Limited-time offer on Hong Kong Dedicated Servers for our Double 11 Carnival: Enter THXPROMO for 50% off the first two months.
Varidata News Bulletin
Knowledge Base | Q&A | Latest Technology | IDC Industry News
Varidata Blog

Fix Partial Outages After Server Switch

Release Date: 2026-11-17
DNS cache splitting traffic after server migration

You finally moved your site to a shiny new US server, flipped the DNS records, everything looks green in your monitoring… and then tickets roll in: “Site is down in city X”, “Mobile network Y can’t open the homepage”, but your synthetic checks say it’s fine. This kind of split-brain outage is almost always a DNS caching side effect, especially when DNS TTL, DNS cache, DNS propagation, server migration, zero downtime are all part of your deployment playbook.

Why Only Some Regions Break After a Server Move

When you switch servers, you are not just moving CPU cycles; you are asking the entire global recursive DNS layer to update its opinion about where your domain lives. That layer is aggressively cached, weirdly heterogeneous, and absolutely does not update everywhere at once. The result is a world map where some autonomous systems still talk to your old box, while others have faithfully followed the new IP.

  • Part of China or Europe still resolves your old IP, which might already be shut down.
  • Other regions resolve the new US server correctly and see a healthy site.
  • Some mobile networks sit in the middle: DNS changed, but old TCP sessions linger.

From a user’s perspective this feels random: neighbors on a different ISP can open the site, but they can’t. From an engineer’s perspective this is the natural outcome of DNS caching being designed for resilience, not synchronized cutovers.

The Minimal DNS Theory You Actually Need

You don’t need to memorize the RFCs to debug this, but you do need a mental model of how the lookup stack works and where Time To Live (TTL) fits in. Think of DNS as a distributed readonly key–value store, with “eventual consistency” and a lot of intermediate caches you do not control.

  1. A user’s device asks its stub resolver to resolve example.com.
  2. The stub resolver forwards the query to a recursive resolver (ISP DNS or a public resolver).
  3. If the recursive resolver has a cached answer that has not expired, it returns it immediately.
  4. Otherwise it walks the DNS tree until it reaches your authoritative DNS and gets the answer plus a TTL.
  5. The recursive resolver caches that answer for the TTL duration and serves it to many users.

TTL is literally the “do not bother me again before this time” hint to the recursive resolvers. If you announce a 3600-second TTL, you are suggesting that resolvers should trust the IP for up to an hour before asking again. That is brilliant for stability, until you change the IP and realize half the planet still trusts the old answer.

Recognizing a TTL Cache Problem in the Wild

Before you start rebooting the new server or rolling back the deployment, you should prove or disprove the “stale DNS cache” hypothesis. The symptoms are distinctive if you know where to look.

  • Outages are geographically or ISP-specific instead of global.
  • International users might be fine while some local ISPs fail, or vice versa.
  • Users report different IPs when they run ping or nslookup from different networks.

The more asymmetric and inconsistent the reports are, the more likely this is a DNS propagation and TTL issue rather than a single-server failure. When you host on US infrastructure, that asymmetry is also amplified by long-haul links and carrier routing policies.

Using Global DNS Checkers Like an Engineer, Not a Tourist

Online “global DNS checker” tools are your first microscope. Most engineers open them once, see a wall of green ticks, and move on. Used properly, they are a time-series debugger for your name records.

  1. Query your A or CNAME record from multiple regions.
  2. Record which locations still return the old IP versus the new IP.
  3. Repeat the same query every few minutes and watch the map converge.

If several locations keep returning the old IP long after the theoretical TTL should have expired, you have two possibilities: either caches really are still sitting on stale data due to nonstandard behavior, or you have more than one DNS zone or record path and updated only part of the hierarchy. In both cases, this is not an HTTP problem; it is a control-plane drift problem.

Terminal-Level Debugging: dig, nslookup, and Friends

Global tools are convenient, but you want ground truth from the actual networks where users are stuck. That’s when you start asking colleagues, friends, or remote servers to run CLI lookups for you.

