Digital Plumber

Archives
Log in
Subscribe
November 14, 2026

Nov 14: Network config change caused 15-hour Microsoft 365 outage

Digital Plumber

Plumbing the information age

Saturday, November 14, 2026  ·  No. 153  ·  16 stories

Today's 3 things that matter

  1. Network config change caused 15-hour Microsoft 365 outage
    Infrastructure operators need to understand cascade failure modes in control-plane upgrades and validation gaps in rollout procedures that can disable authentication and routing across multi-region cloud deployments.
  2. Overloaded air traffic routers grounded all US flights
    Network operators managing legacy systems and hybrid infrastructure need lessons on router capacity planning, backwards-compatibility testing, and graceful degradation when newer systems interact with decades-old equipment.
  3. Oracle releases OCCL collective communications plugin for GPU clusters
    Provides a concrete networking solution for scaling distributed AI training at fault-tolerant scale without stalling jobs on path failures—directly addresses cluster reliability constraints.

Full stories below, grouped by topic.


AIOps & Network Observability

Dynatrace SRE Agent reaches general availability for incident triage

releases.fru.dev · Nov 13, 2026 · Primary source

What happened: Dynatrace SRE Agent is now generally available to automate incident investigation and triage by analyzing incidents to determine root causes, assess severity, and route issues to appropriate teams with remediation proposals. November 2026 release also adds workflow templates for alert reduction, Kubernetes troubleshooting, and database operations.

Why it matters: The SRE Agent automates the MTTR-critical workflow of incident investigation—moving from detection to root-cause determination to team routing automatically without human triage—directly reducing mean time to resolution for network and infrastructure teams.

The Dynatrace SRE Agent is now generally available to automate incident investigation and triage by analyzing incidents to determine root causes, assess severity, and route issues to appropriate teams with remediation proposals. This represents a shift in how observability platforms support NOC operations: rather than presenting raw alerts or correlations to humans, the agent immediately begins investigating—pulling relevant metrics, logs, and traces to build root-cause hypotheses and propose next steps. Dynatrace released the SRE Agent for automated problem investigation and evidence gathering alongside new workflow templates for alert reduction, Kubernetes troubleshooting, and database operations. For network operations practitioners, the SRE Agent model is important because it demonstrates how agentic AI reduces manual handoff steps in incident response: detection → triage → investigation → remediation. Each step now has template-based automation rather than requiring human reading and dashboard navigation.

Read the original at releases.fru.dev


Data Center & AI Networking

Oracle releases OCCL collective communications plugin for GPU clusters

HPCwire · Nov 10, 2026 · Primary source

What happened: Oracle announced the Oracle Collective Communications Library (OCCL), a network plugin designed for distributed AI workloads on OCI-supported GPUs and RDMA-enabled infrastructure. OCCL routes collective communication traffic across OCI's multi-plane backend network to maintain resiliency during network path failures in large GPU clusters.

Why it matters: Provides a concrete networking solution for scaling distributed AI training at fault-tolerant scale without stalling jobs on path failures—directly addresses cluster reliability constraints.

Large-scale AI training depends on hundreds or thousands of GPUs exchanging tensors through collective operations (all-reduce, broadcast, etc.). In hyperscale clusters, a single congested link, NIC failure, or switch port failure can interrupt these synchronous communications, forcing training to stall or restart. Oracle's OCCL network plugin distributes collective communication traffic across OCI's multi-planar backend network, treating the data center fabric as a resilient, redundant mesh. This differs from single-path collective communication libraries (e.g., NCCL over standard Ethernet) that serialize traffic through a single fabric path. The plugin integrates with OCI's infrastructure—meaning it leverages existing RDMA capabilities (InfiniBand or RoCE) and the OCI backend's disaggregated switching architecture. For practitioners, OCCL moves fault tolerance from the application layer (job checkpointing, training restarts) into the network fabric itself, reducing job completion time variance in clusters above ~1,000 GPU scale.

Read the original at hpcwire.com

CoreWeave revenue grows 168% as GPU capacity stays scarce

The Washington Post · Nov 11, 2026 · Industry news

What happened: CoreWeave's Q3 revenue surged 168 percent while older Nvidia chips command higher rental prices. The neocloud provider is managing extraordinary demand for GPU capacity while scaling data center infrastructure and networking across multiple regions.

