git_admin/obsidian-vault
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[2026-10-16] eagle: work/projects/cpm-pr0-episode-telemetry-summary-2026-10-16.md

Browse Source
This commit is contained in:
Alexey Martemyanov committed 2026-10-16 18:04:59 +06:00
1 parent dd2862f25a
commit 942cc65127
1 file changed
+28 -13
Download Patch File Download Diff File Expand all files Collapse all files
work/projects/cpm-pr0-episode-telemetry-summary-2026-10-16.md
+28 -13
View File
@@ -1,5 +1,5 @@
---
title: CPM PR0 — one terminal report per failure episode
title: CPM PR0 — complete telemetry change
date: '2026-10-16'
tags:
- work
@@ -8,26 +8,41 @@ tags:
---
# CPM PR0: one terminal report per failure episode
Local changes for [PR0 #7376](https://github.com/duckduckgo/apple-browsers/pull/7376), in `alex/cpm-telemetry-pr0`, uncommitted.
This describes the uncommitted replacement in `alex/cpm-telemetry-pr0`, based on `origin/main` (`874b465856ca`). It is a proposed revision of [PR0 #7376](https://github.com/duckduckgo/apple-browsers/pull/7376), not a description of that PR's currently published diff.
**What and why.** Replace separate `init_failed`, `stuck`, `recovery` and CPM reload-failed events with `web_extension_cpm_episode`. Each report contains the initial failure and its observed outcome, so diagnosis no longer requires joining a sequence of events.
**Change and purpose.** Replace the initialization-failed, stuck, recovered-with/without-reload and CPM post-reload-failure pixels with `web_extension_cpm_episode`. Put the initial failure and observed outcome together, avoiding reconstruction of a failure/recovery chain from separate events.
**When it fires.** The first timeout opens an episode and captures one diagnostic snapshot. Later failures and reloads update local state. A qualifying dashboard response or a five-minute deadline closes the episode and emits one logical report. The deadline works without further navigations. Existing stuck detection and automatic reload behaviour are preserved.
The first timeout opens an existing monitor episode and captures one diagnostic snapshot. Later failures update local stuck/reload state. A qualifying dashboard response or a five-minute deadline emits the terminal report. The new timer closes quiet episodes without waiting for another navigation; closing an episode cancels its timer and prevents duplicate reports. Stuck detection and automatic reload triggers retain their existing rules.
**What it contains.** Initial failure reason, stuck confirmation, latest reload state and observed outcome. Responses distinguish the original measurement, a later measurement on a failing tab, another tab and ambiguous attribution. Response delay uses three buckets: under 30 seconds, 30–under 60 seconds, and one minute to the deadline.
**Five added summary parameters.**
The initial snapshot records fixed facts about the extension context, background view/process, tab and native-handler registration, controller match and extension scripts.
- `reason`: the first timeout's context — `session_restoration`, `tab_crash`, `extension_reload` or `other`. Previously encoded in failure pixel names.
- `was_stuck`: whether a later eligible non-restoration navigation, begun after the first failure, also timed out.
- `outcome`: `failed_measurement_response` for the measurement that timed out; `failed_tab_response` for a later measurement on a failing tab; `other_tab_response` for another tab; `unattributed_response` for an ambiguous current-document match; or `no_response_before_deadline`.
- `reload_state`: `none` (no reload observed), `pending` (replacement started), `failed` (replacement failed), `loaded` (replacement loaded), or `response_timed_out` (its first eligible messaging measurement timed out). Loaded alone does not prove messaging recovery.
- `response_after`: elapsed time since the first failure, bucketed as `under_30s`, `30s_to_1m` or `1m_to_deadline`. Omitted for the no-response outcome.
**Privacy and counting.** Exclude `bg_events`, raw errors, identifiers, diagnostic counts and app/memory-pressure ages. Rare parameter combinations may still be correlatable.
**Eight retained diagnostic parameters.** All are optional booleans, captured at the first failure; unavailable facts are omitted.
