Asymmetric Routing: When 'the' Path Is Really Two Paths
Every network conversation is a round trip: your packets travel out to the destination, and its replies travel back to you. Here is the part almost nobody tells a new administrator: those two journeys are under no obligation to use the same road. Your request can leave through one provider and cross the country on one route while the answer comes home through a different provider on a completely different route. And everything still works. Usually.
This is asymmetric routing, and it explains a whole family of mysteries. Those include traceroutes that look clean while users suffer, and latency that no hop will admit to. They include transfers that crawl in one direction and fly in the other, and firewalls that drop perfectly legitimate replies. This guide covers why asymmetry exists (spoiler: on the internet it is the default, not a malfunction). It covers why your favorite path tool only ever shows you half of it, and what it does to your measurements and your stateful devices. It covers how to detect it from your side of the fence. And it covers how to decide when it is harmless scenery versus the actual problem.
In this guide
- Two roads, one conversation
- Why networks route asymmetrically on purpose
- Why traceroute shows only half the story
- What asymmetry does to your measurements
- The stateful-firewall trap
- Detecting asymmetry from your side of the fence
- Mapping and documenting both directions
- Fine or fault: judging what you find
- Where to go next
Two roads, one conversation
Think of driving downtown in a city full of one-way streets. The route to the restaurant and the route home share a start and an end and possibly nothing else. Nobody considers that broken. It is just how the streets are laid out. Networks work the same way, for a structural reason: routing decisions are made hop by hop, independently, in each direction. Every router along the way consults its own table and picks the next step toward the packet's destination. The routers your outbound packet consults are one set of decision-makers. The routers the reply consults on its way back are a different set. They are owned by different organizations, optimizing for different things. No referee anywhere compares the two itineraries or cares whether they match.
Why does nothing break by default? Because the packet layer genuinely does not care. Each packet is a self-contained delivery with a destination address on the front, handled on its own. There is no rule anywhere that a reply must retrace the request's steps. The conversation layers above — TCP and friends — only check that packets arrive, not which road delivered them. Two letters between the same two houses can ride entirely different mail trucks, and neither correspondent will ever know.
So when you say "the path to the CRM," you are using convenient shorthand for two paths. Those are the outbound road and the return road, which happen to share endpoints. Most days the shorthand is harmless. The days it stops being harmless are what the rest of this guide is about.
Why networks route asymmetrically on purpose
Asymmetry is not sloppiness; it falls straight out of how the internet's economics work. A few forces produce it constantly:
- Hot-potato handoffs. When one provider must deliver traffic to another, it usually hands it over at the earliest convenient exchange point, minimizing the distance carried on its own dime. Each provider plays the same move in its own direction. So traffic from A to B leaves A's network early. And traffic from B to A leaves B's network early. That produces two different roads through the middle by design.
- Different owners, different preferences. The destination's network chooses the return route using its own policies, costs, and peering relationships, none of which are visible — or answerable — to you.
- Multihoming. Any organization with two internet circuits (you, perhaps) can easily send traffic out one and receive answers on the other. That depends on how addresses are announced and which inbound road the far side's providers prefer.
- Load balancing. Providers spread flows across parallel links, and the spreading decisions differ per direction, adding small-scale asymmetry on top of the large-scale kind.
The practical takeaway: on paths that cross the public internet, expect asymmetry as the normal state of affairs. Inside your own network, by contrast, you control both directions, and symmetry is a reasonable expectation. When internal paths go asymmetric, it is usually because redundancy created a second road. In that case, traffic quietly split across the pair. That redundancy might be two firewalls, two WAN circuits, or a VPN plus a private line. Hold that thought for the firewall section.
Why traceroute shows only half the story
Traceroute discovers the road by coaxing each router along it to reveal itself, hop by hop. The mechanics are covered in reading traceroute output. The crucial fine print: the routers it reveals are the ones handling your outbound packets. The return road contributes no names and no addresses to the display. There is no flag that fixes this; the tool is structurally one-eyed, and so is every pretty visual traceroute built on top of it.
Subtler still: while the addresses are outbound-only, every number is not. Each latency figure is a round trip. It goes out to that hop, then back over whatever return road that hop's reply happened to take. That road can differ from hop to hop (the anatomy of a round trip dissects what hides in such numbers). This is one reason per-hop latencies do strange things — a middle hop reporting more latency than the destination, say. The reply from hop six may ride home on a scenic detour that the replies from hops seven and beyond never touch. The printed road is half the picture; the printed times secretly measure both halves at once.
Sit with that for a second, because it reframes the tool. A traceroute is an accurate list of outbound landmarks, annotated with timings you cannot fully attribute to them. Useful — enormously — but only half a map.
