Cloudflare status: hosting issues and outage reports

Problems in the last 24 hours

The graph below depicts the number of Cloudflare reports received over the last 24 hours by time of day. When the number of reports exceeds the baseline, represented by the red line, an outage is determined.

Most Reported Problems

The following are the most recent problems reported by Cloudflare users through our website.

• Domains (42%)
• Cloud Services (30%)
• Hosting (16%)
• Web Tools (8%)
• E-mail (4%)

Live Outage Map

The most recent Cloudflare outage reports came from the following cities:

City	Problem Type	Report Time
Crisfield	Domains	8 hours ago
Nanaimo	Web Tools	1 day ago
New York City	Web Tools	2 days ago
Istanbul	Domains	5 days ago
Greater Noida	E-mail	7 days ago
Paris	Domains	8 days ago

Full Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

Cloudflare Issues Reports

Latest outage, problems and issue reports in social media:

• 
  EstePrime (@EstePrimeWorld) reported 7 minutes ago

  @uwunetes He thinks that tagging Cloudflare and Cloudflare employees will help him. There is competition between those companies, but he is literally using the shadcn name, which has nothing to do with fair competition.

• 
  whatnicktweets (@whatnicktweets) reported 9 minutes ago

  @boristane DO VPS => Never Cloudflare => Daily

• 
  Mafiova (@mafiova) reported 12 minutes ago

  @ExpLang_Cn having problem with cloudflare

• 
  Augustinas Malinauskas (@amgauge) reported 12 minutes ago

  @adahstwt Research on other platforms, buy on aws or cloudflare - they don’t seem scammy. Squarespace is the worst product in existence, very sad that googledomains sold

• 
  Simon Ábris (@simonxabris) reported 14 minutes ago

  @boristane i use cloudflare and i almost never open the dashboard now, i just use the mcp

• 
  shabingus (@_FaceFTW) reported 15 minutes ago

  As much of a Rust fan that I am, it is not a magical solution for correctness, something that AI can be prone to getting wrong. Remember when Cloudflare went down because of unwrap()? Regardless of who wrote it the burden of correctness is still on the programmer.

• 
  Apollo (@Apollo_onchain) reported 20 minutes ago

  @Kathydotxyz @Cloudflare + you can verify this with @bankrbot and you can claim the fees to help fund your agent. 0x73f29bc14fb50c65c508c77c9e1dbc566b627ba3

• 
  Isaac (@irg111) reported 23 minutes ago

  @JulienHoez @Cloudflare Are you sure this is cloudflare? Wtf

• 
  EVY TECHNO (@EvyTechno) reported 27 minutes ago

  @Oblivious9021 You don’t test harmless-looking bugs by asking: “Does it work?” You test them by asking: “What happens when this scales faster than human intuition?” That Cloudflare outage wasn’t caused by malware. It was caused by computational explosion. A single regex triggered catastrophic backtracking. Which means: the regex engine explored exponentially growing match paths until CPUs hit 100%. The dangerous part? The rule probably looked completely reasonable in code review. That’s the real lesson. Most production disasters are not caused by obviously broken systems. They’re caused by locally-correct decisions interacting globally. So how do you test for that? Not with normal unit tests. You need adversarial testing. You intentionally search for worst-case behavior. For regex specifically: • fuzz inputs • test pathological strings • measure execution complexity • enforce regex timeouts • benchmark CPU cost per request • use linear-time regex engines where possible Because correctness is not enough. Complexity matters. A regex that succeeds in 2ms normally but takes 30 seconds on crafted input is effectively a denial-of-service vulnerability. And this principle exists everywhere in engineering: • hash collisions • recursive parsers • N+1 queries • lock contention • retry storms • cache stampedes • exponential algorithms The bug is often not: “Does it fail?” The bug is: “How badly does it fail under amplification?” That’s why mature systems engineering focuses less on feature correctness… and more on blast radius containment. Because eventually: every harmless assumption meets production scale. And production scale is where hidden complexity becomes physics.

• 
  Remkus de Vries (@remkusdevries) reported 29 minutes ago

  @TheCre8tiveDiva @Within_WP It’s geared towards someone wanting to understand Cloudflare better. There’s tons inside of Cloudflare that people don’t know how to wield towards WordPress. I want to fix that. WooCommerce will be covered!

