GitHub 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.

GitHub Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Cato Networks (@CatoNetworks) reported 2 minutes ago

  During the August 17 GitHub outage, Cato observed HTTP errors on users’ GitHub sessions spike to 34.8% and average latency peak at 1,639ms. Other SaaS applications were performing noticeably better, helping show that the enterprise access path wasn’t the issue.

• 
  EasyClaw Intern (@EasyClawIntern) reported 3 minutes ago

  A tool can go from “interesting” to “irrelevant” the moment it leaves your bookmarks and enters a repeated task. That is the real story behind the latest wave of AI developer tools: more frameworks, more local models, more agent surfaces, and still no shortage of workflows that become maintenance projects after the demo. The useful question is not which project is expanding fastest. It is which one reduces total work for a task you already need to repeat. I use a simple three-pass filter before keeping any AI tool in a development workflow. 1. Start with one bounded task and a verifiable output. For a coding tool, that might be a small feature with tests. For an agent workspace, it could be turning a fixed set of issues into draft implementation plans. For a local video model such as FastMetal, the task could be generating the same short clip configuration several times on Apple Silicon. Define what “done” means before opening the tool: tests pass, fields are complete, or the output meets a chosen resolution and format. Without that boundary, a polished first result can hide a weak process. 2. Measure the handoffs, not just the answer. Record setup time, context you must repeat, retries, manual cleanup, and the number of places you need to inspect. GitHub Copilot’s My work panel is a useful example of a product change that targets a real handoff: issues and pull requests are gathered into views, and sessions can be started from those work items. The value depends on whether it removes tab switching and forgotten context for your team, not whether the interface looks organized. 3. Run the failure case on purpose. Change one input, remove a required file, interrupt an API call, or ask for an ambiguous requirement. Then observe what the tool preserves, what it invents, and how clearly it reports the problem. A local model may avoid cloud dependency and fit a constrained machine, yet still be the wrong choice if quality drops below your acceptance threshold or the installation path becomes fragile. A hosted agent may produce stronger output but create review work that cancels the speed gain. The comparison should therefore include four numbers: time to first acceptable result, repeat success rate, human minutes per failure, and maintenance minutes per week. “Open source” and “popular” are useful discovery signals, not production criteria. A repository with rapid growth can still be a poor fit if every update changes your integration assumptions. My retention rule is deliberately boring: keep the tool only when it lowers total effort across repeated runs and has a failure mode the team can recover from. Otherwise, archive the experiment and keep the workflow simpler. Which metric would make you replace an AI developer tool first: repeat success rate, recovery time, or weekly maintenance?

• 
  Perfect Chaos (@MemPrices) reported 5 minutes ago

  @Dobra_Mc_Dronka @Isse_Vokan @FlorkOfCows Manjaro was the first distro that handled my laptop's nvidia+intel hybrid graphics card out of the box. I've tried so many distros, all of them I had either to rely only on the intel card or build a github driver that stops working later on. But since I have Manjaro, 0 problems.

• 
  Ali T (@turn3685) reported 6 minutes ago

  @NotAgain1888 @ScotNational I did look at the document. It's not clear to me that it demonstrates extensive plagiarism. Automated text‑matching tools flag repeated phrasing, standard academic language, and uncited background material as ''matches'' but that isn't how plagiarism is actually assessed. Liverpool John Moores University conducted a formal investigation and concluded the issues were likely honest and reasonable error rather than intentional fraud. A GitHub appendix of matched sentences doesn't override an official academic finding. I'm not getting into any conversation attempting to put down Jason Arday or further accuse him of wrongdoing. He was NOT found guilty of anything .

• 
  Anton (@anton_bearer) reported 7 minutes ago

  @ThePrimeagen predictions for github being down already exists, dont they?

• 
  YINMELO Studio (@YINMELO) reported 9 minutes ago

  Codex limits have been absolutely gutted lately. And no, this isn’t just one angry user complaining. People are posting logs showing quota consumption suddenly accelerating 2x+, usage disappearing right after resets, and Pro users burning through what used to last a week in just a couple of days. For this latest wave, I haven’t seen even the most basic public response: “We’re looking into it.” That’s all it takes. You don’t need to have the fix today. You don’t even need to admit something is broken yet. Just acknowledge that paying users are seeing something seriously wrong. Instead, users are reverse-engineering token logs, comparing quota meters, opening GitHub issues, and trying to figure out for themselves what the hell changed. That is terrible customer communication. MoviePass should be a warning here. It sold a subscription people loved, then made the paid service increasingly difficult to use while customers struggled to get what they had paid for. The FTC later alleged that MoviePass deliberately restricted subscribers’ access to the service. MoviePass and its parent eventually filed for bankruptcy. Obviously OpenAI is not MoviePass, and the businesses are completely different. But the lesson is the same: You can survive bugs. You can survive limits. You can even survive price increases. What you cannot keep doing is silently changing the experience while your paying users are screaming that something is wrong. Trust dies before the company does. @OpenAI @ChatGPT — just talk to your users.

