Home › Topics › Cloud Network Monitoring › Hybrid Path Health

Hybrid Connectivity: Monitoring the Path Between On-Prem and Cloud


The moment your first workload moved to the cloud, your network quietly grew a new segment. It immediately became one of the most important segments you own. Every login that checks the on-prem directory, every file that syncs, and every backup that comes home rides the same stretch of path. So does every app tier split between the building and the cloud. That path is the hybrid link between your premises and your provider. When it degrades, the tickets do not say "the hybrid path is degraded." They say the cloud is slow, the app is broken, and — unfairly but inevitably — that the network is bad again.


The uncomfortable novelty is that you only own one end of this path. Your edge router is yours. The provider's gateway is theirs. Everything in between belongs to an ISP or a carrier who has never heard of you. Monitoring a path you mostly do not own is a different craft from monitoring a switch you can walk to. It is about evidence per segment and baselines that predate the incident. It is about tests that measure the whole road rather than trusting any single party's dashboard.


This guide covers the two kinds of road between here and there and why a tunnel that says "up" can still be lying. It explains how to baseline latency from your building (and your branches) and how to split responsibility when things go wrong. It covers how to drill your failover on purpose. It also explains how to know — in advance — exactly which applications die when the path does.


In this guide

  • The two roads: VPN and dedicated interconnect
  • Up is not the same as healthy
  • Latency baselines from your building
  • The three segments of responsibility
  • Synthetic tests as the continuous truth
  • Failing over on purpose
  • What dies when the path dies
  • Where to go next

The two roads: VPN and dedicated interconnect


Generically, every provider offers the same two ways to connect your building to your VPC. The first is a site-to-site VPN: an encrypted tunnel from your edge firewall, across the ordinary public internet, to the provider's VPN gateway. It is quick to build — often an afternoon — and its cost pattern is modest: a managed gateway fee and the metered traffic. Its character comes from the road it rides: the public internet. Its latency and jitter vary with weather you cannot see and routing decisions made by strangers.


The second is a dedicated interconnect: a private circuit from your premises (or a carrier facility) directly into the provider's edge. It never touches the public internet. Its latency is steady, and its capacity is a number you chose. Its cost pattern is the opposite of the VPN's: a significant fixed commitment. Provisioning is measured in weeks or months because physical things must be installed. There is typically a gentler rate on the traffic itself. The interconnect behaves, in short, like a WAN circuit — because that is what it is.


AspectSite-to-site VPNDedicated interconnect
Time to buildHours to daysWeeks to months
Cost patternLow fixed, metered trafficHigh fixed commitment, steadier per-traffic pattern
Latency behaviorVariable — internet weather appliesSteady and predictable
What can hurt itAnything on the internet path; tunnel software; MTU gamesPhysical circuit faults; the carrier; the provider edge
Best roleGetting started; backup path; small sitesPrimary path once the cloud is critical

The mature pattern is both: the interconnect as the primary road, the VPN as the standby. Traffic fails over to the VPN when the circuit has a bad day. That design buys real resilience — and it quietly doubles your monitoring obligations, because a standby path that nobody watches is a standby in name only. Both roads need health checks all the time, not just the road currently carrying traffic.


Capacity deserves the same double vision. The interconnect is a circuit with a ceiling you chose. So watch its utilization exactly as you would any WAN link. Sustained high use is your early warning to resize before users feel it. The VPN's ceiling is sneakier: encryption throughput limits on your firewall and on the provider's gateway. That means the standby road is usually smaller than the primary. Knowing that number in advance — and which traffic you would shed to fit inside it — is planning. Discovering it during a failover is improvisation with an audience.


Up is not the same as healthy


Every hybrid connection gives you a state metric: the tunnel is established, the circuit is up. Monitor it — state changes are your fastest, cleanest alerts. Then distrust it, politely, forever. A tunnel can be established while the path through it is broken. Routes beyond the tunnel can be misconfigured so traffic arrives and dies. An MTU mismatch can silently eat every large packet while small pings sail through. Or return traffic can take a different, broken road home. The general version of this lesson is the tunnel that is up and lying about it. It is covered properly in the guide on monitoring site-to-site tunnels. The cloud-specific point is that both ends now publish a state metric, your firewall's and the provider's. They can disagree. When they do, that disagreement is itself diagnostic.


So the rule for the hybrid path is: monitor the state and the through-path. A through-path check is a probe that travels the whole road and touches something real on the other side. It could be a ping to a private cloud address, a TCP connection to an actual service port, ideally both. State tells you the door is open; the through-path check tells you that walking through it still works. You need both, because each fails in a way the other catches. State alerts are instant but gullible. Through-path checks are honest but cannot tell you which layer broke. Together they triangulate.


