Plan: the nixpkgs event log

Status: built in src/events.rs, except two parts: the path prefix filter, and the queues that read new events. It replaces the separate event pipeline (a message broker and five processes) with one log inside Maintainer Mode. The old pipeline stays on the host until this log has run for some time.

What it does

• Collect every public event on NixOS/nixpkgs.
• Keep them in order, with a numeric id.
• Serve them live and from any past point.
• Keep a copy outside the host, which is also a public dataset.

GitHub sends webhooks only for repositories you administer, so there is no webhook receiver. The sources are the Events API and GH Archive.

Sources

• Events API. Poll GET /repos/NixOS/nixpkgs/events with ETag, at the interval GitHub asks for in X-Poll-Interval. A 304 costs no rate limit. The API shows only the last 300 events, so a long gap loses events.
• GH Archive. Each hour, read the file for the past hour and add the events that the poll missed. This closes the gaps from restarts and bursts.

The GitHub event id is the key, so an event that both sources deliver is stored once.

Storage: one SQLite file per month

events/2026-10.sqlite, events/2026-11.sqlite, and so on. The current month takes writes. A past month does not change.

A file per month keeps each file small, makes retention a file delete, and makes the backup a file copy. A query over many months attaches the files it needs.

The seq continues across files: a new month starts at the last seq plus one.

Reading

• GET /api/v1/events is Server-Sent Events. With Last-Event-ID, it replays from that seq, then follows live. Filters: type, number, actor, and a path prefix for pull requests that touch it.
• The queues read new events at once, so a review shows without waiting for the next sync.

Copy outside the host

Any S3-compatible store, at a provider other than the host's. The default is bunny.net Storage, in an EU region with the S3 API, with Bunny CDN in front for the public downloads.

• Each day: the events of that day as YYYY/MM/DD.ndjson.zst, in the GH Archive format. This is the public dataset and the restore path.
• Each past month: its SQLite file, compressed, once.

A restore reads the month files, then the daily files of the current month.

The store holds only public events. It needs no access control for reads, and it can be served in public through the CDN.

Size

Measured on 16 Oct 2026.

GH Archive, one full day (14 Oct 2026, 24 files, 318 MB gzip):

Events API, one page (100 events, read at 12:22 UTC):

• 74 of the 100 events have a time between 12:01 and 12:21 UTC. That is about 220 events per hour, the same as the host figure of 200 to 350.
• 343,873 bytes of JSON (3.4 kB per event); zstd -19 makes 24,465 bytes.
• The response has ETag and X-Poll-Interval: 60.
• The page also holds older events, back to 14 Oct 22:10. Of the 3 events from 14 Oct after 22:10, none is in GH Archive, and the 10 GH Archive events of that time are not on the page.

What this means:

• GH Archive holds about 2 to 3 % of the nixpkgs events. Each hourly file has about 56,000 events for all of GitHub. The gap fill adds the few events that it has, but it cannot close a long gap. The poll is the real source, and a poll that stops loses events.
• About 220 events per hour at 3 kB is about 16 MB of JSON per day. With zstd at 11 : 1, a daily file is about 1.5 MB and a month is about 45 MB. Each row is compressed alone, so a month file on the host is larger.
• Both sources now send a short pull_request object (base, head, id, number, url), not the full pull request.

Order

1. One day of GH Archive for NixOS/nixpkgs: count and bytes.
2. The month files and the poll.
3. The GH Archive gap fill.
4. GET /api/v1/events.
5. The daily copy to the store.
6. Then remove the old pipeline from the host.

Open

• How many months stay on the host.
• Whether the bucket is public, or the dataset gets its own URL.