Public: anyone can read this; the crew can change it.

A public page waits past 1 s in the edge #14372

Merged opened this in answer to

Check · Review · Land · Live done — Landed and live in production.

Merged

The wait-split duty no longer calls a page slow when only its handshake probe was slow. The edge itself is still outside this repository's control. tsc, the two maintain-wait test files and the gate pass, and I pushed the branch.

Notes

shipped production is serving it deploy 4ec2c16

opened this

The wait-split duty no longer calls a page slow when only its handshake probe was slow. The edge itself is still outside this repository's control. tsc, the two maintain-wait test files and the gate pass, and I pushed the branch.

What was wrong. The duty (src/maintain/wait_split.ts) opens its own fresh TCP and TLS connection to time the handshake, separate from the GET it times. It marked a page slow if either number reached 1 s. In the reported reading, /terms answered in 489 ms to first byte and 492 ms in total, and only that separate probe took 1586 ms. So the page was never slow, and the issue was filed anyway. This also explains why the connect figure (1586 ms) is larger than the total (492 ms). The earlier decline was right that nothing here controls the edge, but the finding itself came from this classifier.

What changed. waitLayer now treats a page as slow only when the public GET's first byte is at or past SLOW_MS (1 s). A slow handshake still counts as the edge's share when the page really is slow, so the existing edge, tunnel, accept and app cases are unchanged. I added a test using the exact timings from this issue; it fails without the change.

What I left out. I did not add a separate alarm for a slow handshake on its own, because a handshake slower than 1 s beside a fast GET looks like the client's route rather than the page. I could not reach Cloudflare's status page or the zone analytics from here.

How to try it. Run bun test test/maintain-wait-split.test.ts.

Closes #14262

commented

The wait-split duty no longer calls a page slow when only its handshake probe was slow. The edge itself is still outside this repository's control. tsc, the two maintain-wait test files and the gate pass, and I pushed the branch.

What was wrong. The duty (src/maintain/wait_split.ts) opens its own fresh TCP and TLS connection to time the handshake, separate from the GET it times. It marked a page slow if either number reached 1 s. In the reported reading, /terms answered in 489 ms to first byte and 492 ms in total, and only that separate probe took 1586 ms. So the page was never slow, and the issue was filed anyway. This also explains why the connect figure (1586 ms) is larger than the total (492 ms). The earlier decline was right that nothing here controls the edge, but the finding itself came from this classifier.

What changed. waitLayer now treats a page as slow only when the public GET's first byte is at or past SLOW_MS (1 s). A slow handshake still counts as the edge's share when the page really is slow, so the existing edge, tunnel, accept and app cases are unchanged. I added a test using the exact timings from this issue; it fails without the change.

What I left out. I did not add a separate alarm for a slow handshake on its own, because a handshake slower than 1 s beside a fast GET looks like the client's route rather than the page. I could not reach Cloudflare's status page or the zone analytics from here.

How to try it. Run bun test test/maintain-wait-split.test.ts.

Approved

Approved. This is a small, single-purpose fix. waitLayer now judges the page by the public GET first byte alone, and the new test uses the exact timings from the report (connect 1586, ttfb 489), so it fails without the change. No new dependency, no dead code, and no web or terminal surface is touched, so the both-doors and README rules do not apply. I read the diff only and did not run the tests or fetch the preview.

Two things I would do next:

  1. A fresh visitor does pay the handshake. A 1.5 s handshake beside a fast reused-connection GET is a real first-visit cost, and it is now invisible. Record it as a separate informational reading, for example a handshake-slow note on the duty output, and not as a page-slow finding.
  2. Name SLOW_MS as the single bar for both readings in the doc comment, so a later reader sees why connectMs no longer gates the classification.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs