Back to All Articles
SRE & DevOpsOct 7, 202611 min read

Why Single-Location Uptime Checks Are Dangerous for Production Services

A monitoring probe that reports your site as "down" has only observed one network path from one location. Discover why single-location uptime checks produce false positives, miss regional outages, and how distributed multi-probe monitoring gives you accurate availability data.

P
Prabin Jha
Founder & Lead Architect
LinkedIn

Key Takeaways (BLUF)

Bottom Line Up Front (BLUF): When a single monitoring probe reports your service as "down," it has observed exactly one thing: that your endpoint was unreachable from one network location at one moment in time. It cannot determine whether 100% of your users are affected or zero users are affected. Single-location monitoring conflates "unreachable from here" with "unavailable for everyone" — a logical flaw that produces both false positives (paging your on-call team when the service is healthy) and false negatives (reporting "UP" while an entire continent cannot reach your application). Distributed multi-location monitoring with independent network paths is the only way to reliably distinguish a genuine global outage from a localized network event.

The Fundamental Problem: What a Monitoring Probe Actually Measures

Every uptime monitoring result is an observation made from somewhere. A probe sends an HTTP request, a TCP connection attempt, or an ICMP ping across a specific network path from a specific geographic location through specific autonomous systems to your server. It measures whether that particular network path succeeded at that particular moment.

This is subtly but critically different from measuring whether your service is available.

The difference becomes clear when you consider the structure of the modern internet. Between a monitoring probe in, say, Frankfurt and your origin server in Singapore, there are typically 12 to 25 BGP hops spanning multiple tier-1 carrier networks, multiple IXP peering points, and potentially one or more CDN layers. Any component along that specific path can fail without affecting any other path to your server.

This is the core observation problem of single-location monitoring, and it has two failure modes that experienced SREs encounter constantly:

  • False Positive: The probe reports DOWN, but the service is healthy. The failure exists somewhere on the probe's specific network path — not in your application. Your users are unaffected.
  • False Negative: The probe reports UP, but a significant portion of your user base cannot reach the service. The failure exists on network paths that the probe does not traverse.

False Positives: When "Down" Means "Unreachable From Here"

The false positive case is the one that burns engineering teams the fastest. A 3 AM PagerDuty alert fires. An engineer wakes up, opens dashboards, checks application logs, inspects database health, scans Cloudflare analytics — and finds nothing wrong. Traffic is flowing normally. Error rates are nominal. The service is healthy.

The monitor reported DOWN because something failed between the probe and the server — not inside the server. Here are the most common causes:

  • BGP Route Flapping: The Border Gateway Protocol, which governs how autonomous systems exchange routing information across the global internet, is not designed for millisecond stability. When an ISP or carrier re-converges routes following a link failure or traffic engineering change, packet loss spikes of 2 to 30 seconds are routine. A probe positioned downstream of that route change experiences a timeout and fires an alert. Every other network path to your service continues working normally.
  • Monitoring Provider Transit Failure: The monitoring provider's own upstream internet connection can experience congestion or partial failure. If your monitoring runs from a single VPS in one data center, an outage in that data center's upstream transit provider generates false alerts for every monitored endpoint — none of which are actually down.
  • Anycast CDN Route Settlement: Cloudflare, Fastly, and AWS CloudFront use Anycast routing, meaning the same IP address is announced from hundreds of global Points of Presence simultaneously. When CDN operators perform maintenance or traffic engineering, a small number of network paths temporarily route to a different PoP. During the 5–15 second settlement period, a probe on that specific path may receive a 503 or connection reset. The probe fires an alert. Every real user, routed through a different PoP, notices nothing.
  • Firewall Rate Limiting and IP Blocking: Some WAF configurations, DDoS protection layers, or cloud provider security groups apply rate limits or temporary IP blocks to specific source IP ranges. If the monitoring provider's probe IP is temporarily rate-limited, the probe reports a failure that no real user ever experiences.
  • IPv6 Connectivity Gaps: Many monitoring providers default to IPv4 probing. If your DNS returns AAAA records preferentially and the probe's path has an IPv6 routing problem, the probe reports failure while IPv4 users connecting from elsewhere experience no disruption.

False Negatives: When "Up" Means "Up From Here"

The false negative case is less discussed but often more damaging. Your monitoring dashboard shows green. Your status page reports "Operational." Meanwhile, customers in an entire region are receiving connection timeouts, DNS resolution failures, or HTTP 502 errors.

This happens when the monitoring probe is located in a region or on a network path that does not share the failure affecting real users. Common causes include:

  • Regional Cloud Availability Zone Failures: AWS, GCP, and Azure routinely experience AZ-level failures that affect services deployed in specific regions. If your monitoring probe is located in us-east-1 and your service is hosted there, a failure in eu-west-1 that blocks European users from reaching a regional endpoint would never be detected.
  • DNS Propagation and Resolver Failures: DNS resolution is not globally consistent. A misconfigured DNS change, a TTL-related propagation lag, or a failure in a specific DNS resolver operator (such as a regional ISP's resolver) can cause a subset of users to receive incorrect or absent DNS responses. A probe using a different resolver — or resolving from cache — reports the correct IP and connects successfully, while real users fail at the DNS layer.
  • CDN Origin vs Edge Discrepancy: When CDN caching is functioning, probes may receive cached responses from edge PoPs even when the origin is experiencing failures. The probe measures "is the CDN edge responding?" — not "is the origin healthy?" A probe positioned to hit a warm CDN edge will report UP while the origin is in a degraded state and non-cached requests are failing.
  • Geographic Load Balancer Routing: Services using geographic or latency-based load balancing route different users to different origin endpoints. A probe located in Virginia may consistently reach a healthy us-east endpoint while eu-central users are routed to a degraded eu-central endpoint.

Five Concrete Scenarios That Single-Location Monitoring Cannot Diagnose

Abstract failure modes are useful. Concrete examples are more useful. Here are five real-world scenarios that illustrate exactly where single-location monitoring fails:

ScenarioWhat Actually HappenedSingle-Probe ResultReality for UsersRequired Monitoring Capability
Scenario 1: Monitoring Provider BGP EventThe monitoring provider's upstream carrier rerouted traffic for 12 seconds during maintenance.DOWN alert fires at 2:47 AM.100% of users experienced zero disruption.Multi-provider probing from independent network paths. A single provider's BGP event cannot affect all providers simultaneously.
Scenario 2: EU CDN PoP FailureA Cloudflare PoP in Frankfurt experienced a partial failure affecting European user routing.US-based probe reports UP. No alert.European users received intermittent 502 errors for 23 minutes.Probes from within the affected region (EU) would have detected the PoP failure. Geographic coverage matched to user distribution.
Scenario 3: DNS Resolver FailureA major European ISP's recursive DNS resolver failed to resolve the monitored domain.Probe using Google 8.8.8.8 resolver reports UP. No alert.Millions of ISP customers received NXDOMAIN responses and could not access the service.DNS monitoring from multiple resolver operators. HTTP-only monitoring cannot detect DNS-layer failures affecting specific resolver populations.
Scenario 4: Database Primary FailoverPostgreSQL primary failed. Replica promotion took 45 seconds. During promotion, write requests returned 500 errors.Probe checking GET /health returns 200 from read replica throughout. Reports UP.All write operations (checkout, signup, save) failed for 45 seconds across all regions.Functional health checks that exercise write paths, not just ping/GET endpoints. HTTP 200 on a read-only health endpoint does not confirm write-path availability.
Scenario 5: Genuine Global Application CrashA deployment pushed a circular import that caused the Node.js process to crash on startup across all regions.All probes from all regions report DOWN within seconds.100% of users received connection refused. Complete global outage confirmed.Multi-region corroboration confirms global outage with high confidence. This is the scenario where consensus detection correctly distinguishes noise from real failure.

The Observation Table: What Single vs Multi-Location Monitoring Can and Cannot Determine

The key insight is that a monitoring result is always conditional: it tells you what was observed from a specific location on a specific network path. The table below maps common infrastructure events to what single-location and multi-location monitoring each actually report:

Infrastructure EventSingle-Location May ReportWhat Is Actually HappeningMulti-Location Reveals
Monitoring provider's local network failureDOWN (false positive)Service is healthy globallyAll other probes report UP → local probe failure isolated, no alert
BGP route flap on one carrier (8–30 sec)DOWN (false positive)Transient routing event, users unaffectedProbes on independent AS paths report UP → event filtered as noise
Regional CDN PoP failure (EU)UP (false negative) — if probe is in USEU users experiencing failuresEU probe reports DOWN, US probe reports UP → regional failure detected
DNS resolver failure for one ISPUP (false negative) — if probe uses different resolverISP's customer base cannot resolve domainProbe using affected resolver type detects failure; DNS monitoring catches this separately
Application process crash across all regionsDOWN (true positive)Global service outageAll probes confirm DOWN → high-confidence global outage, alert dispatched immediately
Database replica lag (reads slow, writes fail)UP (false negative) — GET /health returns 200Write operations failing globallyFunctional probes testing write endpoints detect failure; read-only health checks insufficient

Geographic Diversity vs Network Diversity: An Important Distinction

When engineers first encounter multi-location monitoring, the instinct is to think purely geographically: "If I put probes in New York, London, and Tokyo, I have global coverage." This is partially correct but misses a subtler point.