• 
  Raunak (@raunakdoesdev) reported 30 minutes ago

  @benhylak @thdxr @Cloudflare oh really? i’ve never actually used it just had my llms do so and they usually manage

• 
  Furkan (@TheRealFurkan) reported 31 minutes ago

  @hogrbe Why not move it behind Cloudflare? it would fix all your issues.

• 
  BootStepper (@BootStepper) reported 33 minutes ago

  @tan_stack I am trying to use TanStackAI and `cloudflare/tanstack-ai`, but its giving type check issues about "gpt-5.4" not being a valid model -- Is this something you or @Cloudflare can fix?

• 
  Bree Sharp | Local SEO Strategist (@bree_sharp) reported 36 minutes ago

  After moving MWOV from SiteGround to Cloudflare: Ahrefs went from returning 1 URL per crawl to crawling the site normally. The culprit was LiteSpeed's bot management — aggressive enough to rate-limit Ahrefs into giving up after the first URL. GSC showed 59 pages indexed fine. The problem wasn't the site. It was the layer between the site and the crawlers. Sometimes the issue isn't your content or your configuration. It's your infrastructure making decisions you don't know about.

• 
  Vlad P (@orange_boy) reported 38 minutes ago

  @eastdakota @Cloudflare Sorry, cant find better channel to reach support because seems like CF account is blocked from everything for some reason. The limits on acc is lower than free (20 domain redirects instead of 10,000 i.e.), cant add members, cant have API key. Acc is 6-7 yo and there is few errors with no names and few errors stating that email isn't verified. And there is no way to request verification. And when I submit a ticket to support I cant even see it and when I click to my tickets I see only this page and that's it. Cant ask ai because it cannot issue read only API token and crashes with no name error. But if I try to create in another part of UI i see error about email verifications. Looks like edge case and that ac is fully out of coverage. Please, we need to fix some SEO redirects ASAP. Spent all weekend trying to figure out CF issues.

• 
  Pranesh Prakash (@pranesh) reported 38 minutes ago

  Folks from Cloudflare make the case that the "post-Mythos moment" requires defence-in-depth more than raw speed in detection and patching of vulnerabilities. (The post is also self-serving, given CF's service offerings.) #genAI #security

• 
  Prashank Abhishek (@prashank25) reported 39 minutes ago

  @boristane Every day. @digitalocean @Cloudflare give me option to persist session for a week at least or make the 2nd factor remember the device because it's every login right now. Too annoying 🫩

• 
  . (@r2fat3li) reported 39 minutes ago

  @Ahmeed2m @Bashmohandes From my understanding, caddy uses the cf api to renew certs, but the resolving is still done by cloudflare statically (does not change based on if I am connected to my home network or not) and only for public domains. I could buy a domain, but what about the static resolving?

• 
  Johnmark Obiefuna (@jayhemz) reported 43 minutes ago

  Proxying through Cloudflare solves almost all Wordpress security issues.

• 
  Lars (@larsbuilds) reported 46 minutes ago

  @ImLunaHey @Cloudflare With smtp support? :))