  • dig yourdomain.com from several vantage points.
  • nslookup yourdomain.com on Windows boxes behind different ISPs.
  • Compare the “ANSWER SECTION” IP and the “Server” line to see which resolver was used.

When the same domain simultaneously resolves to two or three different addresses depending on the resolver, you are looking at staggered cache expiry. If some resolvers hit your new US server while others still cling to the old machine, you get the familiar partial outage story: error tickets from region A, “works for me” from region B, and a very confused status page.

The Ideal Play: Pre-Lowering TTL Before the Cut

If you are lucky enough to be reading this before your next migration instead of during an incident, the single strongest tool you have is a boring one: reduce TTL in advance. This is the closest thing to a cheat code in DNS operations.

  1. 24–48 hours before you plan to switch servers, lower TTL on the relevant records.
  2. Wait at least one full cycle of the old TTL so legacy caches expire.
  3. Verify current live TTL values via dig or your DNS provider UI.
  4. Only then update the IP to point to the new US server.

That two-step approach lets the ecosystem settle on the new, shorter TTL before you ship the dangerous change. When you finally flip the IP, the maximum time any recursive resolver will keep the old answer is limited by that lower TTL, not by the large value you might have inherited from a previous era.

Already Switched, TTL Still Long: What Now?

Reality is messy: often the IP has already been changed and the TTL on the record is still something like 7200 seconds. You cannot retroactively tell resolvers to forget what they cached, but you can at least stop the bleeding and tighten future behavior.

  • Immediately lower the TTL on the impacted records to a smaller but sane value (e.g., 300–600 seconds).
  • Keep the new IP stable; do not ping-pong between addresses trying to “force” refresh.
  • Communicate a realistic propagation window to stakeholders (for example, up to a few hours).

Lowering TTL after the fact is not magic; resolvers holding the old record with the original long TTL will keep serving it until that initial timer runs out. However, any resolver that refreshes from now on will get the new IP and the shorter TTL, which reduces the lifetime of future inconsistencies.

Emergency Escape Hatches: Alternate Hostnames and Direct IPs

Sometimes business reality does not care about DNS theory: a payment provider in a specific country cannot reach you, and “wait two hours” is not a viable answer. In those situations you reach for ugly-but-effective emergency paths.

  1. Create a temporary hostname that has never been cached before, pointing directly to the new server.
  2. Share that temporary hostname or a short URL with affected partners or internal teams.
  3. Optionally expose a read-only or reduced feature set there to minimize risk.

Because this hostname is fresh, there are no stale caches clinging to the old IP. It cuts directly to the new US infrastructure. Yes, this is operational duct tape, not architecture, but it keeps money flowing while the main hostname catches up with the new reality. Just remember to retire these shortcuts once the main domain is healthy everywhere.

When It’s Not DNS: Line Quality to the US Server

Engineers love the meme “it’s always DNS”, but when your origin is in a US data center and users are far away, sometimes it is just the network between them. DNS might be fully converged while HTTP or TCP traffic silently falls off a cliff.

  • High packet loss or jitter on certain carriers toward the US region.
  • Firewall rules on the new server that treat some geographies differently.
  • CDN configuration that still points to the old origin in some POPs.

The trick is to separate name-resolution issues from path-quality issues. If everyone resolves to the same IP but only some paths show timeouts or huge round-trip times, you are looking at routing and connectivity, not DNS cache weirdness. Tools like traceroute and distributed HTTP checks become more relevant than global DNS viewers in that scenario.

Hosting vs Colocation: Different Failure Surfaces

If your site runs on a managed hosting platform, many network and DNS layers are abstracted away. You might only control the apex domain records, while the provider manages internal health checks and failover. In a colocation setup, more knobs belong to you: BGP announcements, custom resolvers, even in-rack caching infrastructure.

  • On hosting platforms, confirm how their DNS and load-balancing behave during migrations.
  • On colocation setups, audit your own authoritative nameservers and any DIY resolvers.
  • In both cases, track which component is the source of truth for each hostname.

