Skip to content
Crypto.world

Research

Does Client Diversity Protect the Lightning Network? Evidence from the 2026 Core Lightning Shutdown

Client diversity is routinely described as the Lightning Network's insurance policy: if one implementation has a bad bug, the others keep the network alive.

Does Client Diversity Protect the Lightning Network? Evidence from the 2026 Core Lightning Shutdown
arXiv
2610.20952
Published
2026-10-14
Authors
Saurabh Kumar

AI summary

Does Client Diversity Protect the Lightning Network? Evidence from the 2026 Core Lightning Shutdown

Overview

Research area: Bitcoin / Lightning Network measurement, network resilience and software monoculture risk (cs.CR, with a measurement/networking flavour).

Technical level: Intermediate. The paper is written in plain prose with an explicit methods section, but it assumes familiarity with Lightning gossip, channel funding outputs, force closes and graph reachability.

Scope (one sentence): Using 805 daily gossip snapshots and 132,360 on-chain channel closes, the paper measures what happened when Core Lightning maintainers told operators to disconnect on 26 August 2026, and uses that natural experiment to calibrate how far client-correlated failures actually spread.

What This Paper Is About

Client diversity is routinely described as the Lightning Network's insurance policy: if one implementation has a bad bug, the others keep the network alive. But the public graph is not diverse — LND runs 80% of public nodes (84% counting one operator's fleet of about 2,000 tiny nodes) and holds 84% of public channel capacity, a share that has stayed between 84% and 90% for three years. The paper asks two questions that had not been tested against real data: does the surviving diversity protect anything, and is the concentration as dangerous as it looks? A real emergency in August 2026, when Core Lightning (CLN) maintainers told operators who could not upgrade to disconnect, provides the first observed client-wide outage to measure.

Key Contributions

  1. A validated client classifier. Feature-bit rules applied in a fixed order, checked against 795 self-identifying aliases and against an independent on-chain signal — the wallet that builds each node's channel funding transactions. The rules are kept valid across three years of feature-set changes, and a learned baseline ((k)-nearest neighbours on the feature-bit vector) does not beat them (macro-F1 0.92 vs. 0.93).

  2. Three years of concentration measurement. Monthly client shares of public nodes and capacity from May 2023 to October 2026, including the identification of a managed fleet of small LND nodes (1,980 at its March 2026 peak, 1,972 on the latest snapshot holding 1.4 BTC) that distorts raw node counts.

  3. An observed client-correlated outage. A measurement of the CLN shutdown against churn baselines and a control client: who complied, how long nodes stayed dark, how peers and the closing market reacted, and how the event differs from a transport-level failure and from single-operator closing waves.

  4. A calibrated blast-radius model. Failure scenarios on the routable core of 25 August 2026, compared with 94 ordinary two-day windows, targeted baselines, and a null model of how clients connect, showing that hubs rather than cohort sizes determine the damage.

Main Findings

  • The outage was large inside CLN and nearly invisible outside it. Of 624 CLN nodes not dark on 25 August, 137 (22.0%) were dark on 27 August, holding 33% of CLN capacity — about ten times the ordinary daily transition rate of 2.3% (median over 97 days, May–August; 95th percentile 5.6%, maximum 12.3%). LND nodes went dark at 6.4% over the same days, within their ordinary range (median 5.2%, 95th percentile 9.5%).

  • Network-wide reachability barely moved. The share of channel directions toward CLN nodes that peers marked disabled rose from a baseline of 11.8 ± 0.9% to 38.1% on 27 August ((z=29.4)); by capacity from 7.4% to 35.2% ((z=15.5)). LND, Eclair and LDK did not move ((|z|<2)). The difference-in-differences estimate against LND is +14.2 percentage points (95% CI 11.2–16.0) by channel count and +14.7 (10.4–17.1) by capacity, still +12.9 (8.9–15.2) after removing the single node with the most disabled capacity each day.

  • Compliance split by size and by transport. 32.0% of the largest capacity quintile went dark against 8.0–27.2% of the others. By reachability: 28% ((n=310)) of Tor-only CLN nodes, 21% ((n=150)) of dual-stack nodes and 7% ((n=90)) of clearnet-only nodes. The two-week difference-in-differences for clearnet-only nodes is zero; for Tor-only nodes it is +23.1 points (20.8–25.2).

  • The damage was temporary for some, permanent for others. 39% of the 137 nodes were back within three days and 53% within a week; among returning nodes the median time dark was three days (interquartile range 3–7), clustered after CLN 26.06.7 shipped on 28 August. 22% never reappeared by 4 October; counting nodes that returned and went dark again, 32% were dark on that date.