• 
  DataDan｜AI Data Engineering (@ba_niu80557) reported 48 minutes ago

  A function receives a webhook, validates it, queries a database (150ms network round-trip), and returns a response. Total wall-clock time: 170ms. Actual CPU time: 5ms. AWS Lambda bills you for 170ms. Cloudflare Workers bills you for 5ms. Same function. Same result. 34x billing difference — because one platform charges for time your code spent waiting, and the other charges only for time your code spent computing. (Source: Morph, "Cloudflare Workers vs AWS Lambda 2026", April 2026) That billing model difference is the most underappreciated shift in backend infrastructure in 2026. And it's quietly reshaping how production systems are built — not just for edge use cases, but for everything. Here's the thesis I keep arriving at after watching teams migrate over the past 18 months: The "deploy to a region, scale with containers" model that dominated backend engineering from 2015-2024 is being replaced by an "deploy everywhere, scale with isolates" model. And most backend engineers haven't noticed because the migration is happening one function at a time. Baselime reported 80% lower cloud costs after migrating from AWS to Cloudflare. Not 8%. Eighty percent. (Source: Morph, April 2026) The numbers are that dramatic because three structural differences compound: Difference 1: Cold starts don't exist anymore. Lambda cold starts: 100ms to 3,000ms depending on runtime, package size, and VPC config. Java Lambda in a VPC? You might wait 3 full seconds before your code runs a single line. Cloudflare Workers cold starts: under 5 milliseconds. Effectively zero. Because Workers don't spin up containers. They run V8 isolates — the same lightweight sandboxing technology that runs your Chrome browser tabs. For a web API serving human users, 100ms cold start is noticeable but tolerable. For an AI agent making 200 API calls per session, cold starts compound catastrophically. 200 calls × 500ms average cold start = 100 seconds of dead time per session. The agent is waiting for infrastructure, not computing. (Source: Morph, April 2026) This is why every serious AI agent infrastructure team I know is evaluating edge-first deployment. Not because edge is trendy — because their agents are burning money and latency on cold starts that V8 isolates eliminate entirely. Difference 2: Global distribution is the default, not the exception. You deploy a Lambda function. It runs in us-east-1. A user in Tokyo hits your API. Their request travels 11,000 km to Virginia, your function processes it, and the response travels 11,000 km back. Round trip: 200-400ms of pure network latency before your code does anything. You deploy a Cloudflare Worker. It runs in 330+ cities worldwide. A user in Tokyo hits your API. The request reaches a Cloudflare edge node in Tokyo. Your code runs there. Response returns from Tokyo. Round trip network latency: effectively zero. This isn't edge computing as a niche optimization. This is "every function is global by default" as the deployment model. You don't choose a region. There is no region. Your code runs wherever the user is. For a traditional CRUD API, this reduces TTFB by 60-80%. (Source: Digital Applied, "Edge Computing: Cloudflare Workers Dev Guide 2026", January 2026) For AI agent endpoints that serve users across time zones — a customer support agent used by a global company, a coding assistant used by distributed teams — the latency reduction is the difference between "feels instant" and "feels slow." Difference 3: The ecosystem became a full stack. The reason edge computing stayed niche from 2018-2023 was simple: you could run code at the edge, but your data was still in a region. Every edge function that needed a database round-tripped to us-east-1 anyway, killing the latency advantage. In 2026, Cloudflare solved this by building an entire data layer at the edge: → D1: SQLite at the edge. Global read replication. Your queries run where your users are. → KV: Key-value storage with edge caching. Sub-millisecond reads globally. → R2: Object storage. S3-compatible. Zero egress fees. (This alone saves thousands/month for media-heavy applications.) → Durable Objects: Stateful computing at the edge. Strongly consistent, globally coordinated state — the thing that was impossible at the edge until 2024. → Queues: Message queuing with guaranteed delivery. → AI inference: Run ML models on serverless GPUs at the edge. → Vectorize: Vector database for semantic search at the edge. (Sources: Cloudflare Workers docs; Calmops, "Edge Computing with Cloudflare Workers", March 2026) This changes the calculus completely. In 2022, edge was "run your CDN logic there." In 2026, edge is "run your entire application there" — database, storage, queues, AI inference, state management. Full stack. The framework ecosystem caught up too. Hono — under 14KB, zero dependencies, Express-like routing — became the standard routing framework for Workers in 2026. You write code that looks almost identical to Express/Fastify, but it runs globally with zero cold starts. What this means for how you should think about your next backend project: The decision tree has changed: Is your workload I/O heavy (API calls, database queries, webhook processing)?→ Workers bills CPU time only. You pay for 5ms of compute, not 170ms of waiting. The cost difference is 10-34x. This is most web backends. Does your application serve users globally?→ Workers runs in 330+ cities automatically. No multi-region deployment to manage. No cross-region replication to configure. Global is the default. Does your application need zero cold starts?→ Workers uses V8 isolates: sub-5ms startup. Lambda uses containers: 100ms-3s startup. If you're serving real-time AI agents, chatbots, or latency-sensitive APIs, cold starts are unacceptable. Does your workload need heavy compute (video transcoding, ML training, data processing)?→ Lambda. Workers caps at 128MB memory and has CPU time limits. For compute-heavy tasks, Lambda's 10GB memory and 15-minute execution are necessary. Are you deeply integrated with the AWS ecosystem (DynamoDB, SQS, S3 triggers, Step Functions)?→ Lambda. Workers can't trigger on S3 events or consume DynamoDB streams. Migrating away from Lambda means migrating away from the AWS event-driven ecosystem. The honest assessment: 80% of web-facing backend functions are I/O-heavy, globally distributed workloads where Workers is structurally cheaper and faster. 20% are compute-heavy or AWS-locked workloads where Lambda is the right choice. Most teams are running 100% on Lambda because that's what they learned in 2018. The AI angle that ties this back to my usual topics: Every AI agent infrastructure pattern I've written about — MCP servers, tool endpoints, RAG retrieval APIs, model routing gateways, cost tracking middleware — is an I/O-heavy workload that serves global users and needs zero cold starts. These are exactly the workloads where edge-first architecture delivers the largest improvement over traditional serverless. An MCP server on Lambda: cold start + regional latency + wall-clock billing = slow and expensive. An MCP server on Workers: zero cold start + global distribution + CPU-only billing = fast and cheap. The infrastructure layer beneath AI agents matters as much as the orchestration layer above them. Most agent architecture discussions focus on LangGraph vs CrewAI and ignore the fact that the function layer underneath is adding 100+ seconds of dead time per session to cold starts. Three uncomfortable questions for any backend team in 2026: 1) What percentage of your Lambda invocation time is your code actually computing vs waiting for I/O? If you're not measuring this — you're paying for wait time. For most web APIs, CPU time is 3-10% of wall-clock time. The other 90-97% is network round-trips that Lambda bills you for and Workers doesn't. 2) Where are your users, and where is your code? If users are global and code is us-east-1 — you're adding 100-300ms of pure network latency to every request. Workers eliminates this by running your code where your users are. Automatically. 3) When was the last time you evaluated whether your serverless architecture is still the right one? If "when we set it up in 2020" — the infrastructure landscape has fundamentally changed. Edge-first wasn't viable in 2020. It is in 2026. A 2-day migration experiment on a single non-critical endpoint will tell you whether the cost and latency improvements are real for your workload. The thesis: → 2016-2020: "serverless means Lambda" → 2021-2024: "edge is interesting for CDN logic but not real backends" → 2026: "edge-first is the default for I/O-heavy global workloads, and traditional serverless is the fallback for compute-heavy regional workloads" The inversion already happened. Most backend engineers are still deploying to Lambda because that's what the tutorials taught them in 2018. The teams that re-evaluated are running the same functions at 60-80% lower latency and 80% lower cost. Same code. Different infrastructure. Dramatically different bill. The boring infrastructure migration wins. It always does. Especially when the exciting AI agent is waiting 100 seconds for cold starts nobody measured.

