About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.
Health-checked DNS failover can send new visitors to a standby server when the main server stops answering its health check. It cannot move people who are already connected, it cannot copy your data to the standby, and it does not act instantly. For new connections, plan for the health-check detection delay, the remaining DNS-cache time and the time the standby needs to be ready; these can overlap, so they are not a universal sum. Existing sessions have a separate failure and reconnection timeline.
If a provider tells you that DNS failover gives you continuity, the useful follow-up is a number: how many minutes, for which users, and with what data on the standby. This guide explains where the minutes come from, using documentation from Amazon Route 53 and Azure Traffic Manager as worked examples. It is not about recovering a broken DNS service or about a planned move to a new server; it is about automatic, health-driven switching.
Why it matters
DNS, the Domain Name System, is the directory that turns a name such as a website address into the numeric address of a server. Failover at the DNS level means the directory starts giving out a different address when the first one looks unhealthy.
That is attractive because it is cheap and works for many services. It also encourages a shortcut in thinking: that failover equals continuity. Continuity of a service needs a working standby with current data, and a way for users to reach it. DNS supplies only the last of those, and only for new lookups.
Four terms are worth keeping apart:
- Backup is a stored copy of data that you can restore from.
- Replication keeps a second copy up to date, usually with a delay.
- Failover is the switch of users to the second copy or server.
- Disaster recovery is the wider plan, including who decides, how data is checked and how you return.
DNS failover only addresses part of the third item. It says nothing about the first two.
How a health-checked failover works
In the Route 53 model, you create a health check for each resource, and the service answers queries with records whose resources are healthy. In an active-passive setup, Route 53 includes only the healthy primary resources, and if all primary resources are unhealthy, it begins to include only healthy secondary resources.
Route 53's health checkers sit in many places. The documentation says you can choose the interval between checks, every 10 seconds or every 30 seconds. Each checker also applies a failure threshold, which is the number of consecutive failed checks needed to change status. Route 53 then combines the checkers: it considers the endpoint healthy if "more than 18% of health checkers report that an endpoint is healthy". The page explains the low figure as a way to avoid marking an endpoint unhealthy because of network conditions between it and a few checking locations.
Azure Traffic Manager works on the same principle with different settings. Its probing interval can be 30 seconds (normal) or 10 seconds (fast), and its tolerated number of failures ranges from 0 to 9 with a default of 3, according to the Traffic Manager monitoring page, updated 29 December 2025.
Delay one: detection
A health check needs repeated failures before it acts, so the first delay is interval multiplied by threshold, plus the time each check waits for an answer. Microsoft publishes a worked timeline that makes this concrete. With a 30-second interval, a 10-second timeout and three tolerated failures, the page says the system "then tries three more times, at 30-second intervals" and "After a fourth consecutive failure, the monitoring system marks the unavailable endpoint status as Degraded."
Our arithmetic from that timeline: the fault can begin just after a successful check, so the system may not notice for up to 30 seconds, then wait out the first failed check (up to 10 seconds) and three more intervals of 30 seconds. That is on the order of two minutes before the endpoint is marked unhealthy, and the figure is our calculation, not a figure stated by Microsoft.
Route 53 describes its checkers separately: health checkers in different data centres do not coordinate with one another, so the combined view may reach a verdict somewhat faster or slower than a single checker's schedule would suggest (our reading). The Route 53 pages we read do not give a default or range for the failure threshold, so we do not quote one.
Shorter intervals cost more. Route 53 says that choosing the fast 10-second interval incurs an additional charge, and Traffic Manager refers readers to its pricing page for fast probing. That is a pricing matter to confirm, not a recommendation.
Tighter settings also trade stability for speed. Microsoft's FAQ says a tolerated-failure value of 0 can lead to endpoints "being taken out of rotation due to any transient issues", and that with the tightest settings (10-second probing, no tolerated failures, a TTL of 0) its service can provide failovers "under 10 seconds after an endpoint goes unhealthy and a DNS query is made" (Traffic Manager FAQ, updated 5 October 2026). That is a vendor best case for one service, and the trade-off is a switch that may fire on a brief blip (our reading).