Geographic distance is a proxy for network path independence — but it is an imperfect proxy. Consider the following:

  • Two probes 8,000 km apart can still share a common upstream transit provider. If that transit provider experiences an event, both probes report DOWN simultaneously — creating a correlated false positive that looks identical to a genuine global outage.
  • Two probes in the same city but on different ISPs with different Autonomous System Numbers (ASNs) traverse completely independent network paths. They provide stronger evidence of independence than two distant probes on the same backbone provider.
  • Cloud provider probes have correlated failure modes. If your monitoring runs entirely within AWS EC2 instances, an AWS-level event (such as the December 2021 us-east-1 failure that cascaded into multiple services) can affect all probes simultaneously. Running probes across AWS, Google Cloud, and independent bare-metal networks provides genuine path diversity.
  • DNS resolver diversity matters independently of HTTP probe location. A probe that resolves DNS correctly but then performs an HTTP check from an affected region may report a different failure mode than a probe that uses an affected DNS resolver in an unaffected region.

How Distributed Monitoring Changes the Evidence Quality

The value of multi-location monitoring is not mathematical certainty — it is evidence quality. A single probe failure is a weak signal: it could mean your service is down, or it could mean one network path had a transient problem. Multiple independent probe failures are a much stronger signal: the probability that genuinely independent network paths all experience simultaneous unrelated failures converges toward zero as probe count and path diversity increase.

This is the foundation of quorum-based or consensus-based monitoring — the approach used in WakeMetrics and described in detail in our article on how 3-stage consensus eliminates false positive alerts. The principle:

  • One probe DOWN, two probes UP → weak evidence of a problem. More likely a localized network event on the failing probe's path. No alert dispatched.
  • Two probes DOWN, one probe UP → moderate evidence. Could be a regional failure affecting two overlapping network zones. Warrants investigation-level notification, not full incident escalation.
  • Three probes DOWN across three independent regions and ASNs → strong evidence of a genuine global or near-global service outage. High-confidence alert appropriate for full incident escalation.

The Right Monitoring Depth: What "UP" Should Actually Mean

Multi-location probing solves the network observation problem. But there is a parallel issue: what does a successful probe actually verify? An HTTP 200 status code from a single endpoint is not the same as "the application is healthy." The depth of the health check determines the scope of what "UP" actually means.

Monitoring Check TypeWhat It VerifiesWhat It Does Not Verify
ICMP PingHost is reachable at network layerApplication running, HTTP serving, database connected, SSL valid
TCP Port CheckPort is accepting connectionsApplication is processing requests correctly
HTTP GET /health (returns 200)Web server responds with success statusDatabase queries succeeding, write paths functional, third-party integrations available
HTTP GET with response body assertionWeb server responds correctly and application-level content is presentBackground jobs running, queue processing, cache coherence
Functional synthetic transactionFull user-facing flow succeeds (login, search, checkout)Edge cases, concurrent user behavior, data consistency
SSL certificate checkCertificate is valid, chain is trusted, not expiredApplication availability, database health, background processing

What a Better Uptime Monitoring Strategy Looks Like

Replacing single-location monitoring with a more reliable approach does not require a complex multi-tool architecture. The following principles — applied in order of increasing sophistication — dramatically improve monitoring accuracy:

  • 1. Require at least 3 geographically separate probe locations. North America, Europe, and Asia-Pacific covers the three largest internet exchange regions and ensures that a single carrier event is unlikely to affect all three simultaneously.
  • 2. Require probes on independent network paths (different ASNs and cloud providers). Geographic distance is insufficient if all probes share the same upstream transit provider.
  • 3. Apply a consecutive-failure threshold before alerting. Requiring 2 or more consecutive failures from the same probe before it contributes to an alert decision eliminates transient single-packet-loss events from generating pages.
  • 4. Require quorum confirmation across locations before dispatching an alert. One probe DOWN with two probes UP should not page on-call. Two or three probes DOWN with independent corroboration should. This is the consensus principle.
  • 5. Monitor DNS separately from HTTP. DNS failures are invisible to HTTP-only monitors. A dedicated DNS check from multiple resolvers catches resolver failures, propagation issues, and misconfiguration that an HTTP probe would never detect.
  • 6. Use health checks appropriate to your application's actual failure modes. If database connectivity is critical, exercise a lightweight database query in your /health endpoint. If payment processing is critical, synthetic transaction monitoring of the checkout flow provides evidence that HTTP status checks cannot.
  • 7. Align monitoring locations with your actual user geography. A SaaS product with 60% European users should have proportional monitoring coverage in Europe, not just in Virginia where cloud infrastructure is cheapest to run.

Common Mistakes Engineers Make With Uptime Monitoring

Having worked through hundreds of infrastructure monitoring setups, the same misconfiguration patterns appear repeatedly. Here are the most consequential mistakes and how they compound each other:

  • Monitoring a CDN edge instead of the origin. Many teams point their monitors at the same URL their users access, which routes to a CDN PoP. This monitors CDN edge health, not application health. When the origin degrades while CDN serves stale cache, the monitor reports UP while users receive outdated content or eventual errors.
  • Using a single check interval that is too long. A 5-minute check interval means a service can be down for up to 4 minutes and 59 seconds before the monitor detects it. The first failed check then triggers an alert that might self-resolve before confirmation — creating alert noise without useful signal.
  • Treating any non-200 status as a downtime event. Redirects (301/302), authentication challenges (401), and rate-limit responses (429) are not downtime. Misconfigured status code expectations generate false alarms on entirely healthy services.
  • Not monitoring SSL certificate expiry separately. HTTPS monitoring verifies that SSL is currently valid. It does not warn you that the certificate expires in 7 days. Certificate expiry is a predictable failure that should trigger a warning well in advance, not an incident page when users start seeing browser errors.
  • Ignoring timeout configuration. A monitor with a 30-second timeout will not detect slow endpoints — it will simply wait, then report success if the response eventually arrives. Tight, realistic timeout thresholds (typically 3–8 seconds for production APIs) catch latency degradation that HTTP 200 checks miss entirely.

Configuring Multi-Location Monitoring: A Practical Starting Point

The following cURL example shows how to configure a production monitor via the WakeMetrics API with multi-region coverage, realistic timeout thresholds, and consensus confirmation settings that eliminate the majority of false positive alerts:

#!/usr/bin/env bash
# WakeMetrics: Multi-Location Monitor with Consensus Configuration
# This replaces dangerous single-probe monitoring with geographically
# and network-diverse probing from 3 independent regions.

API_KEY="your_wakemetrics_api_key"

# Create monitor with 3-region coverage and consensus confirmation
curl -s -X POST "https://api.wakemetrics.com/v1/monitors" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Production API — Global Health Check",
    "url": "https://api.yourcompany.com/health",
    "type": "http",
    "interval_seconds": 15,
    "timeout_ms": 5000,
    "method": "GET",
    "expected_status": 200,
    "assert_response_body": "ok",
    "confirm_down": 2,
    "regions": ["us-east", "eu-west", "ap-southeast"],
    "alert_channels": [
      { "type": "slack",     "webhook_url": "https://hooks.slack.com/..." },
      { "type": "pagerduty", "routing_key": "your_routing_key" }
    ],
    "ssl_check": true,
    "ssl_expiry_warning_days": 30
  }' | jq '.'

# Also create a DNS-specific monitor to catch resolver failures
# that HTTP checks cannot detect:
curl -s -X POST "https://api.wakemetrics.com/v1/monitors" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "DNS Resolution Check — api.yourcompany.com",
    "url": "api.yourcompany.com",
    "type": "dns",
    "interval_seconds": 60,
    "regions": ["us-east", "eu-west", "ap-southeast"],
    "expected_ip": "your.expected.ip.address"
  }' | jq '.'