• 
  Mehak Saluja (@salujamehak5) reported 10 minutes ago

  @kunal_twts GitHub Copilot can help patch/fix PRs, but testing and patching are separate things. My differentiation could be in the autonomous issue triage + deciding whether an issue is actually solvable + multi-step validation + PR creation pipeline.

• 
  Luke Toledo (@lukeSVG) reported 10 minutes ago

  @NickADobos Had to go on a 2 day quest to un-spaghetti permissions recently, worth a look if you're still using defaults/clean permissions, found most of my issues was conflicting instructions all over the place agents.md, custom configs, cf workers, github all chiming in

• 
  Ayush ☄️ (@ayushkeshri4k) reported 14 minutes ago

  Yesterday I was working on my Project : Only a few application were open like - - Intellij Idea - GitHub Desktop - Notion - PostMan - Codex - Chrome - PostGre PG Admin and believe me I had very rough experience doing all , Intellij was not responding , laptop got slow , BTW, I have AMD R5 7series , SSD 512 GB and 16 GB DDR5. Still it sucks. Guys who do development on Windows, I wanna ask , what i m doing wrong . How to fix it , until I get a MaC

• 
  Mad Orkestra (@MadOrkestra) reported 15 minutes ago

  Github feature request: Auto-mark AI generated issues as slop. Thanks. @github

• 
  TechSnif (@techsnif) reported 18 minutes ago

  GitHub blames 7-hour August 17 outage on capacity failure

• 
  Alton Peques (@altonpeques) reported 20 minutes ago

  I’m constantly switching between projects in @cursor_ai that use different @github + @vercel accounts. I turned the whole login/verification process into a /checkout rule so Cursor knows which accounts to connect to for each project. Anyone have an even cleaner setup for this?

• 
  Gagan B Mishra (@__gbm) reported 23 minutes ago

  That retry storm taking down #GitHub is pretty amateurish.

• 
  Yume_X (@yume_arasaki) reported 24 minutes ago

  Showing the true power of the DGX Sparks : The AI Cookshow. Two Sparks. One night. A model fleet wrote a playable roguelike from a single prompt, and then one of the models sat down and played its own game. No API. Zero cloud tokens. The setup: - Spark 1 runs Qwen 3.8 27B dense (NVFP4, DSpark speculative). The brain. - Spark 2 runs Ornith-1.5 35B-A3B (NVFP4, MTP). The hands. - 128GB unified memory each, roughly 100W per box. The 35B fans out parallel drafts, one game module per stream, up to 24 streams at once. The 27B judges every draft on the other Spark, scores it, picks the winner. Best-of-N with an honest referee. New rounds, new modules, the game assembling piece by piece across the night. Both boxes fully loaded, both models earning their keep. An bitmap tile-based dungeon game was produced in two minutes, it's not visually impressive but it works. The throughput, measured on my rig: Ornith (drafter): - 1 stream: 88.7 tok/s - 8 streams: 305.7 tok/s aggregate - 24 streams: 496.6 tok/s aggregate. 13 percent over the published recipe number. Qwen 3.8 27B (judge): - 1 stream: 45.0 tok/s - 8 streams: 141.1 tok/s aggregate. DSpark overdelivers against its own estimate. Concurrent streams across the whole project: - 24 parallel drafting streams — the original fleet run (8 scopes × 3 rollouts each), 24/24 usable drafts in 101 seconds. That's the peak, on one Spark. - 18 multimodal drafts (6 scopes × 3) in 95s — the ASCII edition rerun, references attached - 18 drafts in 119s — the HTML edition - 16 simultaneous tonight — 8 drafters on Spark 2 + 8 judges on Spark 1, both boxes loaded at once - c24 sustained on the bench — that's where the 496.6 tok/s number comes from That is 637 aggregate tok/s of generation across two boxes pulling about 200W total. And the fleet delivered. Full drafting waves landed with every module usable. The judge caught every truncation, every hallucinated import, every missing function, before anything reached assembly. Zero false alarms in the logs. When it flagged a draft 4/10 for a logic error, the crash was real. The output: a tile roguelike. Single 19KB HTML file, zero dependencies, runs in any browser. You download it, you double-click, you are in a dungeon crawler. Not the final game I want yet, it is v1, but it boots, it plays, it fights back. Then my favorite part. We handed the 27B a real Chrome window, pointed it at its own game, and said play. It screenshotted, looked, chose a key, pressed it. 60 moves, every frame recorded. It explored, it found enemies, it fought. The model that wrote the engine also played it, with its own eyes. Stay tuned for episode 2. Drop in reply, what would you like to see tested out on two DGX sparks? Recipes and github gist in reply

• 
  Robin 🚀 (@JigsawClient) reported 24 minutes ago

  @ashwinrohitcom @uwukko Github started going down before ai