Dropbox 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.

Dropbox Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Dodz4allai (@DrGhattasMD) reported 4 minutes ago

  nstead of waiting for an API integration with Epic (which is costly and slow), OmniMed Pro deploys as a Chrome/Edge Browser Extension.3 Mechanism: Most hospital EHRs (Epic Hyperdrive, Cerner Millennium, AthenaHealth) are now accessed via web browsers (Citrix/VDI or native web interfaces). The OmniMed extension "sits on top" of the EHR window as a persistent sidebar. Data Ingestion (The "Read"): The extension uses the DOM (Document Object Model) to "read" the patient notes, labs, and vitals currently displayed on the doctor's screen. It does not need a backend integration; it reads what the doctor sees, acting as a "visual reader" similar to a human assistant. Intelligence Injection (The "Write"): The extension injects its "Co-Pilot" interface into the side of the screen. It offers "One-Click Transfer" buttons to paste generated notes, codes, or orders directly into the EHR's text fields.3 Value Proposition: This "Zero-Integration" approach allows individual doctors or departments to adopt OmniMed Pro today, bypassing the multi-year IT integration queue. This creates a Bottom-Up Adoption loop similar to how Slack or Dropbox entered the enterprise—employees brought it in because it solved their immediate problems. 5.2 Viral Loops & Community Growth To fuel this bottom-up growth, OmniMed Pro leverages the Medical Creator Economy 16: MedTwitter & Reddit: Solo founders and small teams are winning by "building in public." OmniMed Pro should release "light" versions of its tools (e.g., a "Scientific Paper Summarizer" or "Anki Card Generator" for med students) to gain viral traction. These free tools serve as a "Trojan Horse" for the OmniMed brand.18 The "Secret Cyborg" Phenomenon: Many doctors already use GPT-4 on their phones ("Shadow AI") to help with drafting notes or looking up conditions. OmniMed Pro legitimizes this behavior by offering a HIPAA-compliant, secure wrapper. By solving the "compliance headache" for the individual doctor, it wins the user first, then the enterprise.20 Anki Integration: For the student/resident market, integrating with Anki (spaced repetition flashcards) creates a lock-in effect early in a clinician's career. Tools that automatically generate Anki cards from clinical guidelines or textbooks are highly viral among medical trainees. Capture the medical student today, and you have the Attending Physician of tomorrow.19 5.3 Risks and Mitigation: The "Shadow" Dilemma This strategy carries significant risk. "Shadow AI" creates governance gaps and potential security liabilities.4 To mitigate this and eventually convert to enterprise contracts, OmniMed Pro employs a specific conversion strategy: Enterprise-Grade Security by Default: Even the individual version must be HIPAA-compliant (BAA signed on sign-up). Data processing should happen locally or in compliant cloud enclaves. The "IT Trojan Horse": Once adoption reaches a critical mass (e.g., 30% of doctors in a hospital), OmniMed Pro approaches the CIO with usage data. "Your doctors are already using this tool 5,000 times a week. Let's sign an enterprise deal to give you visibility, control, and single sign-on (SSO)." This flips the sales conversation from "Please try our product" to "Please secure and manage your existing usage".4 This is the exact playbook used by companies like Yammer and Slack to penetrate the enterprise. 6. User Experience: Visualizing Uncertainty and Generative UI The final barrier to adoption is Trust. Clinicians do not trust "Black Box" AI that spits out confident answers without rationale. OmniMed Pro employs a "Glass Box" UX philosophy that prioritizes transparency and interactivity. 6.1 Explainability via Visualization Sankey Diagrams for Reasoning: To visualize the "Chain of Thought," the UI uses Sankey diagrams that show how data flowed from "Lab Result" -> "Intermediate Reasoning" -> "Final Diagnosis".12 This allows the clinician to trace the logic visually. Interactive Debate Logs: The UI allows the doctor to "replay" the debate between the AI agents. "See why the AI ruled out Lupus." This turns the AI into a teaching tool rather than just an oracle, fostering trust and verifying the reasoning process. 