When your website host fails: what moving 11 sites in a week taught us

A lapsed plan on one hosting account took every site on it offline at once. What the symptom looked like, how we diagnosed it, and the questions worth asking about who holds your hosting and DNS.

On this page

When a host fails, the cause is often a billing or plan limit rather than a fault in your site, and the symptom can look like an attack or an outage. On 28 August every Workers-hosted site on one of our Cloudflare accounts started returning a "temporarily rate limited" page, because the paid plan on that account had lapsed. Nothing was wrong with any of the sites. We moved the affected sites over the following days, and we learned several things we would do differently.

What it looked like

All 14 Workers sites on the account returned an HTTP 429 response with Cloudflare's "this website has been temporarily rate limited" page. The page loaded in about a third of a second and came from Cloudflare's own network, before any of the sites' code ran. That is the first clue. A fault in a site's code is slow, or throws its own error page. A response that fast and identical across every site on an account is coming from the layer in front of them.

It was not a Cloudflare outage and it was not specific to any one site. The account's paid Workers subscription had lapsed and the account had fallen back to the free tier's allowance of 100,000 requests a day. The account was actually seeing about 597,000 requests a day, so everything past the cap was refused. Sites we host on Cloudflare Pages, a different product, kept working throughout.

How we diagnosed it, and what we would check first

The symptom looks like a site problem and invites a pointless migration. A single question separates the two: does every site on the account fail the same way? If so, pull the account's total request volume and compare it with the plan's allowance. Cloudflare's reporting interface still answered while the sites were throttled. We checked the numbers before touching any code or DNS.

If you have one site and it shows this kind of page, ask your host or developer which plan the account is on and whether its billing is current. That one question would have saved time.

The recovery, and the mistakes in it

We moved 11 of the affected sites to a different host as an emergency measure and moved one to a new Cloudflare account. The DNS records for those sites stayed in the old account, pointing at the new host. They had to stay unproxied, because turning the proxy back on would send traffic straight back through the throttled network.

Several things went wrong, none of them dramatic:

  • Settings did not travel. Secrets and environment values are held per hosting account. A move leaves them behind. Three form-bearing sites sent enquiries to the wrong place or failed outright until we audited them, covered in the contact form that says sent but delivers nothing. One of our own sites returned errors because a login secret was missing from the new account, and eleven other secrets on it were still unset when we checked.
  • Looking right was not enough. The build, the page checks and the deploy log all looked fine. Only requests to the live sites told us what was broken.
  • We deleted the old destination too early on one site. On 9 October, after rebuilding a client site onto a new Cloudflare account following a failure of its previous host, we removed the old DNS zone and the temporary host straight away. The DNS change had not finished spreading, so the client's own network still resolved the old address and showed a 404 page. We recreated the temporary host as a safety net. The rule we took from it: keep the old destination serving until the change has settled, about 48 hours, and phase migrations rather than cutting over in one step.

What else can take a site down, and what did we learn from each?

The same arrangement failed in two further ways in the weeks that followed. Both were our own incidents, and all three have one thing in common: the code was fine and the part that failed was the join between the site and the account or service it depends on.

What happens when a routine security update breaks every page?

On 12 October a security advisory was published for the framework our sites are built on, and the fix was in a newer version. We upgraded pagecog.com.au to the patched release. The site built without errors and the type checker reported nothing wrong. We deployed it, and the live site returned a server error on every page.

The likely cause was two copies of the same internal routing library in the project: our own pinned copy at one version, and a slightly newer copy that came with the framework. The build and the type checker each looked at one copy and were satisfied, and at runtime on the live hosting the two disagreed. That explanation is inferred. It fits the evidence and aligning the versions fixed the site, but we could not reproduce the broken state afterwards. The other nine sites we upgraded the same way had no pinned copy and went through without trouble. We do not have a reliable measure of how long visitors saw errors, so we will not put a figure on it.

A build shows the code can be assembled, and a type check shows the pieces fit as the code describes. Neither runs your site the way a visitor's request does, on the real hosting with the real settings. What we do now:

  1. Check for duplicate copies of the router libraries after any upgrade, and fix them before anything is released.
  2. Upload the new version without sending it live, then give it 10 percent of real traffic while the old version serves the other 90 percent.
  3. Request real pages from the live site and check the responses before promoting to everyone.
  4. Keep the way back ready. A rollback to the previous version takes seconds.