Monitoring Architecture Comparison

For teams evaluating monitoring approaches, here is an honest comparison of the trade-offs. No approach provides absolute certainty — each one changes what you can and cannot observe:

Monitoring ApproachGeographic CoverageFalse Positive RiskFalse Negative RiskOperational Complexity
Single location, single probeOne network pathHigh (any local BGP/ISP event triggers alert)High (regional outages invisible if probe region unaffected)Low
Multiple locations, alert-on-any-failureSeveral network pathsVery high (more probes = more chances for individual false positives)Medium (regional coverage improves)Low–Medium
Multiple locations, consecutive-failure thresholdSeveral network pathsMedium (transient events filtered by retry)MediumMedium
Multiple locations, quorum/consensus requiredSeveral independent network pathsVery low (simultaneous multi-path failures required)Low–Medium (with correct region placement)Medium
Multiple locations + functional synthetic checks + DNS monitoringNetwork, application, and DNS layers coveredVery lowLow (covers CDN, DNS, and application failures)Medium–High

Further Reading in the WakeMetrics SRE Series

This article addresses the root cause of the problem — single-location observation. The following articles in this series address specific dimensions of the solution:

  • How 3-Stage Consensus Eliminates 99.4% of False Positive Downtime Alerts — deep dive into the SUSPECT state, quorum math, and consensus protocol: /blog/how-3-stage-consensus-eliminates-false-alarms
  • The Math Behind Alert Fatigue: How False Alarms Cost Tech Teams $40,000/Year — financial model for alert fatigue and the engineer productivity cost of false positives: /blog/cost-of-alert-fatigue-sre
  • How to Calculate Mean Time to Detect (MTTD) and Slash It by 70% — practical guide to measuring detection speed and improving it: /blog/how-to-reduce-mean-time-to-detect-mttd
  • Top 7 UptimeRobot Alternatives in 2026 — comparison of monitoring tools by check frequency, consensus engine, and pricing: /blog/best-uptimerobot-alternatives

Frequently Asked Questions

Why does my uptime monitor say my website is down when it is actually up?

The most common cause is that your monitoring probe experienced a transient network event — such as a BGP route flap, ISP congestion spike, or CDN Anycast re-routing — on the specific network path between the probe and your server. From the probe's perspective, the request timed out. From every other network path, your server was responding normally. This is a false positive: a legitimate downtime report from one probe that does not reflect the availability experienced by your actual users. The fix is multi-location monitoring with consensus confirmation — requiring at least 2 independent probes from geographically separate locations to agree before an alert is dispatched.

Can a website be "up" on uptime monitors but "down" for some users?

Yes — this is a false negative and it is one of the most damaging failure modes in uptime monitoring. It occurs when the monitoring probe is located in a region or on a network path that does not share the failure affecting real users. Common causes include regional CDN PoP failures (visible only to users routed through that PoP), DNS resolver failures affecting a specific ISP's customers, geographic load balancer routing to a degraded regional endpoint, or origin failures masked by CDN caching. The solution is to place monitoring probes in locations that match your actual user geography, and to monitor DNS resolution separately from HTTP availability.

What is the difference between single-location and multi-location uptime monitoring?

Single-location monitoring sends probes from one network location, providing one observation from one network path. It cannot distinguish whether a probe failure represents a global service outage or a localized network event. Multi-location monitoring sends independent probes from multiple geographically and network-diverse locations. When probes disagree (one reports DOWN, others report UP), the system can identify the failure as localized to one network path rather than treating it as a global outage. When probes agree (all report DOWN), the corroborating evidence strongly suggests a genuine global service failure. The quality of multi-location monitoring depends on probe independence — probes sharing upstream transit providers can fail together, mimicking a false global failure.

How many monitoring locations do I need for reliable uptime checks?

For most production SaaS services, three geographically and network-diverse monitoring locations provide a good balance between false-alarm reduction and operational complexity. The three locations should span at least two continents and be hosted on different cloud providers or backbone networks to ensure independent network paths. Five or more locations provide stronger evidence but with diminishing marginal returns for most services. More important than probe count is probe placement: monitoring from regions that match your actual user geography ensures that regional failures affecting real users are detected, not hidden behind probes positioned in unaffected regions.

Conclusion

Single-location uptime monitoring is not "good enough with a caveat" — it is architecturally incapable of answering the question you are actually asking, which is: "Can my users reach my service?" It can only answer: "Could this one probe reach this one endpoint from this one network path at this one moment?" For production infrastructure, those are two fundamentally different questions. The solution is not complex: three probes, three independent networks, a consensus requirement before alerting, and health checks designed to exercise your application's actual failure modes. Start your 14-day free trial at /signup?plan=business and configure your first multi-region consensus monitor in under 60 seconds.

Ready to Upgrade from UptimeRobot?

Migrate in under 60 seconds. Get 15-second checks across 450+ global locations with zero false alerts.