Convergence: How Long Until the Network Agrees Again
Every outage has one question at its center, asked by managers, users, and your own stomach: "how long until it's back?" On a network with redundant paths, the honest answer is a word that gets thrown around like it should be obvious: as long as convergence takes. It sounds like physics. It is actually something much more relatable — the time it takes a group to agree on new facts.
Convergence is the process of every router updating its map after a change, until they all agree again. Between the failure and the agreement there is an in-between period where different routers believe different things. During that period the network does strange, temporary, occasionally alarming things. Here is the claim this guide is built on: a redundant network fails over exactly as gracefully as its convergence — no more. The spare circuit, the second router, the backup path: all of it is only as good as the seconds of disagreement it takes to start using them.
By the end of this guide you will know what actually happens during those seconds. You will know how to measure your own network's convergence with nothing fancier than pings and timestamps. You will know how to spot the classic trap — the failover that will not fail over — before an outage finds it for you. You will know how to test failover on purpose without a career-limiting event. And you will know what to honestly promise the business.
In this guide
- What convergence means
- What happens during those seconds
- How long is "normal"?
- Measuring real convergence with pings and timestamps
- Before-and-after route table comparison
- The classic trap: the failover that won't fail over
- Testing failover on purpose
- Setting expectations with the business
- Where to go next
What convergence means
Quick grounding from the plain-language routing guide: routers keep maps of the network, and dynamic routing protocols keep those maps synchronized by passing news between neighbors. When a link or a router fails, the routers closest to the failure notice first and tell their neighbors. The news ripples outward while every router redraws its own map. The network is converged when the ripple is done — when no router is still acting on old news. Think of a caravan of friends driving to the same place when the highway closes. The first driver to see the barricade texts the group. But until everyone reads the text, some cars are still cheerfully driving toward a closed road. Convergence is the time between the barricade and the last driver getting the message.
Two reframings will save you confusion later. First, convergence is not an exotic failure mode — it is the system working. It happens on every change, planned or not: a failed circuit, yes, but also a rebooted switch, a new subnet, a maintenance window. The goal is never zero convergence; the goal is convergence that is rare, fast, and complete. Second, convergence gone chronic has a name you have already met. A network that converges over and over because something keeps flapping is the churn described in the route stability guide. This guide is about a single, honest convergence event — and how long it takes.
What happens during those seconds
A convergence event is four phases in a row, and naming them matters because they fail differently and are tuned differently:
- 1. Detection. A router realizes something is gone. If the physical link drops, detection is nearly instant — the port electrically knows. But the failure may be silent — a device farther down the road dies while your link stays up. In that case, detection waits for the heartbeat mechanism: missed hellos, then an expired hold timer. That can take tens of seconds with conservative default timers, and in many real networks detection is the longest phase by far.
- 2. Announcement. The routers next to the failure tell their neighbors, who tell theirs. The news itself travels fast — typically well under a second across a small network.
- 3. Recalculation. Each router reruns its path math against the updated map. Modern hardware does this quickly; very large maps take longer, and an overloaded CPU takes longer still.
- 4. Installation. The recalculated answers are pushed down into the part of the router that actually forwards packets. Only now does traffic really move to the new path.
In between phase one and phase four, routers disagree, and disagreement has two famous symptoms. Micro-loops: router A has the news and reroutes via B, but B has not heard yet and routes back via A. So packets ping-pong between them until both are updated. A built-in hop limit eventually discards each looping packet, which is why loops waste capacity instead of melting the network. Blackholes: traffic gets forwarded to a router that has already thrown away the old route but not yet installed a new one. The packets are simply, silently dropped. Users experience all of this as a brief dead moment: pages spin, calls garble or drop, file transfers stall then resume. Every bit of it is normal and temporary. The only question — the one this guide teaches you to answer with a number — is whether "temporary" means one second or forty-five.
How long is "normal"?
Honest answer: it depends, and anyone who quotes you one number for all networks is selling something. The mechanics set the range. Best case, a failure that is electrically detected on capable gear with a small map can converge in well under a second — faster than users can perceive. Worst case, a silent failure waiting out conservative hold timers, followed by recalculation on a busy router, can stretch to tens of seconds. Same protocol, same vendor, wildly different outcomes — because timers, topology, map size, and where the failure happens all weigh in. Even on one network, failing the primary circuit may converge in two seconds while failing a campus link takes twenty.