Keep `daily+count`: daily estimates installations with a completed episode that day; count measures completed episodes. The first report each day sends both legs with the same payload; subsequent reports send count only.
- `context_loaded` — extension context loaded; `bg_view_alive` — current background view exists.
- `bg_process_alive` — its process is alive; `bg_process_responsive` — WebKit considers it responsive.
- `tab_in_context` — failing tab registered in WebKit; `handler_registered` — autoconsent native handler registered.
- `tab_controller_match` — tab and context controllers match; `tab_has_ext_scripts` — extension scripts remain registered.
**Fix attribution.** Graveyard attribution is removed from the episode and its diagnostic payload. Add fix-specific application/outcome metrics in the relevant fix PRs when needed. Preventive fixes must also count eligible attempts that succeed without a failure episode. Current diagnostic facts do not distinguish all five failure groups.
Standard metadata remains: `appVersion` on both platforms, plus `pixelSource` and `channel` on macOS.
**Limits.** A dashboard response does not prove document-level `initResp` delivery or recovery of every failing tab. A deadline means no qualifying response was observed. Unfinished episodes can be lost when the process exits.
**Parameters removed from transmission.** `bg_events`, `context_errors`, `network_restarted`, `critical_memory_age`, `bg_view_creations`, `bg_view_leaked_count`, `app_version_change`, `app_launch_age`, `background_graveyard_enabled` and `recovery_from`. Episode state and response attribution replace `recovery_from`. Graveyard attribution is also removed from the diagnostic model and monitor.
Retired pixel definitions and unused parameters are removed on both platforms. The manager still sends the separate `debug_web_extension_reload_failed` event.
For `bg_events`, removal means exclusion from this pixel's payload, not deletion of local lifecycle collection or its existing encoder. This reduces transmitted history but loses event order. The retained booleans are not a lossless replacement for lifecycle signatures and cannot distinguish all five failure groups. Rare combinations of the remaining fields may still be correlatable.
**Validation.** All 742 WebExtensions tests passed after removing graveyard attribution, including a regression test proving the monitor neither reads nor reports its treatment state. macOS PixelKitMock verification, both platform definition validators and strict SwiftLint also passed.
**Frequency and platform wiring.** Both adapters use the same base event name, with no added `_macos` suffix. Keep `daily+count`: the first completed episode of the day sends both legs with identical parameters; subsequent episodes send count only. Daily deduplication is by event name, not reason/outcome, so daily estimates overall affected installations rather than unique installations for every outcome.
Related: [[cpm-status-2026-10-15|CPM status]], [[cpm-readiness-telemetry-tech-design-2026-10-15|Readiness telemetry design]].
**Definition cleanup.** Remove 11 retired CPM pixel definitions and 10 unused parameter dictionary entries per platform. The separate `debug_web_extension_reload_failed` remains emitted by the extension manager.
**Scope and limits.** This local revision does not include the original #7376 additions: `bg_ready_after`, `bg_view_age`, `bg_load_error_after`, `bg_load_error_app_state`, `app_state`, `open_tabs`, `loading_tabs`, `recovery_document` or `recovery_after`. Missing diagnostic facts need an explicit follow-up decision; fix-specific application/outcome metrics belong in the relevant fix PRs. Preventive-fix metrics must include eligible attempts that succeed without a failure episode.
A dashboard response proves neither document-level `initResp` delivery nor recovery of every failing tab. A deadline reports absence of qualifying evidence, not that all tabs died. Unfinished episodes can be lost when the monitor is released or the process exits.
**Validation.** All 742 WebExtensions package tests passed on macOS. Coverage includes first-snapshot preservation, response attribution, reload generations, deadline-only reporting, duplicate prevention, payload allowlisting and absence of graveyard attribution. macOS PixelKitMock checks, both platform definition validators, wide-event consistency checks and strict SwiftLint passed.
Related: [[cpm-status-2026-10-15|CPM status]], [[cpm-readiness-telemetry-tech-design-2026-10-15|Original readiness telemetry design]].