Partial outages after a server switch are often cross-layer bugs: the DNS change happens in one system, the firewall rule lives in another, and traffic policies sit in a third. If you cannot draw the full request path on a whiteboard—from browser, to resolver, to authoritative DNS, to load balancer, to the US machine—you will struggle to localize and fix subtle propagation artifacts.

TTL Strategy as Code, Not Tribal Lore

Most organizations treat TTL like a random integer you punch into a dashboard when creating a record. That is how you end up with one subdomain at 120 seconds, another at 28800 seconds, and absolutely nobody remembering why. Mature teams treat TTL values as first-class configuration, versioned and reviewed like application code.

  1. Define a small set of allowed TTL tiers and what each tier means operationally.
  2. Codify those tiers in your infrastructure-as-code or DNS management scripts.
  3. Require justification in code reviews when someone wants a very low or very high TTL.

By constraining the TTL palette, you make it easier to reason about worst-case cache durations during a migration. You also make post-incident analysis less forensic: instead of guessing what the TTL might have been at the time of the switch, you can check the config history and reconstruct the upper bounds on inconsistency.

A Practical Migration Flow for Your Next Server Switch

Let’s turn this into a concrete runbook you can reuse. Assume you are moving the main site to a new US host and care about minimizing the “some regions can’t open the site” window.

  1. T-minus 7 days: bring up the new environment, sync data, and run end-to-end tests on a hidden hostname.
  2. T-minus 48 hours: lower TTL for the production hostname to a short-but-safe value.
  3. T-minus 24 hours: verify via dig that recursive resolvers now see the shorter TTL.
  4. Cutover hour: update the production record to point to the new US server.
  5. First 6 hours: monitor global DNS, HTTP uptime, and region-specific metrics.
  6. After stability: raise TTL back to a medium value to reduce unnecessary churn.

This flow does not rely on fancy tooling, only on discipline and observability. The payoff is that when someone from a specific region claims “the site is still down”, you can quickly tell whether that is plausible under your TTL policy or whether you should investigate routing, content differences, or browser-level caching instead.

What About SEO and Search Crawlers?

Search engines live behind DNS infrastructure too. When you switch servers and caches disagree, crawlers will also see that split brain. The good news: they are usually more tolerant of transient failures than humans are, as long as you do not make instability a habit.

  • Keep your main hostname stable; avoid bouncing between multiple origins.
  • Minimize 5xx responses during migrations; serve maintenance pages only if necessary.
  • Do not change hostnames purely for SEO reasons during a cutover.

Search crawlers will gradually update their view of your IPs as DNS converges. Your main SEO risk is not the short-term propagation lag; it is extended downtime or repeatedly flaky behavior that trains the crawler to distrust the host. A solid TTL policy and careful cutovers make your site boring from the crawler’s perspective, which is exactly what you want.

Image Alt Text Example for This Topic

If you embed a diagram in this post—say, a flow from user to resolver to old and new servers—the alt attribute should describe its function, not jam in keywords. Here is an example that stays compact and useful:

Diagram showing DNS cache causing split traffic during server migration

Bringing It All Together Without the Magic

Partial regional outages after a server switch look mystical from the outside, but the mechanics are brutally simple: you published a new answer and an entire planet of DNS caches had to decide when to believe you. With a little discipline—treating TTL as code, running pre-cutover rehearsals, and debugging like a network engineer instead of only at the HTTP layer—you can make future migrations to US infrastructure feel more like routine maintenance than crisis response. The same boring tools that solve this once will keep helping every time DNS TTL, DNS cache, DNS propagation, server migration, zero downtime show up in your change calendar.

Your FREE Trial Starts Here!
Contact our Team for Application of Dedicated Server Service!
Register as a Member to Enjoy Exclusive Benefits Now!
Your FREE Trial Starts here!
Contact our Team for Application of Dedicated Server Service!
Register as a Member to Enjoy Exclusive Benefits Now!
Telegram Teams