One asymmetry is worth knowing because it explains a pattern you will see in your own measurements: planned changes usually converge faster than crashes. A router being gracefully shut down or reconfigured announces its departure — "stop counting on me". So the network starts redrawing immediately, skipping the detection phase entirely. A router that loses power announces nothing; its neighbors must wait out the hold timer before they even begin. A device that says goodbye is cheap; a device that vanishes is expensive. This is also why your maintenance windows tend to feel smoother than your outages — it is not luck, it is phase one.
Which leads to the only practical stance for a working admin: stop trying to predict it and go measure it. The number that matters is not the theoretical convergence of your protocol. It is what your network does when that specific link dies, measured from a user's seat. Happily, measuring it requires nothing you do not already own.
Measuring real convergence with pings and timestamps
The principle: experience the failover the way a user would, but with timestamps. Run a continuous ping — one per second — from a user-ish vantage point to a destination on the far side of the failover. Start it before the failure. And let it run through the event. The gap in replies is your convergence time, measured end to end, detection through installation, with no protocol knowledge required:
10:04:11 reply from 10.44.0.10 time 8 ms 10:04:12 reply from 10.44.0.10 time 8 ms 10:04:13 no reply <- primary path failed here 10:04:14 no reply ... 10:04:24 no reply 10:04:25 reply from 10.44.0.10 time 31 ms <- backup path carrying 10:04:26 reply from 10.44.0.10 time 31 ms Result: 12 seconds of loss, end to end.
Three refinements make the measurement trustworthy. Run pings from two or three vantage points at once — convergence is not uniform, and the branch office may see a longer gap than the server room. Note the response time after recovery. In the example, replies came back at higher latency, which is the backup road introducing itself. That is independent confirmation that traffic actually moved rather than the primary blinking back. And accept the resolution: one ping per second answers seconds-scale questions, which is what small networks are asking. If your design targets sub-second failover, you will eventually want purpose-built measurement, but the principle stays identical — a timestamped stream of traffic, interrupted, tells you the truth.
The cheapest habit in this whole pillar: never touch a redundant path without a ping running. Every planned circuit swap, firewall upgrade, or provider maintenance on a redundant pair is a free convergence measurement — if the ping was started first. Admins pay for lab gear to learn numbers their own change windows would have handed them for nothing.
While the pings run, keep one eye on your monitoring console, because a failover test is secretly two tests. The network is being examined — but so is the monitoring itself. Did the neighbor-down alert fire, and how quickly? Did the log message arrive with the right reason attached? Did the dependency rules keep it to one clean notification instead of a storm? Suppose a failover drill ends with "the network converged in nine seconds and monitoring saw all of it." Then it has verified the entire early-warning chain you will be depending on during the real event. If monitoring stayed silent, you just found that out for free, on a Tuesday, in a window.
Before-and-after route table comparison
Pings tell you how long; route tables tell you whether it did the right thing. Snapshot the routing table on your key routers at three moments — before the failure, during it, and after restoration — and compare. Before is your baseline. During answers the important questions. Did the backup routes actually appear? And is the default route — the "everything else goes here" entry — still present and pointing somewhere alive? After answers the question everyone forgets to ask:
before: 0.0.0.0/0 via 203.0.113.1 (primary) during: 0.0.0.0/0 via 198.51.100.1 (backup) <- failover worked after: 0.0.0.0/0 via 203.0.113.1 (primary) <- failback complete
That "after" line is the sneaky-valuable one. Plenty of networks fail over beautifully and then never fail back. The primary is repaired, but the map keeps preferring the backup. The network runs for months on the slow road or the metered one with nobody the wiser. A before-and-after comparison catches it on day one. The mechanics of capturing and diffing tables on a schedule — so this comparison happens automatically, not just on test day — are covered in watching the routing table.
The classic trap: the failover that won't fail over
Now the trap this guide exists to spare you. The scene is always the same. The primary circuit dies, everyone stays calm because "we have a backup," the promised seconds pass, then minutes, then the room gets quiet. The backup — installed years ago, paid for monthly ever since — never takes over. It existed as hardware and an invoice, but not as a route. Nothing in the network's map ever actually offered it as a path. The day of the outage is the day this is discovered.
The ways it happens are depressingly consistent. The backup route was configured but a filter quietly blocks it from ever being advertised. The backup's neighbor session has been down for months — nobody monitored the standby, because standby things are boring. The backup works in your direction, but the far side never learned a return path. So your packets go out and the replies have no road home — failover has two directions, and one of them was never wired. Or a manually configured fallback route sits with the wrong preference value, permanently one notch too unattractive to ever install. Every version has the same shape: the configuration existed; the route never did.