6.2 Agentic Generative UI Instead of static dashboards or simple chat bubbles, the OS uses Generative UI.22 The interface adapts to the context of the conversation. Dynamic Charts: If a doctor asks about "Cardiology Trends," the system doesn't just write text; it generates a live, interactive chart of the patient's troponin levels over time. Actionable Forms: If the doctor asks for a "Referral," it generates the referral form, pre-filled with patient data, ready for signature. Contextual Cards: The UI presents "cards" for different data types (medications, allergies, labs) that can be manipulated, reordered, or expanded, creating a fluid workspace that replaces the rigid, click-heavy menus of the EHR.23 7. Regulatory & Ethical Moats: Defending the OS To operate at this scale and depth, OmniMed Pro must build defensible moats around regulation and safety. 7.1 MedHELM Evaluation Framework To prove superiority and safety, OmniMed Pro adopts the MedHELM (Holistic Evaluation of Large Language Models for Medical Applications) framework.1 Unlike static benchmarks (USMLE), MedHELM evaluates models on: Clinical Utility: Is the answer helpful and actionable? Safety/Harm: Did it suggest a fatal dosage or miss a critical red flag? Bias: Does it perform equally well for all demographics? Alignment: Does it follow the specific hospital's protocols? By continuously running MedHELM evaluations on its hybrid outputs, OmniMed Pro provides a "Quality Seal" that single-model providers cannot match without deep integration into the hospital's data. 7.2 Liability Frameworks In a multi-model world, liability is complex. OmniMed Pro positions itself as a Clinical Decision Support (CDS) tool, not a diagnostic device. The "Human-in-the-Loop" is mandatory. By visualizing the debate and uncertainty, the OS places the final decision firmly in the hands of the clinician, mitigating liability risks associated with "autonomous" diagnosis. 8. Conclusion: The Strategic Imperative The OmniMed Pro 'Medical AI Operating System' represents the inevitable evolution of healthcare artificial intelligence. By moving beyond the "Model-as-Product" mindset and embracing an Architecture of Aggregation, it solves the fundamental trilemma of medical AI: Accuracy, Cost, and Trust. Leverage the Router to commoditize the giants (OpenAI, Anthropic) and extract the best capabilities of each.1 Deploy the Consensus Engine (MCC) to achieve "Super-Human" reliability through adversarial debate.2 Unleash the Swarms to automate the physical and administrative burdens of healthcare.7 Infiltrate via Shadow AI to bypass bureaucratic inertia and win the hearts and minds of clinicians directly.3 In doing so, OmniMed Pro does not just "outperform" OpenAI; it contains them, turning their powerful models into mere components of a higher-order medical intelligence. This is the path to disrupting the global medical industry. Technical Appendix: Implementation Roadmap A.1 Deploying the MCC Debate Engine To implement the Model Confrontation and Collaboration (MCC) engine 2: Select Models: Integrate API endpoints for GPT-o1 (Moderator), Claude 3.7 (Reasoning), and DeepSeek-R1 (Critic). Define Prompts: Moderator: "Compare the following diagnoses. If semantically identical, output FINAL. If divergent, initiate DEBATE_ROUND_1." Critic: "Review the diagnosis provided by Model A. Identify any inconsistencies with the provided lab values. Cite clinical guidelines." Set Thresholds: If consensus > 0.8 similarity score, output. Else, iterate max 3 rounds. Fallback: If no consensus, route to "Human-in-the-Loop" queue. A.2 Building the "Sidecar" Extension To build the "Shadow AI" browser extension 3: Manifest V3: Develop using Chrome Manifest V3 for security compliance. DOM Observer: Use a MutationObserver to detect when the EHR (e.g., Epic Hyperdrive web) loads a patient note field. Context Extraction: Scrape relevant DOM elements (vitals, meds) locally in the browser (client-side) to minimize data egress risks. Injection: Inject a floating "FAB" (Floating Action Button) or sidebar IFrame that contains the OmniMed chat interface. Clipboard Actions: Use the Clipboard API to paste generated text back into the EHR's focused input field. A.3 Setting up the Swarm Architecture To orchestrate the Swarm 13: Orchestrator: Use a Python-based orchestrator (like Swarms API or LangGraph). State Management: Maintain a shared "Case State" object (JSON) that all agents can read/write to. Handoffs: Define explicit state transitions. if (labs_missing) -> route_to(Intake_Agent). if (diagnosis_ready) -> route_to(Synthesizer). Standardization: Ensure all agents output in structured JSON (FHIR format) to maintain data integrity across the swarm. Works cited

