Warm the Start server graph on isolate first fetch - #1681
Merged
Conversation
Page latency was p50 19-49ms daily up to 2026-08-16 and 2728ms the day after the 21:20 deploy. Not load: 2026-08-11 served more total traffic (5.43M vs 4.38M) with pages at p50 41ms, so the volume explanation does not hold. What changed is where MCP work runs. The old hibernatable Agent bridge handed /mcp to the DO almost immediately; the v2 stack authenticates and dispatches in the worker, spinning far more worker isolates. /mcp returns before fetchHandler, so those isolates never load the Start graph, and page requests land on them cold paying loadEntries: p50 3.1s, against 33ms for the request's own work and 9-104ms warm. Re-lands #1628. Its revert came during the Aug 17 storm (11.67M requests against a 2.6-4.3M baseline); volume is back to baseline and exceededMemory is currently zero. START_GRAPH_WARM=false disables it.
Contributor
Cloudflare preview
Sign-in is Cloudflare Access (one-time PIN to an allowed email). The preview has its own database and encryption key; it is destroyed when this PR closes. |
RhysSullivan
added a commit
that referenced
this pull request
Aug 19, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
It is not load, and it is not the graph
Page latency was p50 19-49ms every day through 2026-08-16, then 2728ms the day after the 21:20 UTC deploy. The volume explanation behind #1628/#1634 does not survive the data:
More traffic on Aug 11, fast pages. So the reconnect storm was a coincident symptom, not the cause.
What actually changed
Where MCP work runs. The old hibernatable Agent bridge handed
/mcpto the Durable Object almost immediately. The v2 stack authenticates and dispatches in the worker —mcp.auth.jwt_verifyalone runs ~35k times per 6h there. That spins up far more worker isolates, and/mcpreturns beforefetchHandler, so none of those isolates ever loads the Start graph. Page requests land on them cold.Measured, splitting the costs apart on a cold isolate:
loadEntriesgraph importThe same production bundle in local workerd serves the same path in ~3ms, so this is a cold-isolate cost, not slow code.
Why re-land #1628
#1628 warmed on first fetch and worked. #1634 reverted it when memory kills jumped 500-900 → 3.7k-9.4k per 5min — but that landed during the Aug 17 storm, 11.67M requests in a day against a 2.6-4.3M baseline. Volume is back to baseline and
exceededMemoryis currently zero, so the condition that made warming expensive is gone.I also tried a threshold variant (warm only after an isolate proves persistent, #1679) precisely to avoid that memory risk. It did not work — page requests are ~1 per isolate, so the threshold was rarely met before the request that needed it, and the warm import itself takes ~3s. Reverted in #1680.
START_GRAPH_WARM=falsedisables this without a deploy.Verification
lint0 errors,typecheck44/44,apps/cloudvitest 280 passed. I will watchexceededMemoryand page latency after deploy and flip the switch if kills move.