Why it matters: Signals sustained GPU capacity shortage driving infrastructure buildout at neoclouds—a direct indicator of data center networking demand for high-speed interconnects at scale.

CoreWeave reported $2.6bn quarterly revenue, $104bn contract backlog, 1.5 GW active power as of October 2026. Behind these numbers is a massive networking deployment: CoreWeave's platform includes a networking backbone supporting bandwidth for distributed AI workloads, foundational infrastructure with high-speed interconnects for complex AI workloads, and purpose-built storage with GPU-local caching. At 1.5 GW active capacity with 10× demand through 2027, CoreWeave is deploying hundreds of thousands of GPU interconnects—likely using InfiniBand or RoCE at 400G/800G scale. The neocloud model (CoreWeave, Fluidstack, Lambda building dedicated GPU data centers) drives homogeneous fabric deployments: every cluster runs the same networking stack, avoiding operational fragmentation. Neoclouds often run InfiniBand networking for multi-node training. For infrastructure teams planning cloud-native AI training, this reinforces the cost advantage of dedicated neoclouds over hyperscaler AI services—neoclouds front-load networking and power infrastructure, enabling better unit economics at massive scale.

Read the original at washingtonpost.com


Routing & Internet

Network config change caused 15-hour Microsoft 365 outage

Microsoft Service Health · Nov 14, 2026 · Primary source

What happened: Microsoft 365, Exchange Online, Teams, and SharePoint experienced a 15+ hour global outage starting November 12, caused by a network configuration problem introduced during an earlier capacity upgrade that surfaced after a subsequent network change. Affected 74 countries; root cause was servers failing health checks under load concentration.

Why it matters: Infrastructure operators need to understand cascade failure modes in control-plane upgrades and validation gaps in rollout procedures that can disable authentication and routing across multi-region cloud deployments.

On November 12, 2026, starting between 17:00–18:00 UTC, Microsoft 365 services including Exchange Online, Outlook, Teams, SharePoint, OneDrive, and Copilot went offline globally. The outage lasted over 15 hours and affected 430 million users across 74 countries. Microsoft's investigation identified the cause as a faulty network configuration update introduced during an earlier network-capacity upgrade that had remained dormant until a subsequent network change triggered cascading failures. Many servers on the Microsoft side failed automatic health checks and became unresponsive, concentrating load on remaining infrastructure and overwhelming TLS/authentication handshakes. The company confirmed no evidence of cyberattack or external interference, and the underlying problem was addressed by November 14. This incident illustrates the fragility of monolithic cloud control-plane architectures when configuration changes are not validated across state transitions, and the risk of hidden defects in infrastructure upgrades that surface only under specific operational conditions.

Read the original at thersac.icts.thers.ac.jp

Overloaded air traffic routers grounded all US flights

CNN / U.S. Secretary of Transportation statement · Nov 13, 2026 · Industry news

What happened: On November 13, the FAA grounded all U.S. domestic flights for 4 hours 21 minutes after air traffic control routers became overloaded with data following integration of new technology into legacy infrastructure. Secretary of Transportation blamed outdated routers; Homeland Security attributed it to new tech overwhelming the older framework.

Why it matters: Network operators managing legacy systems and hybrid infrastructure need lessons on router capacity planning, backwards-compatibility testing, and graceful degradation when newer systems interact with decades-old equipment.

On November 13, 2026, at 20:54 UTC (3:54 PM EST), the FAA issued a nationwide ground stop that lasted 4 hours and 21 minutes after a communication outage at Chicago air traffic control centers. The root cause was characterized as old air traffic control routers becoming overloaded and providing inconsistent readings after new technology was integrated into the existing framework. Secretary of Transportation Sean Duffy cited the routers as "incredibly old," while Homeland Security Secretary Markwayne Mullin noted that adding new technology to legacy infrastructure overwhelmed the system. The ground stop was lifted at 8:15 PM EST after the underlying issue was resolved. This represents a classic integration failure where capacity and state-management assumptions in legacy systems were violated by adding modern workloads without re-architecting the control path. Officials confirmed no relation to the concurrent Microsoft 365 outage and no evidence of nefarious attack.