  • A transport failure cost the network twice as much. On 10 July the disabled share toward Tor-only nodes of every client rose from 23.7% to 43.5% overnight (clearnet-only nodes flat), coinciding with a denial-of-service wave against Tor in which relays serving onion-service descriptors fell from 3,874 on 9 July to 1,780 on 12 July and recovered only after 22 July.

  • Reachability effects are modest by ordinary standards. The observed footprint (383 core nodes with no usable channel left on 27 August, 105 of them CLN) reduced all-pairs reachability for 100k-sat payments by 7.8 points. Across 94 ordinary two-day windows from May to August the median loss was 3.6 points, and 4 windows lost more than the CLN shutdown; the two Tor windows cost 15.6–16.0 points. Among CLN pairs the outage cost 24.5 points, more than any ordinary window (maximum 17.7, median 3.2).

  • A single hub outweighs an entire cohort. ACINQ's Eclair node holds 4.9% of core node capacity, and losing it alone cuts reachability by 15.9 points — four times the 3.7-point loss from all 619 CLN nodes in the core. The next most critical nodes, a Lightning service provider and a payment processor running LND, cost 2.8 and 1.9 points; the 20 most critical nodes, 19 of them LND, cost a median of 0.6 points each.

  • Containment is explained by peripherality. The calibrated scenario ((\hat{p}=0.22) of CLN nodes failing at random, 136 nodes) leaves reachability among survivors at 78.8%, against 78.2% when the same number of nodes fails at random across all clients. Drawing failures with the observed compliance of each size quintile and transport class gives the same result ((R=78.7%), (R_{\mathrm{all}}=73.9%)). CLN pairs fall from 79% to 46% connected, while (R_{\mathrm{CLN}}) among surviving CLN nodes is 76%. Losing all 619 CLN core nodes leaves survivors at 76.2%; losing the 11 Eclair nodes, which hold 5.8% of capacity, leaves 63.9%.

  • Minority clients' dependence on LND is what their size implies. 76% of CLN capacity, 92% of Eclair and 81% of LDK faces LND peers, and without LND nodes only 32% of CLN pairs, 51% of Eclair pairs and 12% of LDK pairs remain connected for 100k-sat payments. Under a null model that permutes client labels within capacity deciles (200 draws), CLN would put 88% of its capacity on LND peers and LDK 87%; both put less. CLN nodes keep 15% of capacity among CLN peers against 8% by chance, and LDK 15% against 2%.

  • Node-count diversity statistics mislead. LND runs 90 of the 100 largest nodes of the routable core and 8 of the 10 largest; capacity inequality within every cohort is high (Gini 0.92–0.97). On the 4 October 2026 snapshot, LND runs 84.0% of nodes (80.1% without the fleet) and holds 84.4% of capacity; CLN runs 8.5% of nodes but holds 3.7% of capacity; Eclair's 40 nodes hold 6.1%.

  • Most "incidents" in closing data are single operators. The two largest force-close days were one LND node each: NiceHash with 226 force closes on 3 August and Wallet of Satoshi with 205 on 26 August. Of 37 mass-close events since May, 30 involved LND nodes. Organic force closes involving CLN nodes ran at 2.2 times their July–August rate during 26 August–8 September, 2.6 times relative to LND-to-LND channels (which fell to 0.83 times).

  • The network contracted for reasons unrelated to clients. Of 38,081 channels in gossip on 1 August, 14,547 (38%, holding 44% of capacity) were gone by 4 October; 97% closed on-chain, 10,464 in August, against 6,833 new channels. Channels involving LND, CLN, Eclair and LDK disappeared at similar rates (38%, 42%, 32%, 48%), 79% cooperatively, with ten operators party to 31% of closes.

  • The on-chain cost was negligible. The median force close since May cost 1,222 sat in fees including sweeps, six times a cooperative close (209 sat); all 8,787 force closes paid 0.17 BTC in total. The real cost was liquidity locked behind time-locks and channels lost.

Methodology in Plain English

The author runs a Lightning node that dumps its view of the public network graph once a day, and a Bitcoin node that scans every block for spends of channel funding outputs. From the gossip dumps, each node is labelled with an implementation using its announced feature bits, applied as an ordered rule list, with operator knowledge and CLTV defaults as fallbacks. Two independent checks validate the labels: aliases that name their implementation, and the on-chain fingerprint of the wallet that built each node's funding transactions (LND's btcwallet uses nLockTime 0 and nSequence 0x0; Core/BDK-style wallets use anti-fee-sniping nLockTime and nSequence 0xfffffffd). Batched channel opens, where one node is common to every channel in the transaction, identify the funder exactly.

