Software · Outage

Ten incidents in five days: GitHub's week when the merge button kept timing out

GitHub's own status record shows ten incidents between Monday, November 9 and Friday, November 13, 2026 — a database migration that overwhelmed a new cloud region, an overload in the shared Git infrastructure, failing webhooks, an email outage and pull-request merges timing out. No day was quiet.

THE BUYER'S PULSESoftware8 min read
Two developers review code together at a monitor — the pull-request ritual that GitHub's Friday incident kept interrupting
Code review is a conversation held over a diff. In the week of November 9, GitHub's own record shows that conversation stalling again and again: delayed workflows, timing-out merges, silent webhooks. Photo: Mikhail Nilov / Pexels.

Photo: Mikhail Nilov / Pexels (Pexels License).

There is a document that tells you more about a platform than any keynote: its status page, read whole. GitHub's status record for the week of November 9 reads like a diary kept by an exhausted engineer. Ten incidents in five days — Monday through Friday, no day quiet — spanning the job queues that run the world's CI pipelines, the webhooks that stitch its tools together, the email that guards its logins, and, to close the week, the merge button itself. No single incident was a catastrophe. Together they describe something more interesting: a week in which the friction sat exactly where modern software work happens, in the shared machinery underneath the code.

10 incidentsPosted to GitHub's official status record between Monday, November 9 and Friday, November 13 — every weekday had at least one
65% peakShare of Actions workflow runs affected at the worst moment of Monday's database migration; the mean was 28%
0.09%Of pull-request events on Tuesday never triggered their workflows at all and had to be re-run by hand
3% of requestsHit by Friday's pull-request merge timeouts before mitigation brought the impact under 1%

GitHub is not one website among many. It is where a large share of the world's software is stored, reviewed and shipped, and its Actions service is the assembly line that builds and tests much of it. When its queues stumble, releases stall in companies that have never heard of each other. That is the stage on which this week's ten entries played out — each one time-stamped, quantified and, unusually often, explained by GitHub itself.

A note on how to read the record. The times are UTC, the percentages are GitHub's own, and the operative word is almost never “down.” It is “degraded” — slower, later, intermittent. That distinction is the week's real subject. A platform can be up all week and still tax every team that depends on it, in minutes stolen from builds, deliveries that arrive after the moment they were for, and checks that silently never ran. The ten entries below are that tax, itemised by the tax collector itself.

Monday: the migration that outgrew its new home

The week opened at 09:23 UTC with a degraded cluster in the job-queue infrastructure, delaying webhook deliveries and Actions runs until 11:40. It was the evening entry that set the tone. At 18:07 UTC, during a planned migration of a central Actions database from a data center to a new cloud region, GitHub began moving traffic — and the volume that arrived exceeded both the capacity of the new region and the configured limits. An average of 28% of workflow runs were affected, peaking at 65%. Webhooks, Pages and Copilot code review felt the knock-on effects. The fix was a partial retreat: some traffic went back to the original region while capacity and limits were raised, and by 18:59 the incident closed, with GitHub publishing its root-cause account inside the resolution note itself. Late that night came a third, smaller entry — Copilot models degraded between 23:16 and 00:05 by an upstream provider — the first appearance of a character who returns all week: somebody else's infrastructure.

Tuesday: the big one, in the shared layer

Tuesday's incident, running roughly from 15:00 to 17:55 UTC, was the week's most consequential and the one GitHub explained in most depth. The cause was overload in the shared infrastructure that performs Git operations — the substrate beneath repository pages, pull requests and Actions alike. Recent changes had quietly increased the background processing load and the cost of computing pull-request merges, and under the day's traffic the shared layer saturated. Actions workflows were delayed in about 1.8% of events, peaking near 7% between 16:30 and 16:50; typical delays ran 1.5 to 3.5 minutes, stretching to 10. The detail that should worry every engineering manager sits in GitHub's own numbers: roughly 0.09% of pull-request events never triggered their workflows at all, and had to be re-triggered manually — failures that do not announce themselves, discovered only when someone asks why a check never ran.

The mitigations were the classic triage of an overloaded shared system: switch off a recent merge optimisation, pause background repository analysis, and throttle low-priority traffic — including commit pages for logged-out visitors. GitHub's declared corrective work is equally telling: reduce repeated repository analysis, add load-based concurrency controls, and make pull-request synchronisation more resilient. Translated from incident-speak, the platform's own homework is to make the shared layer degrade gracefully instead of all at once.

