gitoriaLog in with ident

tracker

All repositories: gitoria

ReadmeCodePull requestsReleasesTicketsSettingsCommit5cb85d755cb85d75deploy.sh: a backup taken while a background job writes (tar exit 1) is a warning; archive checked with gzip -tmre5cb85d75/reports/024-tracker-revendor-final.md

6.3 KB

  1. # Report 071 — tracker: Hybriel master 190aa11d measured, NOT deployable under the rule (vendor stays 8efba065)
  3. Repo: loreana:/media/STORAGE/projects/tracker.worldapi.org. The main commit is `c1fa2f2` (on top of 8081350). Not pushed, not deployed. I never read `.env`.
  5. ## Hybriel
  6. - **Commit**: master `190aa11d`. It has #127 d98926c6 + 04df4428, the GC correctness fix fc838894 and the closure-scope fix 038d84b3.
  7. - Built ReleaseFast from a read-only archive in `.scratch/hybriel-071`.
  8. - `bin/hybriel` sha256 `feb597c4b6d416449b33e6feb9c2202d332bdc37d2da85bca642a71785bdedc8`.
  9. - In git, only `plugins/web/{WebFramework,compile}.hl` differ.
  10. - **Lambda audit** (changes since 8efba065): no commit changes lambda semantics. `tools/lambda-audit.py .` lists the same 13 as in 070, so there is nothing to fix.
  11. - **Not adopted.** bin/ and plugins/ are restored from `.scratch/pre-071` (= 8efba065, as in 8081350).
  12. - The candidate is ready in `.scratch/w071/vendor-190aa11d/`: copy its bin/ + plugins/ over the app's and run the 3 gates.
  14. ## Gates (ports 8750–8757)
  15. ```
  16. TRACKER_GATE_SHOTS=/tmp/w071/shots TRACKER_GATE_PORT=8751 TRACKER_GATE_IDENT_PORT=8752 TRACKER_GATE_CHROME=8753-8756 TRACKER_GATE_TMDB_PORT=8757 node tests/browser.mjs → 325 passed, 0 failed
  17. TRACKER_GATE_SHOTS=/tmp/w071/kinds TRACKER_GATE_PORT=8750 TRACKER_GATE_IDENT_PORT=8751 TRACKER_GATE_CHROME=8752-8756 TRACKER_GATE_TMDB_PORT=8757 node tests/kinds.mjs → 32 passed, 0 failed
  18. TRACKER_GATE_SHOTS=/tmp/w071/fr TRACKER_GATE_PORT=8750 TRACKER_GATE_IDENT_PORT=8751 TRACKER_GATE_CHROME=8752-8756 TRACKER_GATE_TMDB_PORT=8757 node tests/franchises.mjs → 50 passed, 0 failed
  19. ```
  20. - All three are green on 190aa11d, and again on the restored 8efba065.
  21. - **One gate fix, committed.** The first kinds.mjs run was 31/1 on "kinds off … no TMDB request was made".
  22. - Cause: the #19 collection seed was on in kinds.mjs and sometimes sent its TMDB request before check 1. This is a timing race, not a Hybriel bug.
  23. - Fix: `tests/kinds.mjs` now sets `TRACKER_COLLECTIONS: '0'`. It then passed 3/3 on new and 1/1 on old.
  24. - The gate screenshots are on Loreana in `/tmp/w071/{shots,kinds,fr}`. I looked at them and they are fine.
  26. ## Big pages: old 8590df63 vs new 190aa11d
  27. Setup:
  28. - Real copy of live `storage/`, taken 2026-10-02 10:50, signed in as az5b2.
  29. - 20 loads per page. Each load is one node fetch (the server's time) plus a full Chrome load (to hydrated, with images; Picard also clicks "Show all" and waits for "Show less").
  30. - Cells show `browser / server` ms average, and RSS before→after the 20 loads (max).
  31. - Loreana is shared with other workers, so repeat runs differ by up to about 30 %.
  33. | page | old fresh | new fresh | old after jobs | new after jobs |
  34. |---|---|---|---|---|
  35. | /my/unwatched (1,288 rows) | 1089 / 974, 7.8→10.3 GB (15.3) | 910 / 793, 1.7→2.1 (2.2) | 3004 / 2688, 13.0→25.6 (26.4) | 2967 / 2879, 9.4→8.2 (12.2) |
  36. | Picard + Show all (212 cast) | 240 / 109 | 245 / 225 (repeat 222 / 135) | 270 / 124 | 333 / 322 |
  37. | /my/series (126) | 292 / 194, 6.5→8.2 | 529 / 398 (repeat 380 / 277), 1.9→2.1 | 1827 / 1834, 17.9→30.1 | **4410 / 4401**, 7.4→10.8 (11.6) |