Read the original at edition.cnn.com


Network Security

Microsoft patches critical LDAP, DNS and DHCP server flaws

OPSWAT / Zero Day Initiative · Nov 12, 2026 · Analysis

What happened: Microsoft November 2026 Patch Tuesday includes unauthenticated out-of-bounds write in LDAP service (CVE-2026-119958, CVSS 9.8) running on every domain controller, plus 10 DNS Server RCE flaws and 9 DHCP Server vulnerabilities (4 RCEs). Only one flaw is under active attack; none are publicly known.

Why it matters: LDAP RCE on domain controllers and DNS RCEs are critical because they affect the Windows infrastructure backbone that networks depend on. DHCP and DNS are network-facing services requiring immediate patching; prioritize patching DC-resident DNS and LDAP before routine systems.

CVE-2026-119958 is an unauthenticated out-of-bounds write in LDAP with CVSS 9.8 (network, no privilege, no interaction); on a domain controller, exploiting LDAP means executing in the directory service context with full AD privileges. CVE-2026-88218 is a Windows DNS Server RCE with CVSS 9.8 affecting critical infrastructure. DNS Server role has 11 total CVEs this cycle; 15 are information disclosure with CVSS 4.9, and 10 are RCEs. DHCP Server has 9 vulnerabilities including 4 RCEs. While only one flaw is under active attack and none are publicly known, the convergence of multiple critical flaws in core infrastructure services means network teams managing Windows DNS and LDAP must stage and test patches urgently in isolated lab environments first, then coordinate with AD/IAM teams on careful rollout sequencing. Check for DNS servers reachable from user VLANs, remote-access ranges, or the internet and patch DNS RCE flaws as one coordinated unit with CVE-2026-81384.

Read the original at opswat.com


People Moves

Jensen Huang resigns as Nvidia CEO, CFO Kress takes over

Nvidia investor relations · Nov 12, 2026 · Primary source

What happened: Jensen Huang resigned as President, CEO and board member effective immediately, with Colette Kress, Executive Vice President and Chief Financial Officer, appointed interim CEO. Kress has served as Nvidia's CFO since September 2013 and has played a key role in the company's growth and evolution.

Why it matters: As the world's leading AI infrastructure provider moves through CEO transition, the interim leadership shift signals continuity in financial strategy and investor relations during a critical period for AI chip supply chain management.

Nvidia's board accepted Jensen Huang's resignation on November 12, 2026, ending a 33-year tenure as the company's sole CEO since its 1993 founding. Colette Kress, who has held the CFO role since 2013 and led Nvidia's finance organization through its AI-driven growth trajectory, stepped into the interim CEO role while the board conducts a search for a permanent replacement. Notably, Kress will continue serving as Executive Vice President and CFO during her interim tenure, retaining dual responsibilities. Prior to joining Nvidia, Kress spent nearly 25 years in finance roles at Cisco and Microsoft. Kress's appointment as interim CEO carries particular weight for infrastructure practitioners: as CFO, she has been intimately involved in Nvidia's capital allocation, supply chain partnerships with cloud hyperscalers and OEMs, and the company's financial commitments to customers unable to immediately afford GPU capacity. Her continuity in financial leadership provides stability for the complex procurement arrangements that define relationships between Nvidia and Meta, Microsoft, Google, OpenAI and other major AI infrastructure buyers.

Read the original at investor.nvidia.com

Meta hires ex-AWS executive Dave Brown to run infrastructure

CNBC · Nov 11, 2026 · Industry news

What happened: Meta's head of infrastructure and co-head of engineering Santosh Janardhan will leave in 2027, with former AWS senior vice president David Brown set to lead Meta's infrastructure efforts of the Meta Compute initiative. Brown was a senior vice president at AWS before joining Meta in July; Janardhan will assist Brown during the handoff, staying on through 2027's first quarter.

Why it matters: Leadership transition in Meta's hyperscale infrastructure operations—managing $135–145B capex for AI data centers—signals strategic shift in data center architecture and AI workload prioritization under AWS cloud infrastructure expertise.