• 
  Sooraj (@suryanox7) reported 51 minutes ago

  @itechnologynet I never seen a request go through for paid models on openrouter without 429. But if you have smart model router.. You don't pay much on cloudflare and having to try it for free is another plus for me

• 
  Adel Bucetta (@adelbucetta) reported 54 minutes ago

  @Omolaranife most web3 projects stall not because of bad ideas but because they can't be built to scale without 10+ developers who know how to deploy to cloudflare, AWS or google cloud.

• 
  Mike Gyi (@mikegyi) reported 55 minutes ago

  Finally sorted out my cloudflare la liga problem.

• 
  dus🍵 (@Mookkaa15221) reported 56 minutes ago

  @gelbooru They say you are hosting 'p but they don't call the cops immediately, methink cloudflare is full of ****.

• 
  PhoneGuy (@phoneguy284) reported 57 minutes ago

  @EWess92 I get blocked by Cloudflare. I use Proton VPN. I've had this problem with other sites.

• 
  Sr Carlos ²³²U (@CJavierSaldana) reported 59 minutes ago

  For me, the best usage of the Codex /goal command: /goal Fork this MIT-licensed xxxxx/yyyy repo and convert it to a Cloudflare-native version. Be destructive: remove all external service dependencies and follow our cloudflare-clone-security-check.

• 
  刘米其丶 (@Liu_Mickey_Ai) reported 1 hour ago

  Due to the Cloudflare outage, all unode services have been affected. All users just need to be patient and wait

• 
  Alper FERUDUN (@AlperTheKing) reported 1 hour ago

  @Cloudflare Retail AI fraud is an event-graph problem across account creation, promo abuse, card testing, inventory hoarding, returns, and checkout velocity. Defense needs identity, behavioral telemetry, and rate limits.

• 
  Anni **** (@aninibread) reported 1 hour ago

  @InderpreetSingh @Cloudflare hey @InderpreetSingh I work on AI Search, happy to help. AI Search now has hybrid search which combines BM25 with vector search. What are you using now for search?