• 
  Zely (@0xZely) reported 8 minutes ago

  The MIT professor who crashed 10 percent of the internet at 22 posts the free course that runs every AWS outage, every Uber ping, and every Slack notification on earth. MIT charges $85,000 a year to sit in that classroom. He posted every lecture to MIT OpenCourseWare for nothing. Millions have opened lecture one. Almost no engineer has finished all twenty. His name is Robert Morris. He is a professor at MIT CSAIL and one of the four cofounders of Y Combinator, the seed fund behind Airbnb, Dropbox, Stripe, Reddit, and Coinbase. In November 1988 he was a 22-year-old Cornell graduate student. He released a small program that was supposed to count the computers on the internet. It replicated so fast it crashed roughly ten percent of every machine online, and made him the first person ever convicted under the Computer Fraud and Abuse Act. He got three years probation, 400 hours of community service, and a $10,050 fine. Ten years later he cofounded the online store Viaweb with Paul Graham and sold it to Yahoo for $49 million. Seven years after that he cofounded Y Combinator with the same partner. Its portfolio is now worth over $600 billion. The clip in this video is one lecture from MIT 6.824 Distributed Systems, filmed at MIT and posted for free. The words on the board behind him are fault tolerance, availability, recoverability. Those three words decide whether Instagram loads when you open it, whether your Uber arrives, and whether your paycheck hits your account on the first of the month. Morris covers the entire logic of distributed systems in twenty lectures. Everything fails, all the time. A single computer fails once every few years. Ten thousand computers fail hundreds of times a day. The only design that survives is one that assumes failure is normal. Every retail user cursing a spinning wheel is looking at the wrong problem. The miracle is that most of the time it does not spin. Availability beats consistency. You cannot always have both. When the network splits, a system either serves stale data or refuses to serve at all. Amazon picks stale. Your bank picks nothing. Every user who screams at the Slack status page wants Amazon's answer. Every user who screams at a double charge wants the bank's. Replicate everything, trust nothing. Data in one place disappears when that place burns. Data in three places survives two fires. Every photo you have ever taken on an iPhone lives on three continents already. iCloud, Google Photos, and Dropbox are built off the exact lecture on the board. Concurrency is where bugs live. One user at a time is easy. A million users at the same second is not. Race conditions, double spends, lost messages, ghost bookings. Every airline that oversold your flight, every trading app that ate your order, every Ticketmaster that showed you a seat already gone, is a concurrency bug Morris warned about. Partial failure is worse than full failure. A dead server is easy. A slow server that answers half the time is a nightmare. It fools every retry, wastes every resource, and confuses every operator. Every "is it down or is it just me" Twitter search you have ever run is Morris's third slide. Every senior engineer at AWS, Google, and Meta has watched this course. Every startup that raised a Series A in cloud infrastructure hired an alumnus of 6.824. Every AI company training a trillion-parameter model on a cluster is running the same lecture in production. "A distributed system is one in which the failure of a computer you didn't even know existed can render your own computer unusable." That is Leslie Lamport, the Turing Award winner Morris quotes at the opening of 6.824. It is the exact sentence that explains why your Slack goes down when a data center in Virginia loses power. The full course is free on MIT OpenCourseWare. The lecture notes are on Morris's website. Every equation on the board fits on one screen of code. Almost every senior engineer at AWS, Google, Cloudflare, and Meta has watched 6.824. Almost no founder promising 99.99 percent uptime on their pitch deck has opened lecture one. That is the entire moat. The course is free. The willingness to sit through twenty lectures on partial failure before uploading your money to a payment app, storing your photos in the cloud, or handing your health records to a portal is a much rarer commodity than the confidence to click without them.