Meta announced November 11 that Santosh Janardhan, head of infrastructure and co-head of engineering since 2009, will depart in early 2027 as Dave Brown assumes full responsibility for the company's global infrastructure operations. Brown, who spent 19 years at AWS focusing on compute, cloud and AI services, joined Meta in July 2026 as VP of Infrastructure but will now formally lead infrastructure and serve as chief technology officer of the Meta Compute initiative spearheaded by President Dina Powell McCormick and technical leader Daniel Gross. Janardhan's tenure saw Meta build out massive data center footprints supporting Facebook, Instagram and WhatsApp; he and Brown will overlap through Q1 2027 for transition. For operations practitioners, this matters significantly: Meta is currently directing $135–145 billion in annual capital to data center expansion for AI training and inference at hyperscale. Brown's AWS background—particularly with AWS Graviton silicon and compute services architecture—suggests potential shifts in Meta's approach to custom silicon strategy (MTIA), network fabric design for GPU clusters, and infrastructure spend priorities relative to competing hyperscalers.

Read the original at cnbc.com


Agentic AI & MCP

Active MCP servers pass 118,000 as registry consolidates ecosystem

DEV Community · Nov 14, 2026 · Industry news

What happened: HostedMCP catalog reported 118,434 active MCP servers and 7,500 skills as of November 14, 2026, with over half now flowing through the official MCP Registry launched in 2025. Of 15,380 scanned servers, 10,327 meet verification standards, representing the practical consolidation of the MCP ecosystem for production infrastructure automation.

Why it matters: Network and SRE teams evaluating MCP server coverage for infrastructure agents now have real-time data on server maturity: verified servers enable trusted integration across network management, observability, and cloud platforms without custom tool-building.

The HostedMCP database snapshot on November 14, 2026 shows the MCP ecosystem has matured significantly since the protocol's 2024 introduction. Of 118,434 cataloged servers, 63,847 have known transport types (majority stdio, followed by HTTP-based), and only a small fraction rely on authentication mechanisms beyond basic API keys. The verification trend matters: 10,327 of 15,380 recently scanned servers (67%) meet the Verified rule, indicating reproducible, stable deployments. The bulk of new servers come through the official MCP Registry (>50% share), which launched in 2025 and now acts as the de facto source of truth. For infrastructure teams, this concentration around a curated registry reduces the burden of vetting third-party MCP servers and lowers the surface area for supply-chain risk. The data also reveals transport consolidation: stdio remains dominant for local agents, while HTTP-based (Streamable HTTP and hosted options) powers remote deployments. This is directly relevant for organizations deploying agents in Kubernetes, cloud functions, or distributed edge environments, as the ecosystem has settled on transport patterns that match modern operational topology.

Read the original at dev.to

Google Cloud adds agentic HPC cluster tools with deterministic checks

CASRAI · Nov 14, 2026 · Industry news

What happened: Google Cloud announced agentic HPC cluster tools and DeepMind science skills on November 13, 2026, with a framework equipping AI agents with deterministic validation routines published through skill repositories instead of granting unconstrained LLM command-line execution. The Declarative Cluster Toolkit addresses historically error-prone manual orchestration of scientific supercomputing workflows.

Why it matters: For network and infrastructure operations teams running complex, multi-stage computational workflows or managing high-performance clusters, this model of constrained-yet-autonomous agentic execution—validated through published skill repos rather than raw LLM access—is directly applicable to production infrastructure automation and incident remediation workflows.

Historically, configuring supercomputing environments for scientific workflows has demanded manual, error-prone orchestration across disparate software tiers. Google Cloud's approach solves this by decoupling agent reasoning from execution: agents invoke pre-validated skills (deterministic routines with known inputs, outputs, and side effects) rather than generating shell commands or configuration patches on the fly. For NetOps and AIOps practitioners, this architectural pattern is highly relevant. Constrained agent execution via skill repositories mirrors governance patterns needed for production network operations: agents diagnose issues and propose changes, but the actual remediation uses only pre-approved, tested runbooks. This avoids the common pitfall where agentic systems generate plausible-but-incorrect configurations because the agent hallucinated a network command or miscalculated CIDR notation. The framework is particularly valuable for multi-step operational tasks—provisioning clusters, scaling workloads, or remediating widespread infrastructure issues—where a single agent mistake can cascade across hundreds of nodes. By publishing skills as code (similar to OpenConfig or Ansible modules), teams build an auditable, versioned library of safe operations that agents can call with confidence. Google's approach scales to both individual teams (publishing internal skills) and organizations (federated skill repositories across business units), making it a practical model for enterprise NetOps adoption.