Delay two: caching of the old answer
Once the directory starts returning the new address, computers that already hold the old answer keep using it. Each DNS answer carries a time to live, or TTL. RFC 1035 (November 1987) defines it as "the time interval that the resource record may be cached before the source of the information should again be consulted". A recursive resolver, which is the server that your office or internet provider runs to look names up for you, may hold the answer for that long.
Route 53's advice for records tied to health checks is: "specify a TTL of 60 seconds or less so clients respond quickly to changes in health status." The same page warns that a longer TTL reduces queries and cost but means a change takes longer to take effect.
Microsoft's timeline puts this in sequence: after the endpoint is marked Degraded, "previous DNS responses that include this endpoint might still be cached by recursive DNS servers and DNS clients" and "Clients continue to use the endpoint until the DNS cache expires." In their example the TTL is 30 seconds.
Two limits on the TTL as a promise (RFC 8767, March 2020):
- RFC 2181, which that document quotes, says "The TTL specifies a maximum time to live, not a mandatory time to live." A TTL is an upper bound on how long a cache may hold an answer.
- Resolvers may cap very high TTLs. RFC 8767 describes "a clamp on the value of TTLs as received from authoritative servers" and says values "SHOULD be capped on the order of days to weeks, with a recommended cap of 604,800 seconds (7 days)". That is a maximum cap, so it does not change a low TTL such as 60 seconds; it does not say what any particular resolver's cap is, and it is not a source for any minimum TTL being enforced.
RFC 8767 also describes serving stale data, where a resolver uses an expired answer when it cannot refresh it. It says that "Stale data is used only when refreshing has failed", so this concerns the case where the authoritative servers cannot be reached, not the normal health-driven change. Whether browsers, operating systems and applications add their own caches beyond this is not covered by the sources we read; in our understanding some do, so measure rather than assume.
A separate timeline: connections that are already open
DNS is consulted when a client looks up a name, not while a connection is in use. Microsoft states it directly: "Because Traffic Manager works at the DNS level, it cannot influence existing connections to any endpoint." It adds that other endpoints might continue to receive traffic via existing connections until those sessions are terminated, and recommends that applications limit session duration so traffic can drain.
For a failing server, our reading is that an open connection will usually break rather than drain, and the application must reconnect. What happens next depends on the application: it may ask the user to sign in again, lose unsaved work or retry silently. Whether a session survives on the standby depends on where session data is kept, which DNS does not control.
What DNS failover does not do
- It does not copy data. If the standby has an old or empty copy, users are switched to old or empty data. Replication or restoration is a separate job.
- It does not check the whole application. Route 53 says that HTTPS health checks "don't validate SSL/TLS certificates", so an expired certificate does not by itself fail the check. Whether the check exercises a sign-in or database call depends on how the check is built (see the related topic on synthetic checks).
- It can mask a bad health check. Route 53's documentation describes a last-resort mode: "when all records are considered unhealthy, the Route 53 algorithm reverts to considering all records healthy". Microsoft describes a similar best-effort behaviour and warns that it can look as though routing is working when failover cannot happen. A check that fails everywhere may therefore leave traffic where it was.
- It does not choose a recovery order. If the main site returns, users go back according to the same health logic. Whether that return is wanted is a decision for people.
Choosing and agreeing a design
The following is our recommendation, not provider guidance.
- Write the timelines separately. For new connections, list the health-check detection delay, the remaining DNS-cache time and the time for the standby to become useful, and note which can overlap; for existing sessions, write the failure and reconnection timeline on its own. If a provider cannot give each item, the delay is unknown.
- Decide what the standby holds. State the age of its data, who keeps it current, and whether applications are installed and licensed.
- Check what the health check tests. A simple page that returns success proves less than a sign-in followed by a data read.
- Set the TTL deliberately. Lower TTLs speed changes and raise query volume. Agree the value in writing.
- Plan for open sessions. Decide what users will see when their connection breaks and whether work can be lost.
- Test the switch. Take a planned, announced failure of the primary in a window that suits the business, and time each phase.
Read-only checks
These commands look up public DNS information and change nothing. They were not tested. In a Linux shell, dig, from the BIND 9 manual pages, accepts +noall and +answer to print only the answer section:
dig +noall +answer www.example.com A
In PowerShell, Resolve-DnsName takes -Name, -Type and -DnsOnly:
Resolve-DnsName -Name www.example.com -Type A -DnsOnly
Replace the example name with your own. In our understanding the TTL shown by a lookup through a resolver is the time remaining in its cache and counts down between repeated queries, whereas asking the zone's own name server (using dig's @server form or Resolve-DnsName -Server) shows the configured value. The pages cited do not describe that output, so confirm it on your systems.
