A data essay on shared infrastructure
CRITICAL
MASS
GitHub keeps the world's code. Ten years of its official status record show a platform that was already stumbling — before AI agents arrived and rewrote the load curve.
Every figure on this page traces to GitHub's own status history, engineering blog, or cited reporting. Sources & methodology.
01 — The record
Ten years of green.
Then the red crept in.
Every day since April 2016, colored by GitHub's own status page — worst reading across ten core services.
- Operational
- Minor
- Degraded
- Major outage
- Severity is encoded by bar height as well as color; major-outage cells are ringed. Arrow keys move between days.
Data table: disrupted days and incidents per year
Named incidents opened per month. Worst month on record: February 2026 — 26 in 28 days.
Data table: incidents per month
2018: 91.8% of days clean. 2023: 67.7%. First 65 days of 2026: 53.8%.
Data table: clean-day share per year
incidents in 2023 alone — five times the 2018 count. 2024 looked like relief. It wasn't.
02 — Anatomy of failure
The scars have names.
Eight incidents that defined the record — cause, duration, blast radius. Read the pattern, not the exceptions.
GitHub's own published figures, y-axis zoomed to 97–100%. The 99.9% enterprise bar is the line Actions kept missing. March 2026: 97.55% — below two nines.
Data table: Actions official monthly uptime
Copilot — the AI workload itself — went from 1.6% of days disrupted in 2022 to 29.2% in early 2026. Actions, the machine's workshop: 20%.
Data table: disrupted-day share per service
03 — The machine load
The new users don't sleep.
AI coding agents now open pull requests, trigger CI, and hammer APIs around the clock. The load curve went vertical in December 2025.
The machine's workshop quadrupled in under three years: 500M → 1B → 2.1B minutes a week.
Data table: Actions weekly minutes
Human traffic follows sleep and weekends. Agent traffic doesn't. Capacity planned around human rhythm has no slack left.
Data table: working hours per week, human vs agent
year-over-year change in human comments on commits in 2025 — while every machine-paced activity record broke. The conversation is leaving the commit log.
GitHub Octoverse, Oct 202504 — Collision
Two curves, one platform.
GitHub's CTO dates the agentic acceleration to late December 2025. The incident record bends at exactly that hinge.
Red bars: incidents per month (status history). Blue line: AI-agent PRs per month (Sep 2025 → Mar 2026). The vertical marker: December 2025, when GitHub says agentic workflows "accelerated sharply."
Data table: incidents vs AI PRs
GitHub's enterprise bar: 99.9%. Official Actions uptime missed it in February (98.70%) and collapsed through it in March (97.55%). The CTO conceded the miss publicly.
Data table: nines target vs actuals
October 2025 plan: 10× capacity. February 2026 conclusion: 30× — a redesign, not an expansion.
Data table: capacity targets
July 2026, officially
- 8 incidents in one month, per GitHub's own availability report.
- 96% peak 5xx error rate in the worst-hit environment on July 8 (7h 04m).
- 113,930 pull-request creation attempts failed on July 24 — 50,904 users, 57 minutes.
- 60% of Actions runs failing with infrastructure errors at peak on July 25.
- August 6: an Actions outage GitHub itself called "unacceptable in both its impact and particularity of its duration."
The platform was sized for people.
The people brought machines.
GitHub is answering: half of monolith read traffic now runs in Azure, authentication is leaving its oldest shared database, and the plan is to have production out of its own datacenters by the end of 2026. The order of operations is explicit — availability, then capacity, then features.
Whether 30× is a ceiling or a floor depends on a curve no one at GitHub controls. The machines don't sleep, don't hesitate, and don't file support tickets. They just retry.
Verification
Sources & methodology
Primary data
- GitHub's Historic Uptime — day-level scrape of the official githubstatus.com uptime-history API (snapshot Mar 3, 2026); underlying dataset. Basis of the matrix, incident counts, clean-day shares, and official monthly uptime figures.
- October 21 post-incident analysis — GitHub Blog, Oct 30, 2018.
- Addressing GitHub's recent availability issues — GitHub Blog, May 16, 2023.
- An update on GitHub availability — CTO Vlad Fedorov, GitHub Blog, Apr 28, 2026 (10×→30×, Dec 2025 agentic acceleration, Apr 23 merge-queue incident).
- GitHub availability report: July 2026 — GitHub Blog, Aug 12, 2026 (8 incidents, 96% 5xx peak, 113,930 failed PR creations, Aug 6 "unacceptable").
- Octoverse 2025 — GitHub, Oct 28, 2025 (986M commits, 43.2M PRs merged/mo, 80% first-week Copilot, 693,867 LLM-SDK repos, −27% commit comments).
Reporting & analysis
- GitHub, Disrupted Completely by AI — GeekPark via 36Kr, Jun 4, 2026 (275M commits/week, Actions 500M→1B→2.1B min/week, Claude Code 4.5% of public commits & 100K→2.6M/week, AI PRs 4M→17M/mo, Feb 9 cache-TTL postmortem, three-nines miss).
- Industry analysis citing SemiAnalysis — Claude Code ≈4% of public GitHub commits, projected 20%+ by end of 2026; see also Signal, Jul 2026.
- Trigger, Not Root Cause: Re-Examining GitHub's 2024–2026 Availability Decline Under AI-Driven Load — Future Internet 18(7):377, MDPI, Jul 20, 2026.
Methodology & caveats
- Matrix cell colors reproduce the official status colors githubstatus.com rendered per day (worst of ten tracked services: API Requests, Actions, Codespaces, Copilot, Git Operations, Issues, Packages, Pages, Pull Requests, Webhooks).
- "Disrupted day" = any service below fully-operational that day. Incidents are counted once per named status event, in the month they began.
- Official monthly uptime percentages are GitHub's own published figures, used verbatim. The status API's opaque per-day counters do not map linearly to those figures and were deliberately not used for downtime math.
- 2019–mid-2020 incident metadata is sparse because GitHub migrated status systems; the public record has a hole there. Pre-2019 entries come from GitHub's legacy status log (April 2016 onward) and cover publicly statused incidents only.
- Datasets end March 3–6, 2026 (snapshot). The 2026 figures cover January 1 – March 6 only (65 days, with March partial at 6 days) and are labeled as such wherever they appear; they are not comparable to full-year totals. April–August 2026 figures come from GitHub's own availability reporting, cited above.
- AI-load figures are as reported by the cited sources; where a number is a projection (14B commits annualized, 20% commit share), it is labeled as one.
An independent data essay. Not affiliated with, or endorsed by, GitHub. Built August 2026.