What asymmetry does to your measurements
Once you know there are two roads, a family of confusing observations turns mundane:
Latency you cannot locate. Ping says the round trip to the branch has grown by forty milliseconds. But your outbound traceroute looks exactly like the baseline, every hop calm. Nothing is wrong with your half of the journey — the growth lives on the return road, which your trace does not display. Whole afternoons have been lost to re-tracing a healthy outbound path while the sick return path sat invisible.
Loss you cannot direct. Ping reports three percent loss. Which direction died — requests that never arrived, or replies that never made it back? From one end, you genuinely cannot tell; the arithmetic of a round trip erases the distinction. The same applies to the loss column of a long MTR run: it counts round trips that failed, wherever they failed.
One-way slowness. Uploads to the branch fly; downloads crawl. The morning video call is crisp in one direction and robotic in the other. These are signatures of two roads with two different personalities — congestion, like traffic jams, has a direction. If your mental model has only one path, one-way symptoms sound impossible and get dismissed as user imagination. They are not; they are the two-road reality showing through.
A fair question at this point: why not just measure each direction separately and be done with it? Because true one-way delay measurement requires both ends to agree on what time it is, down to the millisecond. You are timestamping departure at one clock and arrival at another. Without carefully synchronized clocks at both ends, the measurement is fiction. That is exactly why every everyday tool reports round trips instead: one clock, out and back, honest but blended. Serious two-clock setups exist in the carrier world; for the rest of us, the practical substitutes are the two-ended tests in the detection section below.
The stateful-firewall trap
Measurement confusion wastes time; this next one breaks things outright. A stateful firewall keeps a memory of conversations. When it sees your PC open a connection outward, it files a note. When replies arrive, it admits them because they match the note. Anything that arrives matching no note is treated as uninvited and dropped. This is the cornerstone of modern perimeter security, and it carries a hidden assumption. The firewall assumes it will see both halves of every conversation.
Asymmetry breaks the assumption. Give a network two edges. Those might be two internet circuits on separate firewalls, or a pair of active-active firewalls without synchronized state. They might be a VPN tunnel alongside a private line. Sooner or later, a conversation goes out through edge A while its replies arrive at edge B. Edge B checks its notes, finds nothing (the outbound note lives in A's memory), declares the replies uninvited, and silently eats them. The user experience is maddeningly specific. Connections hang mid-handshake, and some destinations work while others do not. The problem "fixes itself" when a failover forces all traffic through one edge. That is because the failover restores symmetry by accident.
A wrinkle worth understanding is NAT — network address translation. This is the edge device's habit of rewriting your private addresses to its own public one. NAT both protects you from this trap and deepens it. It protects you because a NATed conversation's replies are addressed to the specific public address of the circuit it left through. So the internet has no choice but to return them to that same edge. Simple dual-WAN setups "just work" largely for this reason. NAT deepens the trap when replies somehow do land on the other edge. This typically happens where both circuits carry the same public addresses, as in provider-independent multihoming. That edge holds neither the permission note nor the translation. Even if it wanted to forgive the packet, it has no record of which inside machine the conversation belongs to. The drop is not just policy; it is arithmetic.
If this trap is anywhere in your building, the firewall's own deny log is where it confesses. In that case, look for dropped packets flagged as out-of-state, no-matching-session, or similar wording — legitimate reply traffic condemned as unsolicited. The cures are standard. Synchronize state between the paired devices so both hold every note. Or pin each flow's two directions to the same edge with routing policy. Or accept asymmetry only on segments where no stateful device straddles the split.
Tip: when "some sites hang but ping works" appears shortly after adding a second circuit, VPN, or firewall, suspect this trap first. Ping often survives (some setups state-track it loosely or allow it broadly) while TCP conversations die on the stateless edge. That is exactly the confusing mix users will report.
Detecting asymmetry from your side of the fence
You cannot see the return road directly from your desk, but you can infer it, and occasionally borrow a desk that sees it plainly. In rough order of conclusiveness:
- 1. Trace from both ends. The gold standard. From your side, trace toward the far end. From a machine at the far end — a branch probe, a server you rent, a colleague on the phone — trace back toward you. Compare the two hop lists. Different providers or wildly different roads in the middle: confirmed asymmetry, now mapped. This is one more argument for keeping a foothold at the far side of every path you depend on.
- 2. Read the TTL tea leaves. Every packet carries a TTL (time to live) counter that each router decrements by one. Replies start at a round initial value — commonly 64, 128, or 255 depending on the operating system. So the TTL printed by ping reveals how many routers the reply crossed coming back. Compare that with your outbound hop count from traceroute.
- 3. Borrow a looking glass. Many providers operate public looking glass pages — web tools that run pings and traces from inside their network toward an address you supply. Point one at your own public address and you are effectively tracing the return road from partway along it, no far-end machine required.
- 4. Read your firewall's complaints. Out-of-state drops, as above — less a measurement than a symptom, but within your own estate it is often the first and loudest clue.
