Down in the cloud
Cloud Server Outage: How to Tell a Provider Failure From Your Own, and What to Do While You Wait
Somewhere in a data center, a thing you have never seen has stopped working, and your phone is lighting up. Before anyone reboots, migrates or deletes, answer one question: is this the provider's outage, or yours wearing the provider's name?
November 20, 2026 · 10 min read
- PassProvider health dashboard for this accountno event
- FailApp server answers requeststimeout
- FailApp server disk spacefull
- PassManual snapshot taken before changesdone
- WarnHolding page with 503 from another hostgoing up
- WarnDatabase backup retention1 day
The short answer
When a cloud server goes down, check the provider's own status and health pages first, then your account's resources, then your most recent change. If the provider is at fault, you mostly wait, communicate, and serve a temporary 503 page from outside the failed system. If the provider is fine, the fault is in what you run on it, which is your responsibility under the shared responsibility model.
On this page
- Amazon's problem or yours: the ten-minute sort
- The first hour of a cloud outage
- The provider owns the building; you own the furniture
- An SLA is a coupon, not an insurance policy
- Keep Google's crawler patient while the server is gone
- Panic moves that turn their outage into yours
- After the all-clear: three settings to fix this week
Amazon's problem or yours: the ten-minute sort
Every cloud outage starts with the same rumor: "AWS is down." Occasionally it is true. Just as plausibly, one server, one database or one setting in one account has failed, and the big provider is a convenient villain. The sort takes ten minutes and saves hours of waiting for a fix that nobody at Amazon is working on, because nothing at Amazon is broken.
| Signal | Points to the provider | Points to your setup |
|---|---|---|
| Provider status or health page | Reports an event in your region and service | Shows nothing affecting your resources |
| Other sites and tools on the same provider | Also failing, widely reported | Working normally |
| Your error page | Provider-branded error, or no response at all | Your app's own 500, 502 or 504 page |
| Recent changes | None in the last day | A deploy, a config edit, a resized server, a certificate renewal |
| Scope | Several unrelated services down together | One app, one server, one database |
Our free is my site down checker answers the first question for a public website: down for everyone or just for you. The error code then narrows it. A 502 or 504 means a gateway did not get a good answer from the server behind it, which our 502 bad gateway guide and 504 gateway timeout guide explain in plain terms.
The first hour of a cloud outage
Most of this hour is looking, not touching. The work that matters is cheap and reversible. The work that hurts is expensive and permanent, and it is always the work people do first.
- 1
Read the provider's health view for your account
On AWS, the AWS Health Dashboard "requires no setup, and it's ready to use for authenticated AWS users," according to AWS Health documentation read in November 2026. AWS says it gives visibility into "your resource performance and the availability of your AWS services and accounts," which no public page can match.
- 2
Check the host's public status page
Vercel's status page, read in November 2026, lists Builds, Functions, CDN, DNS, Domain Registration and TLS Certificates as separate components, and offers email, SMS, Slack and webhook subscriptions. Subscribe now so the all-clear finds you.
- 3
List what changed in the last 48 hours
Deploys, resized servers, firewall or security group edits, expired cards on the billing account, renewed certificates. Put times next to each. A change that lines up with the first error is your lead suspect.
- 4
Take a snapshot before you take action
If the server or database still responds at all, take a manual snapshot. For its RDS databases, AWS says "Manual snapshots are not deleted when an instance is deleted." It is the cheapest insurance in this whole page.
- 5
Put up a holding page from somewhere else
A short page that says you are working on it, served from a host that is not the one failing, with the right status code. The section on Google below explains which code.
- 6
Tell staff and customers one true sentence
What is down, what still works, when the next update comes. Our outage communication page covers the cadence.
If the trail ends at a deploy, stop investigating the cloud. Roll back first, as our failed software update page describes. If the trail ends at the database, our database emergency page covers stopping writes and keeping the evidence intact before any restore.
The provider owns the building; you own the furniture
Cloud providers draw a line through every outage, and it is worth knowing where before the next one. The AWS shared responsibility model, read in November 2026, says "AWS is responsible for protecting the infrastructure that runs all of the services offered in the AWS Cloud," while "Customer responsibility will be determined by the AWS Cloud services that a customer selects." The more of the machine you rent raw, the more of the failure is yours.