  38. | /series | 49 / 22 | 82 / 49 (repeat 89 / 30) | 102 / 64 | 166 / 160 |
  39. | /franchises (the index; seeded timelines have no franchise) | 52 / 20 | 70 / 31 (repeat 75 / 20) | 79 / 262 | 126 / 23 |
  40. | a timeline | 47 / 44 | 55 / 9 | 50 / 10 | 85 / 18 |
  42. **For reference**, two more runs on the same copy:
  43. - **Current main, 8efba065, fresh**:
  44. - /my/unwatched 4098 / 3933
  45. - /my/series 407 / 347
  46. - Picard 303 / 177
  47. - /series 70 / 56
  48. - boot RSS 2.39 GB
  49. - **New with `HL_GC_BYTES=0`**:
  50. - Fresh: every page is ≤ 1.2× old (/my/unwatched 577, /my/series 299, /series 59). RSS 3.3→3.9 GB.
  51. - After its jobs: speed is about the same as old (/my/series 1999, /series 135 = 1.3×). But RSS went 15.8→25.9 GB and the jobs peaked at 16.1 GB, which is old-binary memory.
  53. ## RSS and job durations (real copy, all first-start jobs plus a forced delta run)
  55. | | boot RSS (fresh, 60 s after listening) | RSS 60 s into the jobs run | jobs peak / median | job durations |
  56. |---|---|---|---|---|
  57. | new | 1.69–1.74 GB | 4.7 GB | **9.8 GB** / 6.4 GB | 57 min: delta 288 titles 448 s, kinds 1,789 titles 1,188 s, seed 4,774 movies (110 timelines) 2,151 s |
  58. | old (live) | 7.96 GB | 7.7 GB | 13.6 GB / 10.3 GB | 53 min: 430 / 1,082 / 2,020 s |
  60. ## Verdict: not deployable under the mission rule
  61. The rule was "every page ≤ 1.3× the old time and RSS flat". It fails in two places:
  62. - **Speed:**
  63. - After the jobs, new `/my/series` is 2.4× slower (4.4 s vs 1.8 s) and `/series` is 1.6×.
  64. - Fresh, new `/my/series` is 1.3–1.8× and `/series` about 1.7×. The `/series` gap is only tens of ms.
  65. - **RSS:** after the jobs, new swings between 7.4 and 12.2 GB. A fresh new process is flat (1.7–2.2 GB). The old binary is not flat either: 13–31 GB after its jobs.
  67. So 8081350 stays as it is.
  69. **What the lead should know:**
  70. - 190aa11d is better than what main has now (8efba065) on every number except `/my/series`:
  71. - /my/unwatched is 4.5× faster fresh.
  72. - The process that ran the jobs no longer climbs: in 070 it was 8.3 s per load and +25 MB per load; now it is 2.9 s and swinging rather than climbing.
  73. - It uses about half the memory of the live binary.
  74. - The cost that remains is GC by bytes with a large live heap. #127 fixed the growth but not the cost of each collection.
  75. - Options for the creator/architect:
  76. 1. Adopt 190aa11d anyway and restart after `collections done`.
  77. 2. Wait for a Hybriel fix to the per-collection cost.
  78. - Repro: `tests/realdata-071.sh` + `tests/realdata-071-bench.mjs` (STATUS "How to repeat"). Compare /my/series fresh vs after the jobs.
  80. ## Docs and leftovers
  81. - README now has the mission 024 (old 071) summary, the candidate under "Vendored Hybriel", the re-measured old-binary jobs under "Deploy", and the kinds gate note.
  82. - STATUS has the mission 024 (old 071) entry with the full table and "How to repeat".
  83. - Two notes from the runs:
  84. - One Chrome hydration timeout on /series (fresh new) in the first run. I re-ran on a new fresh process and all 120 loads passed. The bench now closes Chrome on an error.
  85. - The deleted copies came from ONE tar of live `storage/`.
  86. - Data copies, the old tree with its `.env` link, and the build cache are deleted. Logs and JSON are in `.scratch/w071/real-*`.
  87. - Nothing of mine is running, and 8750–8759 are free. Every process was killed by PID.

