper guide
A releaseBroke right after a deploy or updateDeploy history, rollbackFailed software update
The databaseTimeouts and 500s on pages that read data, connection errorsDatabase status, storage, connectionsDatabase emergency
An outside APIOne feature fails, with 401, 403 or 429 in the logsThe provider's status page, keys and limitsAPI integration broken
The host or cloudEverything down at once, provider incident listedProvider status page and health dashboardCloud server outage
PaymentsCheckouts or renewals failing, orders not recordedPayment provider status, webhook deliveriesPayment processing failing
Domain or certificateApp unreachable by name, browser security warningDomain expiry, DNS, certificate datesDomain and DNS emergency

A word on outside providers. AWS's shared responsibility model says AWS protects the infrastructure, while on EC2 the customer manages the guest operating system, the application and their data. So even when the cloud is healthy, the app can be your problem. And when the cloud is not, the EC2 service level agreement commits to a monthly uptime percentage of at least 99.99% for instances across two or more Availability Zones, as of November 2026, with service credits as the remedy. Our reading: an SLA pays back part of a bill. It does not pay for lost sales.

Roll back first, then investigate on a copy

When a release caused the outage, the fastest safe move is usually to put the old release back. Vercel's Instant Rollback page describes reverting to a previous production deployment for exactly this, and says the domains point back to the selected deployment. Pro and Enterprise teams can roll back to any deployment previously aliased to production; Hobby users can roll back to the immediately previous one, as of November 2026.

Two details catch teams out. Vercel says that after a rollback it turns off auto-assignment of production domains, so new pushes to the production branch will not replace the rolled-back deployment until someone promotes one. And during a rollback, changed environment variables are not applied. On other hosts, git's own revert command records new commits that reverse a faulty one, which keeps the history honest and makes the fix easy to redo once the bug is understood.

How a two-hour outage becomes a two-day outage

The common thread in long outages is not a hard bug. It is evidence destroyed in the first hour by someone trying to help. Here are the usual ways it happens.

  • Restarting the servers again and again. It can clear the error for ten minutes and erase the logs that explained it.
  • Restoring last night's backup over the live database. A point-in-time restore can reach only as far back as the retention window set before the problem, and a blind restore can overwrite the records that show what went wrong.
  • Deleting and recreating a database instance to clear an error. AWS's RDS backup documentation says that if you don't choose to retain automated backups when deleting an instance, all automated backups are deleted with it.
  • Shipping a quick fix straight to production with no test. Now there are two unknown changes instead of one.
  • Rotating every API key at once because one might be compromised. Google Cloud's API key best practices recommend creating new keys, updating applications to use them, then deleting the old ones, in that order.
  • Saying nothing to users for three hours while the team works in silence.
Restarting a broken server is the software version of shaking a vending machine.

On the database specifically, AWS says RDS backup retention can be set between 0 and 35 days, as of November 2026, and that setting it to 0 disables automated backups. Neon's history window, which limits how far back a restore can reach, depends on the plan. Find out what yours is before you need it, not during the outage. The database emergency page covers restores in detail.

One honest update beats forty worried emails

Within the first hour, tell users that you know, what is affected and when you will update them next. Not when it will be fixed, because you do not know yet. Atlassian, in its Statuspage documentation, describes the product as a communication tool for informing users about outages and scheduled maintenance, with email and text notifications to subscribers. Our advice, drawn from that: one posted update that people can subscribe to saves your team from answering the same question forty times. A shared email or a banner in the app does the same job on a smaller budget. The SaaS outage communication page covers wording and cadence.

Calling in DataCram: response, price and the access we need

We respond the same business day, during the hours on our contact page. We triage in the order above, stop the damage, restore service, find the root cause and write it down so it does not happen twice. Software repairs are from $575, depending on complexity, with a written estimate before work starts. Connections to outside APIs and integrations are restored the same day once we have access. The emergency software repair page explains how the work runs, and emergency pricing explains what moves a quote.

Hosting and cloud

A login with permission to read logs, view deployments and roll back, on your host or cloud account.

The code

Access to the repository and a note of which branch is production. If nobody knows, that is the first finding.

The database

Read access first. We ask before writing to anything, and we take a copy before any restore.

The outside services

A list of the payment, email, login and data providers the app calls, and who holds each account.

Call us today if

  • The app is down and nobody on staff can read the logs
  • A rollback did not fix it, or you have no way to roll back
  • Data looks wrong or missing
  • The developer who built it is not reachable

Wait, or handle it in-house, if

  • Your cloud or host lists a matching incident: nobody can speed that up
  • Your own team has a rollback and it worked
  • Only one outside feature is down and its provider shows an incident
  • The app is internal and can wait for business hours

For the longer view, see web app development and custom software development. If the outage exposed that nobody owns the system, a fractional CTO can take that on. DataCram is a web studio in San Antonio, Texas, building for the web since 1999, working with businesses across the US by phone, video and a shared project page. See our work, the rest of the emergency repair section, or call (210) 346-0848.

Questions owners ask in an emergency

Write down the exact error and when it started, stop all deploys, and check the status pages of your host and outside services. If something shipped just before the failure, roll it back. Then save the logs before anyone restarts anything, and check the database without writing to it.

Roll back when a recent release caused the outage and a known good version exists. It restores service while you diagnose on a copy. Remember that a code rollback does not undo database changes the bad release made, so check the data and confirm the old code can still read it.

No. If AWS, your host or another provider lists a matching incident, the fix is theirs. Use the time to post an update for users and confirm your backups. AWS's EC2 SLA offers service credits when uptime falls below its commitment, as of November 2026, which repays part of a bill, not lost sales.

Save the logs first. A restart can clear the symptom for a while and wipe the evidence of the cause. If you must restart to restore service, copy the logs from before the failure, note the time, and expect the problem to return until the cause is found.

Software repairs are from $575, depending on complexity, with a written estimate before work starts. We respond the same business day, and connections to outside APIs are restored the same day once we have access. The emergency pricing page explains what moves the number.

Usually. We need access to the hosting, the code repository, the database and a list of outside services. If the original developer is gone, we start by securing those accounts, then triage the outage in the same order as any other.

Keep reading

Free · No obligation

Get your app back up

Tell us what users see, the error code if there is one, and what changed before it broke. We reply the same business day with a triage plan and a written estimate.

Which of these sounds like you? (pick any)

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

CallGet a quote