Cloudflare edge: the stream arrives at once, not in pieces

What changed

There is no single announcement here. Four separate edge behaviours interact, and each one is documented as normal, correct behaviour on its own. The failure appears in the combination: a streaming endpoint, a dynamic JSON API and a front-end bundle that all sit behind the same zone.

• Cache admission is by file extension, not by MIME type. Cloudflare does not cache HTML or JSON by default, but it does cache JS and CSS. An unhashed /assets/app.js can therefore be served from cache after a deploy that rewrote it. Anything else is cached only because a Cache Rule said so.
• “Eligible for cache” is not the same as “cached”. A matched rule still yields BYPASS when the origin answer is not cacheable, and HIT when it is. The rule’s Edge TTL mode decides whether origin Cache-Control is respected or ignored, so one dropdown can turn an authenticated JSON route into a shared cache entry.
• Compression is content-type driven and can be forced. Cloudflare’s default compressed-type list does not include text/event-stream, but it asks your origin for br or gzip on every request, and Compression Rules can target the stream path anyway. Cache-Control: no-transform is the documented way to stop Cloudflare from altering compression.
• Rocket Loader is a zone-level rewrite of JavaScript loading. It defers all scripts, inline ones included, to speed first paint. Its own documentation says to disable it if you see JavaScript issues and to opt individual scripts out with data-cfasync="false". A CSP-enabled app needs header changes before it can even run.
• Auto Minify is deprecated, not gone. The feature was deprecated on 2024-08-05; the API endpoints are deprecated, and the documented remedy for a zone that still has it on is to switch css, html and js off through the zone settings API.
• Tunnel-side buffering has its own switch. Cloudflare Tunnel buffers proxied responses unless the origin sends Content-Type: text/event-stream; the header is what tells cloudflared to stream instead of buffer.

Detect: are you affected?

All three checks are read-only. Check 1 is about the edge decision, check 2 is about time-on-the-wire, check 3 reads your zone configuration.

Check 1 — what the edge decided for the endpoint

curl -sI "https://app.internal/api/chat" \
  -H "Accept: text/event-stream" \
  | grep -iE "^(HTTP/|cf-cache-status|content-type|content-encoding|cache-control|age|vary):"

Check 2 — first byte versus last byte

curl -sN "https://app.internal/api/chat" \
  -X POST \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"prompt":"count to twenty, one word per line"}' \
  -o /dev/null \
  -w 'first_byte=%{time_starttransfer}s total=%{time_total}s\n'

Check 3 — zone settings that rewrite or cache

curl -s "https://api.cloudflare.com/client/v4/zones/$ZONE_ID/settings/rocket_loader" \
  -H "Authorization: Bearer $CF_API_TOKEN" | jq -r '.result.value'
curl -s "https://api.cloudflare.com/client/v4/zones/$ZONE_ID/settings/minify" \
  -H "Authorization: Bearer $CF_API_TOKEN" | jq -r '.result.value'
curl -s "https://api.cloudflare.com/client/v4/zones/$ZONE_ID/rulesets/phases/http_request_cache_settings/entrypoint" \
  -H "Authorization: Bearer $CF_API_TOKEN" \
  | jq '.result.rules[]? | {description, expression, action}'

Not affected? If check 2 shows bytes arriving throughout and check 3 shows a clean zone, this runbook is not your problem. Stop here and do not buy it. If you want a second opinion first, email [email protected] with the two command outputs and the tunnel or proxy in front of the origin.

Failure modes

Five failures, each with its trigger, the observable evidence and the fix. Two are visible here.

Migration: what the paid chapter covers

The rule is simple: streaming and identity-bearing routes must be explicitly excluded from caching and from content optimization; static assets must be cacheable safely. The chapter splits the work by component:

The bundle closes with a rollback plan (which single change to revert when the stream stops after the fix) and a 24-hour watch list: CF-Cache-Status counts per path, first-byte percentiles, and 5xx around the purge window.

Appendix

• Default cache behavior — cached extensions, HTML/JSON not cached, Edge TTL defaults (reviewed 2026-11-08; page updated 2026-09-14)
• Cloudflare cache responses — HIT / MISS / BYPASS / DYNAMIC and the Age header (reviewed 2026-11-08; page updated 2026-09-04)
• Cache Rules available settings — cache eligibility, Edge TTL modes (reviewed 2026-11-08; page updated 2026-10-15)
• Cache Everything example — the dynamic-content warning (reviewed 2026-11-08; page updated 2025-10-13)
• Content compression — compressed content types, end-to-end behaviour, no-transform (reviewed 2026-11-08; page updated 2026-04-17)
• Rocket Loader overview and Ignore JavaScripts (reviewed 2026-11-08; pages updated 2026-08-14 and 2026-04-17)
• Turn off Auto Minify via API with the deprecation dated 2024-08-05 (reviewed 2026-11-08)
• Tunnel troubleshooting — “Cloudflare Tunnel is buffering my streaming response” (reviewed 2026-11-08; page updated 2026-09-02)
• cloudflared issue #199 — “Server-sent-events are buffered” (secondary: thread opened 2020-05-12, closed 2025-05-14)
• cloudflared issue #1496 — “Cloudflare proxy buffers SSE requests” (secondary: opened and closed 2025-06-24)