Branches

Latest commits

  • 5cb85d75deploy.sh: a backup taken while a background job writes (tar exit 1) is a warning; archive checked with gzip -tmre
  • 9abda75dtracker: report 035mre
  • 12595e47mission 035: theme re-vendored from layouts.worldapi.org 0222f67 (two corner radii: radiusSmall 5px, radiusLarge 10px); the tracker's 18 own radii -> radiusSmall/radiusLarge (--layout-radius is gone); check-theme 0; gates 379/0, 32/0, 53/0, 229/0, 26/0; real copy: every computed radius in {0, 5px, 10px, 50%}mre
  • e7305014tracker: report 032 (art + photos)mre
  • fbb903cctracker#33/#37 (mission 032): a title without a TMDB poster gets its backdrop (w780, posterFromBackdrop, shown 2:3 centre-cropped); movies store runtime, the page shows release date + runtime; art backfill (public titles without poster file / movies without runtime) and person photo backfill (tmdbProfile / photoCheck) as the last start jobs (TRACKER_ART, TRACKER_PHOTOS; off in every gate start); gates 379/0, 32/0, 53/0, 229/0, 26/0, check-theme 0; live copy: art 4977 titles in 42 min (115 posters, 15 backdrops, 4609 runtimes), The Remaining shows its backdrop + 7 minmre
  • 4ecc67b2tracker: STATUS/LOG for the t38 + t40 merge (gates 374/0, 32/0, 53/0, 229/0, 26/0, check-theme 0; live copy checks)mre
  • e5945d2fMerge t40 (tracker#40 curated franchises, Franchises menu, superseded collections) into main: deploy.sh lists all six gates (browser, kinds, franchises, pager, franchiseseed, check-theme); search.hl keeps #37's personPhotoOf + #40's importWithCredits; pager gate runs with TRACKER_FRANCHISE_SEED=0; gates 374/0, 32/0, 53/0, 229/0, 26/0, check-theme 0mre
  • b638e99dMerge t38 (tracker#38 pagination, #35 Returning/Airing label) into main: LOG/STATUS keep both sides; pager gate follows #37's /people (everyone, last updated first: seed updatedAt); gates pager 229/0, browser 374/0mre
  • f8ffa4dbtracker: report 032 + The Remaining + #39mre
  • 7565a863tracker: LOG timemre
  • b10f00c8tracker#39: double episodes — migrated episodes whose TMDB id TMDB replaced are adopted by their number in the sync (old id -> migratedTmdbId); merge.hl step 3 merges each season's doubles at start (keeper: most watches > synced > first; watches moved/parked; tombstones into mergedEpisodes, nothing deleted); tools/count-duplicate-episodes.hl; gate fixture + paths-m039; live copy 850 -> 0 in 64 s; gates 373/0, 32/0, 52/0mre
  • 8f1d4542tracker#40: superseded collections — a TMDB collection timeline whose titles are all in one curated timeline is hidden (supersededBy; kept: own page + editor finder), set by the collection seed when it makes one and by the curated build (lifted when the cover is gone); partly covered ones join that franchise; no second widget (First Contact: only Star Trek — Prime); gate franchiseseed 26/0, browser 365/0, kinds 32/0, franchises 53/0, check-theme 0; real copy 24 supersededmre
  • 9448d643tracker#35 follow-up (mission 033): TVmaze 'Running' is labelled 'Returning', or 'Airing' while a non-special episode of the two newest seasons is released within today +-7 days (data unchanged); gate fixtures Running/Airing/Aired + a special; browser 366/0, kinds 32/0, franchises 52/0, pager 229/0, check-theme 0mre
  • 7d4b293dtracker#40: "Franchises" in the main menu (desktop header after People, phone sidebar) → /franchises, marked on franchise and timeline pages; gates 365/0, 32/0, 53/0, 24/0, check-theme 0mre
  • 8751adb8tracker: report 032mre
  • 9bce1f65tracker mission 032: STATUS gate files + the hour-boundary flakemre
  • 718bfb89tracker#37 (mission 032): /people = everyone, last updated first (updatedAt stamped by the person fill; view built at boot, touched people first at once), photo + name tiles (person colour) with the /movies pagination, /people/<letter> removed; photo = our file, tmdbProfile, a cast/crew entry's profile (in-memory map at boot), else the new 'no photo' placeholder; new cast/crew/created_by people keep tmdbProfile; search people rows with the photo; /settings = the heading only; util.hl sortDesc starts from sorted runs (same result, 105k: 1.6 s -> 0.15 s); gates 369/0, 32/0, 52/0, check-theme 0; README/STATUS/LOGmre
  • 93dfb0batracker#40 (mission 034): the curated franchises — data/franchises.json (17 franchises, 31 timelines, 285 TMDB titles, movies + series, in-universe/release order, 12 TMDB collections attached); lib/franchiseseed.hl + jobs.hl franchiseSeedTick (last start job, imports missing titles via details.hl importWithCredits = the search's Add, one per step paced, then one build; franchiseseed.db: editor changes win, the creator's same-name franchise adopted / timeline left alone, 404 remembered, resumable, idempotent); timeline heads 'N titles · in-universe order' (orderKind) and wrap on a phone; series pages show the widget; new gate tests/franchiseseed.mjs (5th in deploy.sh), the others run with TRACKER_FRANCHISE_SEED=0; gates 365/0, 32/0, 52/0, 24/0, check-theme 0; real copy 196 imported, 0 failed, 7 min, restart unchanged=31mre
  • 03ec792ftracker#38 (mission 033): pagination goes exactly to the clicked page — tilelist read the clicked button's text after pagination.hl's own handler had rebuilt the buttons (real clicks only); now li.current, else the button's own text; new gate tests/pager.mjs (5 lists x 11 pages, 390/1280, real + script clicks, Back/Forward) 229/0; browser 365/0, kinds 32/0, franchises 52/0, check-theme 0mre
  • 96ba683adeploy.sh: a gate without a 'passed,' line (check-theme) no longer ends the scriptmre