Theirs: the infrastructure
Data centers, hardware, the network between them and the managed layers of each service. When that fails, the provider fixes it and you wait.
Yours on a raw server
For EC2, AWS lists the customer's job as "management of the guest operating system (including updates and security patches)," the software you install, and the security group firewall.
Yours everywhere: the data
AWS says "Customers are responsible for managing their data (including encryption options)." Backups that were never configured are not the provider's oversight.
Yours in practice: the design
One server in one zone with no tested backup is a choice someone made, often years ago and often without knowing it was a choice.

This is the uncomfortable part of a cloud outage. If your server's disk filled, its operating system crashed or its firewall rule was edited, the status page will stay green all day, because from the provider's side nothing went wrong.
An SLA is a coupon, not an insurance policy
Owners sometimes take comfort in the uptime number on the provider's contract. Read it closely. The Amazon EC2 service level agreement, last updated May 25, 2022, commits to a Monthly Uptime Percentage of at least 99.99% for each region (for instances spread across two or more Availability Zones) and an Instance-Level Uptime Percentage of at least 99.5% for a single instance, as of November 2026. The remedy when AWS misses is a credit on your bill.
| Monthly uptime for the region | Service credit |
|---|---|
| Below 99.99% but at least 99.0% | 10% |
| Below 99.0% but at least 95.0% | 30% |
| Below 95.0% | 100% |
Our reading, as reasoning rather than a quote: an SLA is the provider's written uptime commitment, and its remedy is a service credit, not payment for the sales you lost. Note also the gap between the two promises. A single server carries a lower commitment than a setup spread across zones. That difference is the cost of having one machine, written into the contract.
Keep Google's crawler patient while the server is gone
A long outage can follow you into search results. Google's documentation on HTTP and network errors, read in November 2026, says server errors make its crawlers slow down, and that indexed URLs "are preserved in the index, but eventually dropped" if they keep returning errors. Google gives no number of days. It also says that once the server returns 2xx again, Google "gradually increases the crawl rate for the site."
The point is to say "temporarily unavailable" on purpose instead of letting a dead server time out. The standard itself, RFC 9110, describes 503 as an overload or maintenance condition "which will likely be alleviated after some delay." Our 503 service unavailable guide covers the setup on common hosts. If the outage is a whole public website rather than a cloud app, the website down emergency page has the first fifteen minutes.
Panic moves that turn their outage into yours
Provider outages end. The permanent damage usually comes from what someone did at hour one to feel useful. These are the moves to forbid in the group chat.
- Deleting and recreating a database instance. AWS says "If you don't choose Retain automated backups when you delete a DB instance, all automated backups are deleted with the DB instance." A reset button can erase every restore point you had.
- Restoring a backup over the live database before you know what failed. You may overwrite the newest good data with older data, and lose the evidence of what went wrong.
- Failing over to a region or server nobody has ever tested. An untested standby is a second outage with a different address.
- Repointing DNS in a hurry. Cloudflare's TTL documentation notes that a longer TTL "means that updates to your records take longer to go into effect," so a rushed change can strand half your visitors on the wrong server for hours.
- Creating access keys for the AWS root user so a helper can "just get in." AWS says "We strongly recommend that you do not create access keys for your root user"; give helpers their own limited access instead.
If you cannot get into the cloud account at all because the person who set it up has left, that is a different emergency with a different first step, covered on our developer disappeared page.
After the all-clear: three settings to fix this week
When the provider posts its resolution, test the money paths yourself: logins, checkout, forms, the phone number on the contact page. Then fix the three settings that decide how bad the next outage will be. Uptime checks are first. Google Cloud's uptime check documentation describes checks sent "from multiple locations around the world," alerting policies that notify you when a check fails, and a certificate-expiry metric that can warn you before a certificate lapses.
Backup retention is second. AWS says an RDS retention period can be set "between 0 and 35 days," as of November 2026, with a default of one day through the API or CLI and seven days in the console. A restore only reaches back as far as the window set before the problem, which is our reasoning from those numbers, not an AWS sentence. Account recovery is third: AWS advises a root user email "managed by your business" that forwards to a group, so recovery does not depend on one person being reachable.