Notice, too, what kind of failures these were. Almost nothing in the week's record is a hard outage — a service simply gone. The vocabulary is degradation: delayed, throttled, timing out, running behind. Degradation is the harder failure mode for the customer, because it does not trip the alarms a clean outage trips. Pipelines do not fail; they finish twenty minutes late — or they never start, and nobody notices until the deployment that depended on them is missing. Tuesday's 0.09% is the purest form of the species: a failure whose only symptom is an absence. A platform of GitHub's size cannot honestly promise a week without degradation. What it can do — and what this record shows it doing — is measure the degradation in public, in percentages, while it is still happening.

The supporting cast: webhooks, email and the model supply chain

The rest of the week was a tour of the dependencies. Wednesday brought two short, sharp entries: webhooks degraded for four minutes at 20:22 UTC, then a multi-product incident from 20:48 to 21:21 that degraded Actions, Pages, Webhooks and Copilot while the Actions job backlog drained. Thursday belonged to the connective tissue: webhooks again, from 18:43 to 19:33, with timeouts and errors in about 1% of traffic — a small percentage of a very large number, felt by every integration waiting on a delivery that never came. Then, at 19:36, email: an upstream provider's disruption delayed sign-ins, address verifications, password resets and notifications until 21:43 — a reminder that even the front door of the platform hangs on a vendor's hinges.

And twice in the week, the failing layer was the newest one. Copilot's models degraded on Monday night and again on Friday evening, both times attributed to an upstream provider, with model names that read like a catalogue of the AI supply chain — GPT-5.3 Codex, GPT-5.6 Luna and Sol, GPT-6.1 Terra, GPT-6 Luna. On Friday, GitHub notes, its automated detection raised warnings inside the product itself. The platform that hosts the world's code now depends, for its flagship assistant, on models served from outside its own walls; the status page has started recording that dependency in the same flat tone it uses for its own databases.

Friday: the merge button times out

The week's closing incident was the most symbolic. From 15:47 to 18:29 UTC on Friday, pull requests degraded: merges began timing out. The initial impact touched 3% of requests; a mitigation at 16:12 UTC brought it under 1%, and the backlog worked itself out over the following two hours. Merging is the ceremonial centre of the pull-request workflow — the moment reviewed code becomes the product — and it was the week's last thing to break. GitHub closed the incident with a promise: a detailed root-cause analysis “will be shared as soon as it is available.” At the week's end, it was still owed. Saturday and Sunday, the record shows nothing at all. After ten entries, the silence reads less like calm than like everyone involved finally getting some sleep.

There is a quieter story in the document itself. The status-page genre is not famous for candour — its clichés are “some users may have experienced issues” and resolution notes that resolve nothing. This record is different in kind: impact stated in percentages, mitigations named while they are being applied, an upstream provider blamed plainly when the fault is upstream, a root cause published inside a Monday-evening resolution note, and a second one promised by name for Friday. Enterprise customers do not read these pages for entertainment; they read them the way auditors read filings, and they price what they find into renewals. A bad week, documented this well, can end up arguing for the platform almost as effectively as a quiet one.

Read whole, the week is not an indictment; it is an X-ray. A platform fails where it is shared — queues, Git operations, merge computation, delivery plumbing, and now model supply — and each of this week's incidents landed on one of those commons. GitHub's record is unusually candid about percentages, mitigations and unfinished homework, and that candour is itself part of the product enterprise customers are buying. The tests that remain are the two promises the record leaves open: Tuesday's corrective work, and Friday's root-cause analysis. The status page will tell us — it always does, in public, whether the week is quiet or not.

𝐆𝐄𝐓 𝐅𝐑𝐄𝐄 𝐅𝐔𝐋𝐋 𝐌𝐀𝐆𝐀𝐙𝐈𝐍𝐄 𝐀𝐓 thebuyerspulse.com

Sources

  • GitHub Status (Nov 9–13, 2026), the official incident record (githubstatus.com): all ten incidents and their timelines, Monday's Actions database migration (28% of workflow runs affected on average, 65% at peak), Tuesday's shared Git-operations overload (about 1.8% of Actions events delayed, roughly 0.09% of pull-request events never triggered), Wednesday's and Thursday's webhook degradations, Thursday's upstream email disruption, Friday's pull-request merge timeouts (3% of requests, mitigated below 1%) and the two Copilot model degradations attributed to upstream providers.
  • UptimeRobot, Pingoru and PulsAPI (Nov 2026), independent tracking of the same record: Friday's pull-request incident ran 2 hours 42 minutes across eight updates; trackers counted fifteen GitHub incidents between November 4 and 14.
  • Photo: Mikhail Nilov / Pexels — two professionals reviewing code together at a monitor (Pexels License).