Read the original at casrai.org

NTT DATA publishes framework for production agentic AI

NTT DATA · Nov 13, 2026 · Primary source

What happened: NTT DATA published a framework on November 13, 2026 addressing the operational, security and governance risk of agentic AI, presenting practical frameworks to assess readiness, strengthen architecture and measure trust for structured paths toward secure, reliable and scalable enterprise deployment. The guidance covers strategic, reasoning, runtime, governance and security challenges that prevent agents from scaling.

Why it matters: Network and SRE teams moving agentic operations from pilots to production need practical assessment frameworks for readiness across agent design, governance, infrastructure, and evaluation—this paper directly addresses the operational trust and control requirements that determine whether agents can safely handle production changes without human intervention.

NTT DATA's framework covers: identifying strategic, reasoning, runtime, governance and security challenges that prevent AI agents from scaling; assessing production readiness across business value, data, agent design, governance, infrastructure and evaluation; applying graduated autonomy, guardrails, human oversight, observability, cost controls and secure identity management; evaluating architectural compliance across governance, security, responsible AI, resilience and regulatory requirements; and measuring operational trust through task success, constraint adherence, intervention rates and execution complexity. For infrastructure operations, the distinction between advisory agents (that recommend actions) and autonomous agents (that execute changes) hinges on these readiness dimensions. A well-governed agent architecture requires: (1) runtime execution sandboxes that enforce policy-as-code limits (e.g., OPA), (2) human approval gates for high-risk actions scaled by impact radius, (3) observability that traces every agent decision back to its reasoning chain, and (4) evaluation frameworks that measure not just whether the agent achieved its goal, but whether it did so safely and cost-effectively. Graduated autonomy is particularly relevant: starting agents in advisory mode (they diagnose and recommend, humans approve), moving to conditional mode (they execute if confidence > threshold and impact is low), and only advancing to autonomous when the organization has achieved 99%+ task success and zero safety incidents over a sustained period. The paper's focus on measurable trust—intervention rates, constraint adherence, execution complexity—gives ops teams concrete metrics to track during rollout.

Read the original at nttdata.com


Telco & Cable AI

GSMA and ITU agree to collaborate on open telecom AI

RCR Wireless News · Nov 12, 2026 · Industry news

What happened: GSMA and ITU signed a non-binding letter of intent to collaborate on the Open Telco AI program, addressing operator concerns about over-dependence on a handful of AI providers for network operations. AT&T reports its telco-trained models reach roughly 90% on network operation benchmarks.

Why it matters: Operators are standardizing on open telecom AI models to avoid vendor lock-in for mission-critical network functions, directly reducing operational risk and enabling interoperability across vendor stacks in autonomous network deployments.

The GSMA and ITU collaboration on Open Telco AI addresses a strategic vulnerability: only four or five large AI providers currently control most telecom network intelligence. This dependency creates operational risk if access is withdrawn or models diverge. The framework covers telco-grade AI for network operations, infrastructure and standards, with workstreams addressing domain-specific model development, benchmarking, and data sharing. AT&T's telco-trained models reportedly achieve roughly 90% accuracy on industry benchmarks, topping the Open Telco AI leaderboard—demonstrating that operator-developed models can match or exceed general-purpose AI for network tasks. GSMA Intelligence estimates AI currently contributes roughly 2.5% of operator revenue, split 62/38 between cost-saving and revenue-generating objectives. Participants include China Telecom, China Unicom, and China Mobile, with Huawei as a sponsor and data contributor. The effort establishes specifications and evaluation frameworks so operators can swap AI models without retraining entire automation pipelines—critical for networks moving toward autonomous operations where model errors directly impact service quality and customer trust.

Read the original at rcrwireless.com


Research, Standards & Industry

Janus framework unifies digital twins and network orchestration

arXiv · Nov 13, 2026 · Research

What happened: Paper accepted to IEEE GLOBECOM 2026 proposing an agentic LLM-based framework unifying digital twin construction and network orchestration for next-generation wireless networks. Demonstrates AI-native autonomous wireless network orchestration.

Why it matters: Addresses dynamic model updates for network operations; enables adaptive orchestration for NetDevOps engineers building autonomous network management systems.