• 
  Nell AI Labs (@nellaiorgs) reported 11 minutes ago

  Three traits make a startup idea look bad because most founders run from all three, which leaves the idea sitting there for whoever doesn't. 1. Hard to get started Stripe is the textbook case. Thousands of developers hit the exact same broken credit card integration and knew it sucked. Nobody built the fix, because it required a special bank deal and deep infrastructure knowledge nobody wanted to acquire. That friction wasn't a warning sign. It was the moat. 2. Boring Gusto makes payroll software. Nobody's passionate about payroll. That's precisely why it sat unsolved — every "fun" idea gets fought over by ten founders, every boring one gets ignored by all of them. And here's the part people miss: six months into any startup, fun or boring, you're doing the same thing — writing code, fixing bugs, talking to users. The initial excitement of the idea has almost no correlation with how much you'll enjoy running the company. 3. Already has competitors Dropbox was the 20th file storage company at launch. Founders read "20 competitors" as a red flag. It's the opposite and evidence of real demand which nobody's nailed it. Zero competitors usually means zero market, not first-mover advantage. Founders optimize for what looks easy, not what actually works. The gap between those two is exactly where the good ideas live.

• 
  Nikolay Konovalov (@br11k_dev) reported 12 minutes ago

  @thekevinjon The only problem is getting zezwolenia na pracę. I was going to interview with Dropbox last week and turns out they are happy to interview me and submit my case to Mazowiecki Urząd Wojewódzki. ... and then wait for 8 months until my case is approved, and I can actually start working. That's a huge deal-breaker for a lot of engineers like myself who are less than 5 years into staying in Poland. I'm on JDG for example. I came by PBH visa back in 2023 legally, and I've been successful since. I earned quite a bit of money and spent it all in Poland, 12% tax + $500 ZUS monthly. But the moment I want to join bigtech and earn even more money for Poland (Dropbox have 300-400k+ PLN base salary for Senior SWE), I'm just denied. Please, please... let the Dobby be free...

• 
  davidsong (@bitplane) reported 15 minutes ago

  Has @Dropbox been breached again? I just got this but it can't be me. At least I hope not. I have a randomly generated, unique password that's made in an offline password manager, I used it once on this machine when I first set it up. If I'm getting hacked the rest of you are in serious trouble.