Probe in both directions if you can. A check from the building into the VPC and a check from an agent in the VPC back toward the building sound redundant. But paths can fail asymmetrically — return routing has its own opinions. A pair of checks that disagree is one of the fastest "which direction is broken?" answers available anywhere in networking.


Tip: give the MTU problem five minutes of prevention now. Tunnels add encapsulation overhead, which shrinks the largest packet that fits. A path that passes pings but stalls file copies and hangs TLS handshakes is the classic signature. Test once with large, do-not-fragment pings through the tunnel, set clamping where your firewall recommends it, and file the working numbers in your runbook. Future-you will otherwise rediscover this on a bad day, slowly.


Latency baselines from your building


"The cloud feels slow" is unanswerable without one number: what it usually feels like. That number has to be measured from where your users sit — the building — against the cloud resources they use. Measure it continuously, so that history exists before anyone needs it. A scheduled probe every minute or so from your monitoring system to a couple of private addresses in the VPC does the job:


ping 10.0.2.14   (database instance, via the hybrid path)
  reply from 10.0.2.14: time=38 ms
  reply from 10.0.2.14: time=39 ms
  reply from 10.0.2.14: time=38 ms

same check, logged every minute, graphed for months

Three probes tell you almost everything. One goes to your own edge router (proves the building side). One goes across to a cloud instance (the whole road). One goes to the same instance from inside the VPC by an agent or probe there (proves the cloud side). When users complain, the three graphs assign the slowness to a segment in about thirty seconds.


What should the baseline look like? On an interconnect: flat, boring, and tightly clustered — distance sets the floor (the speed of light remains stubbornly non-negotiable), and variation should be small. Over a VPN: a visible texture of variation, mild at quiet hours, wider at busy ones, because the internet is a shared road. Learn your typical value and your typical spread. Alert on sustained departure from your own normal rather than on any absolute number borrowed from someone else's network. That means a multiple of your usual latency held for several minutes, or loss above a small steady threshold.


And if you have branch offices, remember each site has its own road to the cloud. Some of those sites go via your headquarters and the main hybrid link, and some go straight out their local internet. A probe running at each site, reporting back centrally, is the only way "the cloud is slow in the warehouse" becomes a graph instead of an argument. This is exactly the job remote monitoring probes exist for.


Write the resulting numbers down somewhere the whole team can see — a small table like this one turns every future "is this normal?" into a lookup:


FromToRoadTypicalInvestigate when
Head officeCloud databaseInterconnectSteady, tight spreadSustained rise above the usual spread, or any loss
Head officeCloud databaseVPN (standby)Higher, visibly variableLoss beyond a small steady trickle; spread widening for hours
Branch officeCloud appLocal internet + VPNIts own number — measure itDeparture from that site's own baseline, not headquarters' baseline

The three segments of responsibility


When the hybrid path degrades, the first organizational question is "whose problem is this?" The honest answer is that the path has three segments with three different owners. Your edge is the firewall or router terminating the tunnel or circuit, plus your building's wiring behind it. It is yours to fix. The general craft of watching that device and its link is covered in WAN circuit health. The middle: your ISP or the carrier hauling the circuit — not yours to fix, but very much yours to prove things about. Their edge is the provider's gateway and everything behind it. It is theirs entirely, visible to you only through their status page and the tunnel metrics they publish among their native metrics.


A traceroute through the path makes the segments concrete. Annotated, it reads like a property map:


hop 1   192.168.1.1     1 ms     your firewall            (your edge)
hop 2   100.64.17.1     4 ms     ISP first hop            (the middle begins)
hop 3-5 carrier hops    9-31 ms  the middle, in transit
hop 6   * * *                    provider edge, silent    (their edge begins)
hop 7   10.0.2.14       38 ms    the instance, answering  (through!)

Read it with the usual traceroute caveats. Middle boxes often deprioritize or ignore probes (a silent hop is not a broken hop), and per-hop times can mislead. The reliable signals are the ends — where responses stop entirely, and what the final through-path number says. The real power move is evidence per segment, gathered continuously. Your edge device's health and interface counters prove segment one. Your own independent latency and loss measurements across the middle prove segment two, in exactly the spirit of independent circuit measurement. And the provider's own published tunnel metrics cover segment three. Suppose you open a ticket with the carrier holding a week of loss graphs, timestamps, and a clean bill of health for your own edge. In that case, the conversation is short. When you call and say "it feels slow," the conversation is long and circular, and it ends with them suggesting you reboot something.


