[Karanvir1729]

Summary

• ingest() registered its counting with waitUntil only after awaiting the R2 write. A client that reset its connection during that write (about 150 ms; for example a browser being shut down) cancelled the handler, so the batch stayed in R2 but never reached the dashboard's counters, the export index or forwarding. A reset before the write lost the batch altogether.
• handle() now passes the whole of ingest() to opts.waitUntil from its first line. The response is unchanged and other routes are not touched.

Evidence

Local workerd with the real entry point (src/index.ts). R2's timing is simulated in the two reset rows, where the client resets 100 ms in. Each cell is batches stored in R2 / counted on the dashboard:

Client behaviour	Batches	Before	After
Waits for the reply	20	20 / 20	20 / 20
Closes normally	10	10 / 10	10 / 10
Resets, write landed but not yet acknowledged	10	10 / 0	10 / 10
Resets before the write starts	10	0 / 0	10 / 10
Resets with half the body sent	5	0 / 0	0 / 0

A batch sent twice still counts once. A plain close (FIN) does not cancel the handler under workerd, so reproducing this needs a connection reset (socket.resetAndDestroy() in Node).

In production the cancellation itself can't be observed, because logging is off and the error was swallowed. What points to it:

• One app had 13 of 198 page loads in R2 and not in the dashboard. All 13 were solo visits by one automated browser, and every batch of the other 185 page loads was counted.
• Replaying those 16 batches locally over an open connection counted all 13 page loads, so their content is not the cause.
• Cloudflare recorded 72 client disconnects on the collector from 20:00 to 03:00 UTC, all during that browser's active hours.
• Across all apps since Nov 2, 7 of 355 page loads with a session event are missing from the dashboard.

Test plan

• cd telemetry-worker && npx vitest run: 363 pass. The two new tests in test/index.test.ts failed before the fix.
• The local workerd reset test above, before and after. The harness is not committed.
• After deploy, compare the R2-fed admin sessions with the dashboard's sessions rows for new page loads. The gap should close.

Deploying

Not deployed. The live worker runs 931b6f45, so deploying from main after this merges also ships e8767ebf and 22590994, which change text in pages.ts and studio.ts. There are no migrations. Counts already missed are not backfilled.

🤖 Generated with Claude Code