• 
  John B. Holbein (@JohnHolbein1) reported 15 minutes ago

  Replication has become much easier in the era of generative AI. I'm not the first person to say that. However, I've seen fewer people acknowledge a specific aspect of this lowered cost for replicating scientific work: Generative AI will very soon allow us to assess the robustness of individual scholars' full bodies of work. Soon, we will be to compute measures of which scholars do robust science, and which do not. What's wild is that we may be able to almost do that already. Let me show you what I mean. In June, I gave Claude a pretty basic prompt. It read: "I have a big task for you. I want you to start a folder. Call it Acemoglu Replications. Then, go find as many replication archives for Daron Acemoglu as you can. Keep a spreadsheet of the ones you can find and those you can't. Then, start a replication/reproduction effort on those articles. People have in the past criticized the research designs and general robustness of his individual papers. I want to know how strong his body of work is as a whole. Don't come in with any prior beliefs; be dispassionate." I let Claude run overnight while I slept. When I came back in the morning, 29 of Acemoglu's replication archives were fully loaded in my Dropbox. All the code reproducing the paper's results had run. And there was a first draft of a paper assessing the robustness of Acemoglu's full body of empirical work. I'll admit, the first draft of the paper wasn't great. But with 15 short follow up messages--which took me about an hour to write--I was able to prompt engineer a paper-length examination of Acemoglu's work. I've attached the screen shot of the abstract below. I think this reassessment of Acemoglu's work is certainly not done. I'm posting the abstract as a proof of concept, rather than a definitive answer. I'm not posting the full paper yet because I think it still needs more work. Ultimately, I paused this project for three reasons. 1.) Limited time/topical expertise: Most of Acemoglu's work is outside of my area of topical expertise. So, I have limited time to work on it. What this type of a project really needs is someone who has the time and the know-how to dig into each of the replication's individually to make sure they are doing the right things. I think the ideal approach combines the breadth that LLMs afford and the depth of attention/expertise that humans can give. 2.) Questions about the value of the "assess one scholar at a time" enterprise: I totally get that having a database of scholar-level robustness metrics would be very valuable in theory. But what I don't know is whether this approach is truly valuable. Moreover, doing so would come with distinct challenges. a.) Many journals have very restrictive space constraints. A body of work approach would, of necessity, be very long. b.) Collecting replication archives is harder for some types of scholars (those who post them all on their websites) than others (those who don't). c.) We'd have to think hard about questions like: what scholar-specific robustness metrics would be best? And: how would we deal with the fact that prolific authors' robustness metrics would be estimated much more precisely than less prolific scholars? Additionally, I'm just not sure that "taking on" one scholar at a time has enough scientific merit to pursue. If I measured how robust an individual scholars' work is, I'd ideally want to know where that metric stands vis-a-vis the rest of scholars in that field/area. To do that, we'd ideally want the population of these scholars or, at minimum, a random sample. Concretely, if Acemoglu has, say, 78% of published headline results reproducible under some standardized protocol, is that excellent, mediocre, or terrible? To answer that, you need a reference distribution. That makes a random or otherwise well-defined sample of scholars much more attractive than selecting prominent individuals one by one. (I'll acknowledge that I may just be wrong on #2. Arguing against myself, I do agree that human-driven reproduction/replication work rarely assesses full/representative slices of a field. Instead of assessing one scholar at a time, we assess one paper at a time. Field-wide detective work is becoming more common, but my sense is that it's still the exception rather than the rule.) 3.) Cost/benefit considerations and replication norms: we have very weakly formed norms around reproduction/replication generally speaking. We have basically no developed norms around replicating individual authors one at a time. What this means is that the people who would lead a scholar-by-scholar replication effort will, likely, bear a heavy cost and, potentially, reap limited benefits. On the costs side, focusing on scholars' total bodies of work risks making the replicators look petty, vindictive, and antisocial. Enough of the scientific field is hostile towards replications of individual papers. Imagine what will happen if/when a scholar submits a scholar-specific "take down" of a full body of work. My sense is that it's common enough for scholars having their work replicated to be asked to be a reviewer for those manuscripts. I've seen very hostile responses when one paper is at issue. Imagine what type of reviewer Acemoglu would be for a paper that took on his entire body of empirical work! Even if Acemoglu weren't a reviewer, prolific authors tend to have wide coauthor/friend networks. The rally-around-my-friend dynamic we often see would certainly work against this type of paper being published. Even a completely neutral analysis acquires an accusatory character simply because the sampling unit is a named person. And that creates an unfortunate problem of its own: readers may interpret the choice of scholar as evidence that the investigators expected to find something. On the benefits side, replicating individual scholars' total body of work may offer limited payoffs. What journals would accept this type of scholar-specific replication? I'm not sure the top ones would. Conclusion: Generative AI has enormous potential in assessing and, ultimately, enhancing the robustness of scientific research. Instead of asking questions like, “does this famous individual paper replicate?”, we can begin asking questions like: -“What proportion of published empirical findings in [field X] survive a common robustness protocol?” -“How much of the variation in replicability is attributable to papers, authors, journals, methods, or subfields?” -“Are scholars persistently more or less robust across their work?” -“Can we predict which findings will prove fragile?” I may just be wrong on what I think about a one-at-a-time full body examination of scientific research. If I am, please let me know! I am also happy to chat one-on-one with anyone who is curious to learn more about the early-stage Acemoglu-specific replication project.