To detect outages, the author uses a peer-side signal rather than node self-reports: a node that is offline cannot announce anything, so its peers disable their direction toward it. For each client and day, the paper computes the share of channel directions pointing toward that client's nodes that peers mark disabled, unweighted and capacity-weighted. Client labels are frozen on 1 August 2026 so the incident cannot reshuffle cohorts, and the baseline is 1–25 August.

For the blast-radius analysis, the author takes the routable core (channels enabled in both directions) on 25 August 2026, models failures as a fraction of a client's nodes going unavailable, and measures the share of pre-failure node pairs still connected by a path of channels able to carry a 100,000-satoshi payment. These scenarios are compared against 94 ordinary two-day windows, against failures targeted by degree and by capacity, and against a null model that permutes client labels among nodes of similar capacity to remove any preference of a client for its own kind.

Why This Matters

Impact on research. Prior Lightning resilience work removed random or highly connected nodes from a snapshot, which says little about realistic failures that are partial, version-scoped and temporary. This paper supplies the first measured instance of a client-correlated outage in production and uses it to calibrate scenarios. It also reframes the diversity question: the useful metric is which software the hubs run, not how many nodes run each client, and a transport-level failure can be a common-mode failure too.

Real-world applications:

  • Incident response: monitoring peer-side disabled shares per client and per transport as a cheap early-warning signal, which was unambiguous for both the Tor event and the CLN shutdown.
  • Reporting: publishing client diversity weighted by hub capacity or reachability rather than by node count, so that concentration is not hidden by thousands of tiny nodes.
  • Network stewardship: encouraging channels between minority-client hubs (CLN, Eclair, LDK) so their users are not connected only through LND.
  • Operational guidance: scoping emergency advice by version and duration, and preferring --offline to shutdown because an offline node keeps watching the chain for revoked states and expiring HTLCs.

Industry relevance. Node operators, wallet and service providers, and exchanges that depend on Lightning liquidity now have evidence that the network absorbed a real client-wide failure, while the fragility sits with a handful of hubs — one of which (ACINQ) matters more to everyone else than the entire CLN cohort. The finding that 22% compliance left a third of those nodes still dark five weeks later is directly relevant to anyone deciding whether to follow the next emergency call.

Future Directions

  • Making diversity metrics hub-aware. The paper recommends weighting diversity by hub capacity or reachability; the natural next step is a standing metric that tracks the software of the most critical nodes over time and flags when a single operator's client choice becomes a single point of failure.

  • Seeing the unannounced graph. The gossip view excludes unannounced channels, where much LDK (wallets) and Eclair (Phoenix) usage lives, and those users reach the network through their providers' nodes. Measuring hub failures against payment success rather than connectivity would require data the paper says it does not have.

  • Modelling liquidity, not just connectivity. Balances are private, so the model does not capture how rerouted payments drain liquidity on surviving paths, nor fees, HTLC limits or private routing hints. The paper explicitly notes that absolute reachability levels overstate what payments achieve, leaving room for a liquidity-aware version of the same experiment.

  • Detecting hidden fleets and version-scoped exposure. The author found one managed fleet through its alias and hubs and notes that smaller fleets without a shared alias would go undetected. Gossip does not reveal software versions, so version-scoped risk — which is what emergency advisories actually target — remains inferred rather than observed.

Target Audience

Lightning protocol developers and implementation maintainers; node operators, routing node operators and Lightning service providers; researchers in blockchain measurement, network resilience and software monoculture risk; and analysts or risk teams at custodial wallets and exchanges who need to reason about common-mode failure in payment infrastructure. Readers wanting the headline result can take the findings section alone; readers wanting to reuse the classifier or the failure model will need the methods and validation sections, and the released code and aggregate datasets.

Authors’ abstract

Client diversity is often cited as the Lightning Network's defence against a bug in any one implementation, yet LND runs 84% of public nodes and holds 84% of public channel capacity, a share that has barely moved since 2023. We test what diversity buys with a natural experiment. On 26 August 2026, amid security fixes in every implementation, Core Lightning (CLN) maintainers told operators who could not upgrade to disconnect their nodes. Using 805 daily gossip snapshots,…

The full abstract on arXiv
Read the original paper