Synthetic tests as the continuous truth


Latency probes watch the road; synthetic tests watch the errands. A synthetic test is a small scripted transaction that behaves like a user — resolve the name, open the connection, perform one honest unit of work. It runs on a schedule, from where users actually are. For the hybrid path, a sensible battery is layered. It includes a ping for the road itself and a TCP connect to each critical service port (directory, database, file service, application). It includes a DNS resolution of the cloud names your users rely on. And it includes one application-level check — an HTTP request, a small file read — per app that matters. Each layer catches what the one below cannot. Routing can be perfect while the service port is closed. The port can be open while the application behind it returns garbage.


Run them frequently enough to notice trouble in minutes. Require two or three consecutive failures before alerting so one lost packet does not page anyone. The payoff compounds during incidents. When the ping is healthy but the database connect fails, the path is fine and the problem is in the cloud. When everything fails at once, the path itself is the story. In that case, your per-segment evidence from the previous section tells you which owner to call. Synthetic results are also the record that survives the meeting afterward. "Authentication checks failed from both sites for eleven minutes" is a sentence with timestamps behind it. That beats every sentence built from memory.


Failing over on purpose


If you run the mature design — interconnect primary, VPN standby — then somewhere in your configuration lives a promise. That promise is that when the circuit fails, traffic will move to the tunnel. An untested failover is not a capability; it is a wish with paperwork, exactly like the backup nobody has restored from. The only way to know is a drill. The drill is also a monitoring exercise — arguably the best one you will run all year:


  • 1. Schedule a window and tell people. A failover drill during business hours without notice is called an outage, whatever your intentions were.
  • 2. Snapshot the baselines. Capture current latency, loss, and per-road state so "back to normal" is a comparison, not a feeling.
  • 3. Pre-check the standby. Tunnel established, routes ready, and its own through-path check passing. Discovering the standby is dead is the drill's most valuable possible outcome — better now than during the real thing.
  • 4. Fail the primary deliberately. Administratively down the interconnect at your edge, in the way your change process approves and can instantly reverse.
  • 5. Watch the monitoring, not just the traffic. Did state alerts fire promptly? Did through-path checks fail and then recover as routing converged? How many seconds of interruption did synthetic tests record? Did dependent-app checks behave as your dependency model predicted? If the path failed over so smoothly that monitoring stayed green throughout — lovely, and also suspicious: would you have known this happened at midnight? A drill your monitoring cannot see is a drill that found a monitoring gap.
  • 6. Live on the standby briefly. Latency will typically be worse over the VPN; measure how much worse, so expectations during a real event are pre-set numbers, not hopeful guesses.
  • 7. Fail back and re-verify. Failback is its own transition with its own opportunities to go wrong; watch it with the same attention.
  • 8. Write it down. Convergence time, alert behavior, standby performance, surprises, fixes filed. Next drill, beat the numbers.

What dies when the path dies


The last discipline is knowing the blast radius in advance. Hybrid environments accumulate quiet dependencies. These include authentication that checks the on-prem directory, file shares mounted across the link, and printing that transits it for reasons lost to history. They include app front-ends in the cloud calling databases in the building — or the reverse. The day the path fails is a terrible day to start discovering the list. The discipline of mapping what depends on what is covered in dependency thinking, and the hybrid link deserves to be the first arrow you draw.


Put the model into the monitoring itself. If your system understands that a dozen cloud checks all depend on the hybrid path, then a path failure produces one loud, correctly-named alert. In that case, the path failure does not produce a storm of forty tickets that buries the cause under its own consequences. And keep the model honest with reality: the failover drill is its natural test. If the drill kills an application your model said was safe, the model just taught you something. Drills that only confirm what you already knew are drills you have earned the right to run less often. The list of path-dependent applications also has an audience beyond your team. It is the plain-language answer when leadership asks what the interconnect is for. It also explains why the standby VPN — and the monitoring around both — is worth what it costs.


Where to go next


The hybrid path is the strangest segment you own: critical to nearly everything. You own one end, a provider owns the other, and strangers own the middle. The craft is to monitor it like you own all of it anyway. That means state plus through-path, baselines from every place users sit, and evidence per segment. It means synthetic errands running continuously, failover proven by drills, and a dependency map that turns path failures into one honest alert. The rest of the Cloud Network Monitoring pillar surrounds this road. It offers the VPC mental model for what lives at the far end and provider metrics for the far end's own gauges. It offers egress visibility for the traffic riding this very path home.


Put this into practice with SysPeep Monitor

Most of this guide is checks run from your side of the road. That is exactly where SysPeep Monitor lives. It offers ping and latency checks from your building to cloud addresses, plus SNMP health on the edge firewall and WAN interfaces. It offers remote probes for branch sites and dependency-aware alerting that reports one path failure instead of forty consequences. All of this comes from one web console. The free trial installs on any Windows machine in minutes.

Download Free TrialLearn About SysPeep