Two defenses, and you need both. First, monitoring: treat the backup path's neighbor session and route presence as first-class checks with real severity, exactly as the neighbor state guide prescribes. "Redundancy lost" must be an alarm, not a diary entry. A backup that broke quietly in spring is a trap that springs in autumn. Second, testing: a backup that has never carried traffic is a rumor with a purchase order. The WAN-side version of keeping backups honest — circuits, providers, and edge routers — is covered in failover readiness; the on-purpose test below is the routing-side companion.
Testing failover on purpose
Deliberately breaking your own network sounds like a prank, but the logic is simple: the failover will be tested either way. The only choice is whether the test happens in a scheduled window with you holding a rollback plan, or during a Tuesday outage with the sales team watching. Here is a runbook shape that keeps it boring:
- 1. Book a window and announce it. A real maintenance window, communicated to stakeholders, with the honest caveat that brief blips are the point of the exercise.
- 2. Write the rollback first. Write the exact steps to restore the primary. Have a management path to your routers that does not depend on the link you are about to break — console access or an out-of-band path. Test that access before you start.
- 3. Snapshot the baseline. Route tables and neighbor lists on the key routers, saved to files.
- 4. Start the pings. Continuous, timestamped, from two or three vantage points, to destinations beyond the failover point.
- 5. Fail it realistically. Shutting the interface down in software is the gentle version — it is detected instantly, so it measures your best case. Pulling the actual cable or powering the actual circuit equipment off is closer to real life. Know which one you tested and label the result accordingly.
- 6. Observe and record. Length of the ping gap from each vantage point, the during-failure route snapshot, confirmation that the backup path is genuinely carrying traffic.
- 7. Restore, and watch the failback. Bring the primary back, confirm the map returns to baseline, and note how long that took — failback has a convergence time too.
- 8. Write it all down. The measured numbers, the surprises, the date, and any fixes filed. This document is now the source of truth for what "redundant" means on your network.
Warning: test one assumption at a time, and never schedule a failover test when a failed rollback would be unsurvivable. The afternoon before payroll runs is not the moment. If step 2's independent management path does not exist, building it is your prerequisite project. A test that can lock you out of your own routers is not a test; it is a hostage situation.
Setting expectations with the business
The word "redundant" performs a magic trick on non-technical ears: it transforms into "seamless." Nobody says seamless out loud, but it is what gets heard. The gap between seamless and reality is where trust evaporates during the next incident — even when the network did everything right. Your measured number is the antidote. The honest message sounds like this: "If a circuit fails, expect roughly a ten-second pause. Calls in progress may drop, some sessions will ask you to sign in again, and file transfers will resume on their own." That message, delivered before an incident, converts a future "the network is broken" into "ah, this is the ten seconds they mentioned."
It helps to translate the blip per activity, because "brief loss" lands differently depending on what someone is doing when it hits:
| What users are doing | During the blip | After convergence |
|---|---|---|
| Browsing, email | A spinner, maybe one retry | Back to normal; most people never notice |
| Voice or video call | Garble, freeze, or a dropped call | Redial works; quality may differ on the backup road |
| Large file transfer | A stall | Resumes on its own; slower if the backup is thinner |
| VPN and remote sessions | May disconnect | Reconnects automatically, or asks for a sign-in |
A measured number also changes your credibility in the room. "We tested failover in a window on the night of the last maintenance; loss of the primary circuit costs nine seconds" is a sentence with entirely different weight than "it should be quick." The measured number may turn out to be unacceptable to the business. Perhaps forty seconds of blackout during a circuit failure genuinely will not do. Then you have a funded design conversation held on a calm afternoon with data, instead of a shouting match held during an outage with none. Timers can be tuned and designs can be changed, but only a measurement tells anyone whether they need to be.
Where to go next
Convergence in one breath. After any change, the network takes real time to agree on the new map. During that time expect loops, blackholes, and a blip. The length of the blip is a measurable, testable property of your network. And the deadliest failure mode is the backup path that was never really on the map at all. Measure in a window, monitor the standby like it matters — because it is the thing you will be standing on next.
From here, watching the routing table turns the snapshot-and-compare trick into a scheduled, automatic habit. The rest of Routing Protocol Health covers the heartbeats and stability signals that tell you a failover is coming before it arrives.
Put this into practice with SysPeep Monitor
A planned failover test goes better with monitoring watching alongside you. SysPeep Monitor's live dashboards show device and interface state as you pull the plug. Its SNMP polling records the before-and-after numbers. And its alerts confirm which checks caught the event and how fast. Set up the free trial on a Windows machine before your next window.
Download Free TrialLearn About SysPeep