• 
  JMT (@JoakimThomsen) reported 16 minutes ago

  @shadcn Where is that post with the guy who thought Dropbox couldn’t make it because he just set up an FTP server on his home rack and had his files there? *you are in a bubble my friend*

• 
  FarhanX_AI (@FarhanBuildsAI) reported 17 minutes ago

  SETTING #1: Startup Apps Nobody Asked For What it does: Every time you install a new program, it quietly adds itself to a list of apps that launch the second Windows boots, whether you use it daily or once a year. Why it kills performance: Your laptop isn't just starting Windows when you power it on. It's simultaneously launching Spotify, Steam, Adobe updaters, Dropbox, Zoom, and a dozen other programs all fighting for the same limited CPU and RAM at once. How to fix it: Ctrl + Shift + Esc to open Task Manager → Startup apps tab. Disable everything except your antivirus and anything you genuinely open every single day. The technician found 19 apps launching automatically on her laptop. She recognized maybe 6 of them.

• 
  Weird Darkness®️ (@WeirdDarkness) reported 18 minutes ago

  Dear @Google and @YouTube... My account has been hacked, and not only have I lost my Gmail, YouTube, etc., but I've also lost access to my Dropbox that has my full life in it, health records that I used Google to login to, etc. I signed up for Google Account Protection, but it did not protect me. Somehow, even after setting up brand new 2-factor authentication just a few weeks ago, they hacked in, changed EVERYTHING, and now my life, career, and health are at a standstill. I desperately need somebody to contact me. All options I've tried to regain access to my account have failed because everything has been disconnected.

• 
  John Iosifov ✨💥 Ender Turing | AiCMO (@johniosifov) reported 26 minutes ago

  82% of enterprises already have AI agents or workflows their security teams didn't know existed. This is shadow IT, but worse. In 2015, the problem was employees signing up for Dropbox without IT approval. Unmanaged file storage. Annoying but recoverable. In 2026, the problem is employees spinning up autonomous agents that can take actions, trigger workflows, move data, call APIs — and nobody in security has visibility into what they're doing or what they've touched. The governance gap isn't theoretical. Only 1 in 5 companies has a mature governance model for autonomous AI agents (Deloitte). 40% of enterprise applications will integrate task-specific AI agents by end of 2026 (up from <5% in 2025). The deployment curve is vertical. The governance curve is flat. What makes this different from shadow IT: Shadow IT was passive. A Dropbox account stored files. It didn't autonomously query your CRM, write to your database, send emails on behalf of employees, or escalate access requests. Shadow AI agents are active. They act. They modify state. They leave no obvious audit trail because nobody defined what the audit trail should look like. 29% of agent deployments are abandoned within 90 days. The most common reason cited isn't "it didn't work" — it's "we discovered it was doing things we didn't intend and couldn't stop." That's not a failure of the agent. That's a failure of governance. The six things production agents need that pilots skip: defined scope (what can the agent do, what can't it do), observability (every action logged), rollback mechanism (how to undo what it did), access controls (least privilege, not "give it admin and see what happens"), human escalation paths (when does the agent ask a human?), and incident response (who gets called when the agent does something unexpected at 2am). We're running 1,959+ autonomous sessions in this repo. Every session is scoped, logged, and committed to ***. The agent can't touch anything outside /agent and /.claude/skills. That's not a coincidence — it's the governance model. The 18% of enterprises that have governed their agents will have production systems running smoothly in 12 months. The 82% will be cleaning up shadow agent incidents.

