#178 is the last of the 0.73.0 → 0.77.0 train: it bumps backwards now, and its probe kill switch does not exist

[harper-joseph]

The 0.73.0 → 0.77.0 train shipped. This issue now tracks the two items that outlived it, and records what the train actually did — because what shipped differs from the plan in one way that matters for the next bump.

What shipped (merged; deploy is a separate, still-pending human action — kohls-pr#116)

		landed as
plugin #165	conditional sitemap fetching	merged
plugin #177	new-target fast path	merged
plugin #181	raw-document cache	merged
plugin #182	explain cadence	merged
kohls #108	sitemap anchor → 03:00 ET	merged via kohls #114
kohls #111	PDP departureAction, dry-run	merged via kohls #114
kohls #113	raw cache on catalog	merged via kohls #114

The reserved versions collapsed. Only prerender-v0.77.0 was tagged — 0.73.0, 0.74.0, 0.75.0 and 0.76.0 exist as commits on main but never as releases, so the four PRs reached kohls in one tarball. main has since moved to 0.79.0 (0.78.0 raw-cache TTL, 0.79.0 readiness contracts).

Still open

1. #178 now bumps backwards

It sets packages/plugin to 0.75.0 against a main at 0.79.0, so merging it as-is is a downgrade of the kind that hit #59. It needs renumbering to the next free version — check gh release list at the time, don't reuse the number below.

2. #178's probe kill switch does not exist

It removes if (stats.aborted) triggers.stop() and states "Only disabling the probe clears it", but resetTriggerQueue's only caller is resetChangeProbeState, marked "Tests only", and syncProbeTimers never touches the queue. So with the probe disabled the queue keeps hard-expiring pages and filing renders for up to maxPending / ratePerSecond — ~4h37m at the new defaults. Verified in the diff. This is the blocker; the renumber is bookkeeping.

3. #178 also ships inert on kohls

A live override row holds changeProbe.maxTriggersPerSweep = 90000, and the override layer wins over both file and default. Merging changes nothing until that row is deleted — and deleting it later moves behaviour 90,000 → unlimited with no deploy and no PR. Its description needs an explicit deploy-then-delete step.

4. kohls #112 — the 81% figure is still unsourced

The threshold change rests on "a real mass reprice measures 81%", which came from the note on the live override row, not a measurement. Project notes from the same week record a promo-night reprice as 7.8–18.9% of products. If that is right, 0.4 sits above the event it is meant to catch and never fires. I could not settle it from the cluster: /prerender_admin/analytics ignores end, and a distant start hits the scanCap override, so historical per-pass probe data is not reachable. Needs whoever wrote the 81%. #112 stays draft until then.

The override-row dance, which is where an incident would come from

Still live, still order-dependent:

• changeProbe.canary.threshold = 0.7. Deploy feat(console): relative freshness, a miss taxonomy, and a bot filter on Traffic #112's file first (inert while the row exists), then delete the row. The reverse order leaves the threshold at the package default of 0.1 in between — the value behind the 2026-09-10 false trip and ~18h of degraded cache.
• changeProbe.maxTriggersPerSweep = 90000. Deleting it is what activates Queue every detected change instead of deferring past a budget (v0.75.0) #178. Do it deliberately, not as cleanup.

Read the live set before assuming anything: GET /prerender_admin/config. The repo does not describe the running system.

What to watch once the train is deployed (merged to kohls-pr main 2026-09-19; NOT yet deployed — the cluster runs a pre-0.73.0 plugin; see kohls-pr#116 for the rollout)

from	metric	expectation
#165	prerender_ops sitemap_not_modified	non-zero, or the feature is doing nothing
#177	sitemap_created_soon	bounded by maxPerRun; watch the 15-min render burst
#113	raw_cache by outcome	read the refusals; has-cookie climbing is the one worth an alert
#113	route_serve /catalog/ miss rate	the number this exists to move (74–90% before)
#111	sitemap_departure_would_render / _reattached	read together, see below

Neither sitemap_not_modified nor sitemap_created_soon has a console surface — raw prerender_ops queries for now.

Before flipping sitemap.departure.dryRun: false

• Run the observation window entirely under the final Conditional sitemap fetching — skip the reconcile on a 304 (v0.73.0) #165 + Resource cache stores only third-party tag/personalization scripts (0 first-party) — it should be restricted by host #108 config, with sitemap_not_modified confirmed non-zero. Conditional sitemap fetching — skip the reconcile on a 304 (v0.73.0) #165 converts re-attributions that would have come from a 304'd child into would_render, so departure_reattached improves because of Conditional sitemap fetching — skip the reconcile on a 304 (v0.73.0) #165, not because the corpus stopped shearing.
• Know the false-departure shape: a URL listed by two children (the code records 95 in one real 800k corpus) where the owner drops it and the other child 304s → scores departed → one wasted hard-expiry and render for a product that never left.
• maxActions overflow is dropped permanently, whatever the schema doc says: Sitemap.js:601 nulls sitemapUrl before :603 records the candidate, and the feeding scan queries sitemapUrl == <child>. At ~4,173 removals/day against the 5000 default we sit at 83% of the cap.

Rollback

• Plugin: revert the component to the previous tarball. A rename leaves the old component loaded, and harper restart does not load staged components — use docker restart.
• render.raw: enabled: false. Stored rows expire on their own (@table(expiration:)), nothing to clean up.
• departureAction: sitemap.departure.enabled: false is the incident switch; it stops the behaviour without editing the route list.
• A canary bulk invalidation: delete the Invalidation row.

Conflict traps, kept for the next train

Both were hit during the trial merge and both fail far from the conflict:

• configSchema.js, Conditional sitemap fetching — skip the reconcile on a 304 (v0.73.0) #165 × Render a newly declared sitemap URL soon, not a full interval later (v0.74.0) #177. A union resolve silently drops the closing brace of Conditional sitemap fetching — skip the reconcile on a 304 (v0.73.0) #165's conditional group and the file fails to parse ~500 lines later. Keep both groups and close conditional explicitly.
• test/routeClass.test.js, feat(plugin): cache origin documents for URLs outside the render rotation; v0.76.0 #181 × feat(plugin): explain a URL's render cadence, including the ladder rung; v0.77.0 #182. Both append; feat(plugin): explain a URL's render cadence, including the ladder rung; v0.77.0 #182 also edits the import block. Take feat(plugin): explain a URL's render cadence, including the ladder rung; v0.77.0 #182's file as the trunk and append feat(plugin): cache origin documents for URLs outside the render rotation; v0.76.0 #181's block.
• Keep BOTH METRICS.md rows in a Conditional sitemap fetching — skip the reconcile on a 304 (v0.73.0) #165 × Render a newly declared sitemap URL soon, not a full interval later (v0.74.0) #177-shaped resolution. metrics.sitemapRun takes a free-form series string, so test/metrics.test.js cannot detect a dropped row — a one-sided resolution ships a series nobody can read, which is the DYNAMIC_SERIES_SLOT shape.

🤖 Generated with Claude Code