Janus combines agentic LLM capabilities with digital twin construction and network orchestration in a single closed-loop system. Existing approaches decouple twin construction from orchestration, limiting adaptability. The framework demonstrates effectiveness, scalability, and generalizability for AI-native autonomous wireless network orchestration. Results show practical fixed-time leader-following consensus of true unknown nonlinear multi-agent systems. This addresses a critical gap in network operations by allowing dynamic model updates tied directly to orchestration decisions, relevant for NetDevOps engineers building autonomous network management systems in large-scale, heterogeneous, dynamic environments.

Read the original at arxiv.org

LLM quality and cost vary widely on optical network tasks

arXiv · Nov 13, 2026 · Research

What happened: Paper evaluating LLMs for optical network automation with human grounding, examining how output quality and inference cost vary substantially across LLM families for operational network tasks.

Why it matters: Critical for NetDevOps teams evaluating LLM-based automation tools; shows which LLM families produce reliable outputs at acceptable cost for network operations.

Research provides human-grounded evaluation of LLMs for optical network automation, addressing practical deployment questions about LLM reliability and inference cost. Cross-listed between cs.NI, cs.AI, and cs.LG on arXiv, the work bridges academic rigor with operational reality. The human-grounding aspect ensures results reflect actual network automation requirements—configuration generation, troubleshooting, intent translation—rather than synthetic benchmarks. Results show performance variance across LLM families on optical domain tasks. Directly applicable to teams considering LLM-based tools for network operations, intent-driven configuration, and automated remediation. Essential baseline for organizations evaluating whether LLM adoption improves MTTR or introduces operational risk.

Read the original at arxiv.org


AI Industry & Policy

White House urges easing liability to spur AI incident reporting

Bloomberg · Nov 14, 2026 · Industry news

What happened: White House AI task force officials recommended easing cyber liability policies to incentivize AI companies to report critical incidents, advising that previous regulatory standards may have a chilling effect on information sharing.

Why it matters: This marks a policy shift affecting how AI companies report critical incidents to government agencies and could reduce regulatory enforcement pressure on disclosure.

At a White House meeting on November 14, officials from President Donald Trump's artificial intelligence task force recommended that agencies ease their cyber liability policies to incentivize AI companies to report critical incidents. The officials advised that previous standards for taking regulatory action in response to cyber episodes may be outdated and have a chilling effect on important information sharing. This represents a significant policy adjustment from the Trump administration, suggesting a regulatory approach prioritizing disclosure over enforcement. The shift acknowledges tension between compliance and transparency: stricter liability rules discourage self-reporting, while relaxing them could accelerate incident disclosure but potentially reduce accountability. For enterprises operating AI systems, this signals a more liability-friendly environment from federal authorities, though it may not affect state-level enforcement in California, Colorado, or New York.

Read the original at bloomberg.com

Nebius plans $40 billion AI cloud capex for 2027

CASRAI · Nov 14, 2026 · Industry news

What happened: Nebius Group announced November 12 capex projections of $40-45 billion for 2027 to expand GPU clusters and data centre operations, following 2026 capex revision to $27 billion, reflecting accelerating industrial demand for AI training and inference capacity.

Why it matters: This $40-45 billion capex commitment signals major infrastructure consolidation and GPU supply chain strategies; enterprises should evaluate capacity availability and pricing impact for 2027 AI workloads.

AI cloud infrastructure provider Nebius Group announced November 12 an ambitious capex roadmap projecting $40-45 billion in planned spending for 2027 to expand GPU clusters and data centre operations, revised from $27 billion for 2026. This reflects accelerating industrial demand for specialized AI model training and inference capacity backed by customer prepayments and long-term enterprise contracts. However, Nebius disclosed operational delays at third-party co-location sites affecting deployment schedules. This capex commitment is significant for enterprises evaluating GPU supply chains and inference costs. The customer prepayments indicate strong confidence in sustained AI workload demand, but infrastructure bottlenecks at co-location sites may affect 2027 deployment timelines and GPU availability—important considerations for organizations planning AI infrastructure procurement or cost forecasting.

Read the original at casrai.org


Read this edition on the web · The week in network intelligence · Vendor Radar

Digital Plumber is AI-curated and AI-summarized, with no human review before publishing. Verify before acting on anything here. How it works.

Don't miss what's next. Subscribe to Digital Plumber:
← Newer Nov 15: Juniper router feature caused nationwide FAA ground stop Older → Nov 13: Critical FortiOS flaw lets attackers forge BGP and OSPF routes
Powered by Buttondown, the easiest way to start and grow your newsletter.