When to call us: the provider says it is fixed and your app is still down, nobody can say what runs where, or the restore options look like a trap. We respond the same business day (hours on our contact page), and cloud and server repairs are from $575, depending on complexity, with a written estimate before work starts. See emergency pricing for what moves that number and our emergency software repair page for the full offer. Ongoing care starts with hosting and maintenance from $55 a month, and redesigns of fragile setups fall under web app development. For a cloud app that is down for reasons that have nothing to do with the provider, start with the web app down checklist, or browse the emergency repair hub. DataCram is a San Antonio studio (about us) at (210) 346-0848.
Questions owners ask in an emergency
Sign in and open the AWS Health Dashboard, which AWS says is available to all customers at no additional cost and gives visibility into your own resources and accounts. If it shows nothing for your region and services, and other sites on AWS are working, the problem is almost certainly in your own server, database or configuration.
The EC2 SLA, last updated May 25, 2022, offers service credits on your bill: 10%, 30% or 100% depending on how far regional uptime fell below 99.99%, as of November 2026. Our reading is that an SLA is a credit, not compensation for lost sales. We did not read the SLA's exclusions or claim process, so check those with AWS.
Yes, if you can serve it from somewhere other than the failed server. Google's guidance on pausing a site says to return an informational page with a 503 status for short outages, use the retry-after header, and never return a 503 for robots.txt. A dead server that simply times out tells visitors and crawlers nothing.
Google says URLs that keep returning server errors are preserved in the index but eventually dropped, and it gives no number of days. When the server returns normal responses again, Google says it gradually increases the crawl rate. Keeping the outage short and serving a proper 503 is the practical defense.
It depends on what failed. AWS says it protects the infrastructure that runs its services, while customers handle what they choose to run. On EC2, that includes the guest operating system, its patches, the software you install and the security group firewall. Your data and its backups are your responsibility either way.
DataCram's server, cloud and software repairs are from $575, depending on complexity, with a written estimate before work starts. We respond the same business day. If the outage is the provider's, part of our job is telling you honestly that the right move is to wait, and helping you communicate while you do.
Keep reading
- The website offerEmergency Website Repair: Who to Call When the Site Is Down, and What Happens After You DoWebsite down and no one to call? A same-business-day reply, fixed WordPress repair prices, from $575 for other platforms, and the checks before we sign off.
- The software offerEmergency Software Repair: Triage, Containment and a Written Root Cause for Broken Custom SoftwareCustom software or a web app broken today? Triage, stop the damage, restore, find the root cause, and a written estimate first. From $575, depending on scope.
- The mobile offerEmergency App Repair: An iOS and Android Developer for Crashes, Store Trouble and Expired CertificatesApp crashing, rejected or pulled from the store? Crash triage, store fixes, certificates and the access we need, from $575 with a written estimate first.
- Not legal advice. Read anyway.Data Breach: What to Do in the First Hours, Before Anyone Wipes the EvidenceThe first hours after a data breach: contain it without destroying evidence, call your attorney, insurer and police, and know the Texas notice deadlines.
- Signing, expiredApple Distribution Certificate or Provisioning Profile Expired: What Stops, What Keeps Running, and Who Can RenewApple certificate or provisioning profile expired? See what Apple says happens to live apps, enterprise apps and push, who can renew, and the safe order.
- Search traffic emergencyWebsite Traffic Dropped Overnight: The First Hour, Before Anyone Touches the SiteTraffic fell off a cliff overnight? Rule out broken tracking, an outage, a stray noindex and a manual action first, in the order that costs least.
- Everything in Emergency RepairSee the section
From the Insights desk
- The Optimizely Hack Was Not an Outage. It Was a Delivery.On November 16, a flaw in Optimizely's sign-up flow let an attacker slip card skimmers, wallet swappers and iPhone exploits into the testing scripts of 178 organizations, the BBC and Zoom among them. How it happened, what it did, and the fixes, in order.
- Why Your WordPress Form Gets Spammed — and the Fix That Doesn't Annoy Real CustomersForm spam isn't the problem — it's what the spam does: buried leads, a poisoned CRM, and an inbox you've learned to ignore. Why CAPTCHAs make it worse, and the silent fix that stops bots without costing you real customers.
Find out whose outage this is
Tell us your provider, what is down and what changed recently. We reply the same business day with what we would check first and a written estimate if repair is needed.
Which of these sounds like you? (pick any)
Rather talk now? Call (210) 346-0848.