For a site with logins or a database we do not release a dependency upgrade straight to everyone. Ask whoever looks after your site three things: whether they apply security updates, whether they check the live site afterwards because "the deploy succeeded" and "your site works" are different statements, and how long a bad release would take to undo.

What happens when a domain expires with auto-renew on?

Auto-renew only works if the payment behind it goes through. On 3 October a client's .com.au expired with auto-renew switched on at the registrar, because the prepaid balance on the reseller account that pays for renewals was below the cost of a renewal. We found the site dead on 5 October. The code was fine and the hosting was running, but the domain's registry status was serverHold with the reason given as expired, so it was pulled out of DNS and every network answered that the domain does not exist. Email on the domain stopped too.

That is different from a hosting fault, which usually gives an error page or a timeout. An expired domain gives a "not found" from everywhere at once. Two details make it hard to catch from outside. Deploys keep working while the domain is dead, so nothing in a normal update flags it. And a status of "server renew prohibited" on a .au domain is a normal registry setting, not a warning. The one to worry about is serverHold.

We topped up the balance and renewed on the evening of 5 October, and the registrar moved the expiry to October 2027 and cleared the hold. Some resolvers kept answering "does not exist" for a while, because networks remember a negative answer, which for .com.au can be up to an hour. In an earlier incident, on 30 August, two domains went on hold after renewals lapsed. After renewal the .com.au zone republished their delegations about 16 minutes later, and neither site needed any change. During that outage we confirmed the sites were healthy by sending a request straight to the server's address with the domain name attached, which proved the fault was DNS only.

On the one site where we can measure it, an August expiry and restore left rankings untouched: daily impressions through the week of the hold were 231, 198, 181, 137, 127, 183 and 183, flat against the weeks either side. That is one site and one week, and visitors and enquiries were lost for those days regardless of what Google showed. Four domains we look after lapsed between August and October, one of them our own. To protect yours:

  1. Log in at the registrar and read the expiry date, not the auto-renew switch. For a .au name, a WHOIS lookup shows the status and expiry for anyone.
  2. Check what is paying. For a card, look at its expiry date. For a prepaid balance, check it covers the next renewal.
  3. Use an email address that is not on the domain for registrar notices, since the email inside a dead domain dies with it.
  4. Put the date in a calendar a month ahead with a reminder to check payment.
  5. Know who the registrant is, as covered in who actually owns your website. Changing your domain name covers the opposite case, where you move on purpose.

What to ask about your own site

  1. Who holds the hosting account, and who is billed? If the answer is someone you cannot contact, you have a single point of failure.
  2. Who holds the domain and the DNS? See who actually owns your website for the four separate accounts most sites have.
  3. What happens if the host's plan lapses? A good provider can tell you, and knows whether they would be told before you are.
  4. How quickly could the site move? If the code and content are in a repository the developer controls, a move is days. If the site lives inside a hosting product that cannot export it, it may be a rebuild.
  5. When was the last time someone requested the live site and checked the forms?

Why we say this openly

It would be easier to leave this out, since the lapse was on our side. We include it because the questions above are the ones we would want a customer to be asking about us, and because the way to show you handle a failure is to describe one. It also explains why we are consolidating our hosting onto one account that we watch, rather than spreading sites around.

If you are comparing providers, a monthly arrangement means the person hosting your site is paid for as long as it stays up, which is a reasonable reason to care. It does not make the hosting immune to mistakes, as this post shows. Our pricing covers what the base plan includes, and the web design service page describes how we build and host.

Common questions

What does 'temporarily rate limited' mean on a website?
It is a response from the hosting or network provider saying the site or account has exceeded an allowance. It is usually a plan or billing limit, not a fault in your site's code.
Who should hold my hosting account?
Whoever is responsible for the site's uptime should be able to see the account, its billing status and its limits, and you should know who that is and how to reach them.
Can my domain expire even though auto-renew is on?
Yes. Auto-renew only records that you asked for renewal, not that the payment succeeded. An expired card, a declined payment, a low prepaid balance or a renewal notice sent to an unread address can all let the domain lapse while the setting still says on.
Why did my website break after a security update?
A build and a type check do not run your site the way a visitor does. Two copies of the same library can pass both and still disagree at runtime on the live host. Release updates to a small share of traffic first and request real pages before promoting them.
How long does a DNS change take to settle?
Resolvers cache old records, so some visitors can see the old destination for up to about 48 hours after a change. Keep the old destination serving until the change has fully settled.

Want a website that turns readers into customers?

A custom site, hosting and upkeep for one monthly fee, with your leads and Google results in a dashboard you can check any time.