Questions to put to your provider
- What does the health check request, how often, and how many failures trigger a switch?
- What is the TTL on the failover record today, and who can change it?
- Which locations or networks do the health checks run from?
- What data does the standby hold, how old is it, and who confirms it is usable?
- What happens to users who are signed in when the switch occurs?
- If every health check fails, what does the DNS service return?
- How and when was the failover last tested end to end, and what times were measured?
A worked example (illustrative)
The following is invented. It is one specified, serial, illustrative scenario only: it assumes the old answer stays cached for the full TTL after detection, which is the worst case, and real timelines may overlap. A small firm hosts its customer portal with a health-checked DNS record. The check runs every 30 seconds and acts after three consecutive failures. The record's TTL is 300 seconds, the default left in place by the setup. The standby server holds a copy of the portal restored from last night's backup. The main server fails at 10:00. The check notices at about 10:02. Some users' resolvers hold the old answer until about 10:07, and those users see errors until then. Everyone who reaches the standby sees data as it was at last night's backup, and anyone mid-order has lost it. The firm lowers the TTL to 60 seconds, agrees to replicate the database every 15 minutes, and adds a sign-in step to the check.
Read next
Related articles you may want next.
- HostingSynthetic sign-in checks for a hosted application: what a probe provesWhat a scripted sign-in test catches that server monitoring misses, where it falls short, and how to choose one user journey to probe.Read article
- HostingWhat monitoring should a managed hosting customer expect?What a managed hosting customer should ask a provider to define behind 24/7 monitoring: items watched, who is alerted, and what action follows.Read article
- HostingPlanning identity, DNS and application cutovers for hosted servicesA staged plan for hosted identity, DNS and application changes, with a reporting-API example, fresh-connection tests and rollback gates.Read article
Risks and limits
- Our timeline figures are estimates built from documentation, not measurements of any real service.
- We read Route 53 and Traffic Manager documentation. Other DNS services differ in intervals, thresholds and defaults.
- Resolver, browser and operating system caching beyond the TTL is not sourced here.
- Health checks from outside cannot see internal faults, and checks that pass may not mean the application works.
- DNS failover is not a substitute for backups, replication or a recovery plan.
Next step
Measure the TTL on your important records and ask for the health-check interval and failure threshold in writing, then work out the new-connection timeline from them, noting which parts overlap, and ask separately what existing sessions will do. If you are reviewing a hosted network design, the Virtual Private Cloud networking page is a starting point for that conversation; ask for the scope in writing.
Sources
- How Amazon Route 53 determines whether a health check is healthy, Amazon Route 53 Developer Guide, Amazon Web Services, date unavailable.
- Active-active and active-passive failover, Amazon Route 53 Developer Guide, Amazon Web Services, date unavailable.
- Values specific for failover records, Amazon Route 53 Developer Guide, Amazon Web Services, date unavailable.
- Values that you specify when you create or update health checks, Amazon Route 53 Developer Guide, Amazon Web Services, date unavailable.
- How Amazon Route 53 averts failover problems, Amazon Route 53 Developer Guide, Amazon Web Services, date unavailable.
- Azure Traffic Manager endpoint monitoring, Microsoft Learn, last updated 29 December 2025.
- Azure Traffic Manager - FAQ, Microsoft Learn, last updated 5 October 2026.
- RFC 1035: Domain names - implementation and specification, IETF, November 1987.
- RFC 8767: Serving stale data to improve DNS resiliency, IETF, March 2020.
- Resolve-DnsName, Microsoft Learn, date unavailable.
- BIND 9 manual pages (dig), Internet Systems Consortium, version 9.20.30 as served, date unavailable.
All accessed 15 November 2026.
Learning notes
- DNS failover changes the answer to new lookups. It does not move open connections.
- For new connections, plan for detection delay, remaining DNS-cache time and standby readiness; these can overlap. Existing sessions have a separate failure and reconnection timeline.
- A low TTL helps but is a cap, not a promise.
- A standby with old data is a failover to old data.
- Ask what the health check actually tests.
This article was originally published on .