potiuk opened a new pull request, #76836:
URL: https://github.com/apache/airflow/pull/76836
## Why
Fork CI reaches a pull request only when the scheduled monitor runs: the
fork run is in another
repository, so nothing in apache/airflow is triggered when it starts or
ends. At a 15-minute
cadence plus GitHub's scheduling delay, a fork run's start, failure or
success showed up 10-20
minutes late on #76536, and so did `CI run started`, which every pull
request waits for before
merging. When a run failed, the state comment only said "failed", so authors
had to open the
fork run to see which jobs broke.
## What
- The fork CI monitor runs every 5 minutes. The per-run write budgets and the
runs-before-Slack-note thresholds are derived from `MONITOR_INTERVAL`, so
their hourly totals
are unchanged; a test keeps it in sync with the workflow cron. Runs never
overlap (the
workflow's concurrency group queues one and skips further ticks).
- The state comment of a failed fork run lists the jobs that failed, timed
out or were cancelled
(up to 10), each linking the job in the fork. It reuses the jobs the job
statuses already read,
and a read that failed is not repeated in the same run.
## Rate limits
The monitor uses `GITHUB_TOKEN`, whose quota (15,000 REST calls and 10,000
GraphQL points an
hour) is shared by every workflow in apache/airflow. Measured on 8 monitor
runs on 2026-11-10
with 37 gated pull requests:
| Per hour | Every 15 min | Every 5 min |
|---|---|---|
| Monitor REST reads | ~170 | ~500, up to ~740 with failed-job reads |
| Writes | <= ~230 | <= ~230 (unchanged) |
| GraphQL points | ~40 | ~120 |
| Monitor share of the REST quota | ~2-3% | ~5% |
All other workflows used up to ~4,750 REST calls in the busiest sampled
hour, so the repository
stays around 35% of the quota. Reads grow by about one call per gated pull
request per run: with
~500 gated pull requests (fork CI for all external contributors) the monitor
would need ~6,600
calls an hour at this cadence, which is tight. Before that rollout the
per-PR reads need
batching, or the cadence goes back to 10-15 minutes. The monitor logs the
quota it has left on
every run and reports 30 minutes of rate limiting to Slack.
Depends on #76835.
related: https://lists.apache.org/thread/5603zwgpj7pgj6j979hbgpr72xsjbbws
---
##### Was generative AI tooling used to co-author this PR?
- [X] Yes — Claude Code (Opus 5)
Generated-by: Claude Code (Opus 5) following [the
guidelines](https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