The TTL trick deserves a worked example, because it costs ten seconds and works from one end:
$ traceroute -n 203.0.113.30
...
9 203.0.113.30 88.4 ms <- destination reached: about nine hops OUT
$ ping -c 1 203.0.113.30
64 bytes from 203.0.113.30: icmp_seq=1 ttl=52 time=88.1 ms
<- reply arrived with TTL 52
# Reply TTLs start at a round number: 64, 128, or 255.
# Nearest above 52 is 64, so: 64 - 52 = about twelve hops BACK.
# Nine out, twelve back: the return road is a different road.
Treat the result as a strong hint, not a court exhibit. You are guessing the initial TTL (pick the nearest round value above what you received). And matching hop counts do not prove matching roads — two different twelve-hop routes count the same. Mismatched counts, though, settle the headline question in seconds: there are two roads.
The same tea leaves also work over time. The reply TTL from a given destination is normally boring — the same value, ping after ping, day after day, because the return road's length is stable. So note it in your records. When it shifts — 52 yesterday, 47 today — the return road just changed length. That means it changed, full stop. That is a return-path change detected from one end, with no far-side machine involved. It is the one-eyed cousin of the outbound diffing in detecting path changes: far less detail, but watching the direction your traceroute never sees.
Mapping and documenting both directions
For paths that matter — office to branch, office to data center, anything site-to-site — do the two-ended capture during a healthy week. For those paths, write both roads into the path register you built in building a path baseline. One entry, two directions, honestly labeled:
| Field | Outbound (office → branch) | Return (branch → office) |
|---|---|---|
| Captured from | Office probe | Branch probe |
| Egress | Fiber circuit via fw1 | Branch cable circuit |
| Providers seen | example-isp.net, transit.example | othernet.example only |
| Hop count | Nine to ten | Twelve |
| Round trip from this end | Roughly 30 to 38 ms | Roughly 30 to 38 ms |
| Verified | Date and initials | Date and initials |
Two practical notes. First, the round-trip row is the same journey measured from opposite ends, so the numbers should roughly agree. When they stop agreeing, something one-sided is happening to the probes themselves, which is its own clue. Second, the moment both directions are in the register, your path change detection deserves the same upgrade. Schedule captures from both ends, because the return road can change on its own schedule while the outbound road sits innocently still.
Give the map a picture, too. On whatever network diagram your team keeps, draw site-to-site paths as two arrows, not one line, each labeled with its egress and provider. It looks fussy right up until an incident at two in the morning. The on-call person may be future you, or someone who has never read this guide. They need to grasp in one glance that "the road there" and "the road back" are different objects with different owners. One line teaches the wrong model; two arrows teach the right one silently.
Fine or fault: judging what you find
Detection always ends with the same question: so is this bad? Judge it with the table, not with instinct:
| What you found | Verdict | Why |
|---|---|---|
| Internet destination, two roads, both performing within baseline | Fine | This is the internet working as designed; document and move on |
| Odd per-hop latencies mid-trace, service healthy end to end | Fine | Replies riding varied return roads; an artifact, not a fault |
| One direction clearly slower or lossier than the other | Problem worth pursuing | Users live on the worse road; map both directions and escalate with two-ended evidence |
| A stateful device sees only half of some conversations | Fault, fix now | Legitimate replies are being dropped; sync state or pin flows symmetric |
| Asymmetry between your own redundant edges or cores | Design smell | You own both roads; unmanaged internal asymmetry mostly buys you the firewall trap |
| Site-to-site traffic out the encrypted tunnel, replies outside it | Fault, fix now | The conversation is half-protected and usually half-dropped |
The pattern behind the table: asymmetry across networks you do not own is weather; asymmetry across stateful devices you do own is a bug. Between those poles sits performance asymmetry, which is worth exactly as much attention as the slower direction's effect on real users. Sometimes that means no attention, sometimes an escalation with both traces attached.
Where to go next
You now carry the two-road model, and it quietly upgrades every other path skill you have. Baselines become two-column records. A clean outbound trace during a lossy incident stops being a contradiction and starts being a pointer at the return road. And when you work the localization playbook, you will remember that half the evidence is structurally invisible from any single vantage point. That is precisely why that playbook leans on testing from both ends of the pain.
From here, the rest of the Path & Hop-by-Hop Analysis guides put the model to work. And if the phrase "round trip" still feels like a single number rather than two journeys wearing a trench coat, the network latency guides will finish the job.
Put this into practice with SysPeep Monitor
Asymmetric problems yield to measurements taken from both ends — which is easiest when both ends are already instrumented. SysPeep Monitor's lightweight remote probes measure latency, loss, and availability from each site simultaneously. They feed one central console where a one-way problem stands out at a glance. The free trial installs on a Windows machine in minutes.
Download Free TrialLearn About SysPeep