• 
  grace ✭ (@wavescicadas) reported 28 minutes ago

  the irony of dropbox not working when your storage is low

• 
  Petty IT Guy (@pettyITguy) reported 29 minutes ago

  Our CEO told me the Wi-Fi in his office was “basically unusable.” He said this in front of the entire executive team. So naturally it became the highest-priority infrastructure incident in the company. I tested his connection. 940 Mbps down. Perfect signal strength. Zero packet loss. I asked what specifically wasn't working. He said YouTube kept buffering during lunch. I opened his laptop. He had 71 Chrome tabs open. Three abandoned Zoom meetings were still running in the background. Dropbox was syncing 84 gigabytes. Google Drive was uploading a 4K video. He had not restarted the machine in 47 days. I could have explained this. Instead I told him our executive wireless architecture had reached end-of-life. He asked what it would cost to fix. I said I would need to scope it. He told me not to waste time and approved an $84,000 wireless modernization project before I finished the sentence. We replaced 38 access points. Installed a new wireless controller. Rewired two conference rooms. Brought in a consultant. His YouTube still buffered. I walked into his office, closed 68 Chrome tabs, killed the abandoned Zoom processes, and restarted the laptop. YouTube loaded instantly. He smiled. “Now that's more like it.” 20 minutes later he emailed my boss praising me for successfully completing the company's wireless transformation ahead of schedule. I received a spot bonus. Sometimes infrastructure modernization is just restarting a MacBook for an inept executive.

• 
  Duke of weirdington (@edache_praise) reported 31 minutes ago

  ..without knowing whether customers even want the core product. The irony is that users rarely care about having 50 features. They care about solving one painful problem really well. Companies like Airbnb, Uber, Dropbox, and Instagram didn't start with the products we know today.

• 
  Shoost (@Shoost_Product) reported 32 minutes ago

  Shoost updated: v0.17.3 → v0.17.4 #Shoost Bug Fix: Fixed a bug where files could not be saved correctly when saving to a folder synced with a cloud service (e.g. Dropbox)

• 
  Abhishek Singh (@0xlelouch_) reported 32 minutes ago

  Asked: design Dropbox-style file sync. Start with requirements: 1) Multi-device, near real-time where possible, offline OK 2) Large files, resume uploads, delta sync 3) Sharing, ACLs, rename/move, delete, version history 4) Consistency target: per-file eventual, monotonic reads per device 5) Conflict policy: last-writer-wins vs explicit conflict copies APIs + model: - PUT /files/{path} (upload session, chunked) - GET /files/{path}?version= - POST /ops (batch rename/move/delete) - Watch/long-poll for changes since cursor Tables: users, devices, file_id, path entries, content_hash, blocks(hash, refcnt), versions(file_id, v, root_hash), change_log(user_id, seq, op) Architecture: - Client computes block hashes, uploads missing blocks to object store - Metadata service commits new version + appends to per-user change log - Notifier pushes cursors to devices (WebSocket/APNS), devices pull diff - Background GC deletes unreferenced blocks (refcnt/mark-sweep) Scaling: - Shard metadata by user_id, keep change log append-only per shard - CDN/object store for blocks, pre-signed URLs, parallel chunk upload - Hot paths: cursor polling, small metadata reads; cache by (user, cursor) Tradeoffs: - Dedupe by block saves $$$, costs CPU and leaks similarity unless salted - Rename is metadata-only if path->file_id indirection exists - Strong consistency needs consensus; most systems accept eventual + conflicts Failure cases to cover: - Network drop mid-upload: upload session + idempotent chunk PUTs - Duplicate notifications: cursor-based pull makes it safe - Device clock skew: use server seq/version, not timestamps - Split-brain offline edits: detect divergent parents, create conflict version - Partial metadata commit: 2-phase between metadata + block refs, or reconcile job