Peninsulas and Islands in Internet Connectivity

• Peninsulas and islands are connectivity phenomena where peninsulas exhibit partial reachability and islands are fully isolated from the Internet core.
• Algorithmic detection using tools like Taitao and Chiloe quantifies these events with high precision and practical scalability across large datasets.
• Measurement platforms such as RIPE Atlas and CAIDA Ark validate the impact of these phenomena on outage metrics, guiding operational troubleshooting and policy decisions.

Persistent partial reachability in the public Internet manifests through two principal connectivity phenomena—peninsulas and islands—each reflecting fragmentation in the global routing substrate despite the foundational goal of universal end-to-end communication. Recent research formalizes these concepts via reachability-driven definitions applied to the Internet core, develops detection mechanisms, and quantifies their prevalence across major longitudinal datasets. Peninsulas constitute network blocks exhibiting persistent, partial connectivity as seen by distributed vantage points, while islands denote partitions fully disconnected from the Internet core yet mutually reachable. These forms of fragmentation, rooted in policy, misconfiguration, and commercial dispute, occur with frequency comparable to conventional outages, directly impacting measurement fidelity, operational troubleshooting, and Internet governance.

1. Formal Definitions and Conceptual Foundations

The Internet core is defined as a unique, strongly-connected component (SCC) in the directed graph $G$ of active, public IP addresses, wherein nodes $x, y$ satisfy bidirectional routability ($x \rightarrow y$ and $y \rightarrow x$). The core is required to contain $|C| > 0.5 \cdot |A|$ active addresses for uniqueness and global coverage. Peninsulas and islands are formalized through observer disagreement:

• Peninsula: For block $b$ in measurement round $i$, with $O_{i,b}$ being the set of probing vantage points and $O^{up}_{i,b}$ those seeing $b$ up, $b$ is on a peninsula if $0 < |O^{up}_{i,b}| < |O_{i,b}|$ (Baltra et al., 17 Jan 2026, Baltra et al., 2024, Baltra et al., 2024).
• Island: Observer $o$ is on an island at round $i$ if $0 \leq |B^{up}_{i,o}| < |B^{dn}_{i,o}|$, i.e., $o$ reaches fewer than half the core blocks ($B$) while retaining local connectivity. For long-lived events, $|B^{up}_{i,o}| \rightarrow 0$ is required (Baltra et al., 17 Jan 2026, Baltra et al., 2024).

This reachability-centric foundation rejects authority-based or registry-rooted definitions and supports a neutral, peer-driven model for Internet partitioning and core membership.

2. Algorithmic Detection of Peninsulas and Islands

Two primary algorithms operationalize these phenomena:

• Taitao (Peninsula Detector): Iteratively, for each round and block, aggregates per-vantage-point reachability, reporting peninsulas where partial observer disagreement persists beyond one window. High-noise or unreliable blocks (response $<85\%$ or $>10$ distinct response patterns) are filtered. Long-lived peninsulas require $\geq5$ hours duration to match daily-probing systems (Baltra et al., 17 Jan 2026, Baltra et al., 2024).
• Chiloe (Island Detector): For each vantage point and round, computes the fraction of the core reachable. If this falls to at most half the core (or zero for persistent islands), the VP is classified as an island for events lasting $>1$ round (Baltra et al., 2024).

The computational complexity for both is tractable ($O(|T|\cdot|B|\cdot|O|)$ for Taitao, $O(|T|\cdot|O|\cdot|C|)$ for Chiloe), with practical implementations on datasets involving up to $10^4$ vantage points and $5\times10^6$ monitored blocks.

3. Measurement Systems and Empirical Validation

Three complementary infrastructures underpin recent large-scale studies:

Platform	Vantage Points (VPs)	Measurement Type
Trinocular	6	ICMP pings, /24
RIPE Atlas	~12,000	DNS root queries
CAIDA Ark	171	Traceroutes, /24

Cross-validation leverages CAIDA Ark’s daily traceroutes to confirm peninsula events observed by Taitao, reporting recall $\approx0.94$, and bounding strict precision $0.42$ and loose precision $0.82$ (Baltra et al., 17 Jan 2026). RIPE Atlas and Trinocular provide high-frequency, multi-year time-series for both phenomena, while RouteViews BGP feeds contextualize prefix-level policy effects.

4. Prevalence, Durations, and Root Causes

• Peninsula Frequency: Across $3.7 \times 10^6$ observed blocks (2017Q4), the fraction of time blocks are peninsulas ($\approx0.075\%$) parallels the outage fraction ($\approx0.074\%$). Statistical convergence of estimates is achieved with $\geq3$ independent VPs (Baltra et al., 17 Jan 2026, Baltra et al., 2024).
• Event Durations: 65% of peninsula events last 20–60 minutes; 2% persist $\geq24$ hours, yet these dominate total peninsula-time (57%). A small subset (7%) of long-lived peninsulas account for 90% of the total lost connectivity time (Baltra et al., 17 Jan 2026).
• Peninsula Sizes: 10% of peninsulas fully cover a routable prefix (interdomain policy impact); 30% are small ($<$5% coverage). 20% of peninsula-time is in full-prefix partitions.
• Islands: RIPE Atlas finds 2.1% of VPs as IPv4 islands and 3.4% IPv6 islands over 90 days; most events last $\leq20$ min and are local (70% within one hop of VP) (Baltra et al., 17 Jan 2026).
• Root Causes: Policy-level filters at AS ingress, peering disputes (e.g., Cogent vs Hurricane Electric for IPv6), routing transients (BGP flaps), corporate firewalls, and local misconfigurations are established drivers of persistent partial reachability (Baltra et al., 2024).

5. Impact on Measurement Fidelity and Operational Monitoring

Peninsulas and islands introduce “noise” in outage and reliability metrics, often exceeding operational signals by $5\times$ to $9.7\times$ (e.g., RIPE DNSmon query loss rates). Removing island and peninsula contributions sharpens system sensitivity to genuine connectivity transitions, notably in DNS root server monitoring (Baltra et al., 17 Jan 2026, Baltra et al., 2024, Baltra et al., 2024). Measurement tools (Trinocular, DNSmon) are recommended to explicitly classify partial reachability, rather than subsume disagreement events into “up” or “down” bins, enhancing diagnostic precision. Utilizing only 3–4 independent, geographically-distributed VPs suffices for high-confidence prevalence estimation.

6. Policy, Governance, and Architectural Implications

The majority-based definition of the Internet core asserts no single provider, country, or RIR controls $\geq50\%$ of active addresses (ARIN $\approx45\%$ IPv4, RIPE $\approx47\%$ IPv6) (Baltra et al., 17 Jan 2026, Baltra et al., 2024). Secession, sovereign “kill-switch,” and fragmentation scenarios require coalitions exceeding this threshold, anchoring policy discussions in neutral technical criteria. Cloud, NAT, and firewall-induced conditional reachability further underscore partial connectivity’s structural permanence—rather than its treatment as anomalous failure. The reachability-first principle provides a rigorous foundation for both future innovation and resilience quantification.

7. Limitations, Measurement Artifacts, and Future Directions

A key distinction is drawn between genuine islands (site partitioned yet locally alive) and measurement artifacts, as local LAN-only configurations can produce false-positives, especially in densely instrumented platforms such as RIPE Atlas (Baltra et al., 2024). Algorithmic refinements, including all-but-one or two-round voting, mitigate observer locality errors. The operational landscape is enriched by recognizing partial partitions as a pervasive and tractable class of events—warranting attention equal to classic outage detection. Ongoing research aims to generalize these tools for next-generation Internet measurement and to inform robust policies as fragmentation risk evolves.