tracker

6.0 KB

1. # Report 025 (w074) — tracker#30 /my/series speed + three small fixes
3. **Commit**: `2667da0` on main in loreana:/media/STORAGE/projects/tracker.worldapi.org (parent 3909810). Not pushed, not deployed. I never read `.env`. Every process was killed by its PID, and nothing of mine is still running.
5. ## 1. #30: the /my/ pages after the jobs
6. - **Cause**: the rows were already small. The cost was in reading the data behind them on every load:
7. - /my/series fetched every followed title's whole record (cast + crew) plus one episode per watch (11,648 fetches).
8. - /my/unwatched and /my/schedule fetched every season and episode of all 128 followed titles.
9. - After the jobs the heap is large, so all that allocation drives the byte-triggered GC.
10. - **Fix**: three slim caches in `shows.hl`, plus a kept watch set in `watches.hl`:
11. - `cardOf`: a title's card (title, year, kind, link, poster).
12. - `episodeRowsOf`: a title's episodes as `{ id s e title release }`.
13. - `episodeKeyOf`: a watched episode's `{ show s e }`.
14. - `watchedSetOf`: the user's watched set, kept until that user's next watch.
15. - They are filled at boot (~460 ms) and refreshed in `catalog.hl refreshCatalogShow`, which every writer of a title already calls.
16. - The poster tiles use them too (`/`, `/movies`, `/series`, genres).
17. - **Times** on a real copy, signed in as az5b2, 20 loads per page, `server / browser ms` averages:
19. | page | before, after jobs | after, after jobs | after, fresh |
20. |---|---|---|---|
21. | /my/series | 1767 / 1927 | **63 / 123** | 53 / 113 |
22. | /my/shows | 3539 / 3397 | 39 / 73 | 64 / 73 |
23. | /my/unwatched | 4063 / 4540 | **176 / 329** | 172 / 336 |
24. | /my/movies | 114 / 155 | 8 / 38 | 6 / 36 |
25. | /my/schedule | 4518 / 4543 | 14 / 52 | 13 / 58 |
26. | RSS during the bench | 4.7 → 21.3 GB peak | 1.36 → 1.68 GB | 1.83 → 2.13 GB |
28. - The copy was taken while live was in the middle of its seed. Its first start therefore ran a forced delta sync of 316–333 titles plus the remaining 1,802 seed movies, about 20 minutes in total.
29. - So "before" here is 1.8 s, not the 4.4 s from report 024, because the heap was smaller.
30. - The before process was benched while the after process was running its jobs. Its RSS growth pushed Loreana into swap, and 3 GB of the after process was swapped out during its own bench.
31. - **Output is unchanged**: the page text of /my/series, /my/shows, /my/unwatched, /my/movies, /my/schedule, /series and / is identical in three comparisons:
32. - before vs after, both fresh;
33. - the after process that ran the jobs vs a fresh process on the same storage, so the cache refresh after 333 synced titles is right.
34. - **One catch on the way**: my first version of "last watched" went through the title's seasons and showed Reacher S02E08 instead of S03E08. The creator's S03 watches sit on a migrated duplicate season that is not in the title's season list. It now goes by the watched episode's own `show`, as before.
36. ## 2. The timeline page shows its name once
37. The name is now only in the h1. The list head keeps the count and the sort toggle.
39. ## 3. The franchise widget on a phone
40. - The widget is now inside the show header:
41. - on a phone, right under the poster and before the genres, plot and cast;
42. - from 40rem, in the right column under the title and links.
43. - It is hidden when the title is in no timeline.
45. ## 4. Star Trek: First Contact
46. - **The tmdbId is not wrong.** 199 is the movie (`/movie/199`).
47. - **Cause**: the old #9 sync asked `/tv/199` (the bug fixed in mission 006). It stored that TV show's overview ("Kabuhayang Swak na Swak…"), status `Ended` and seasonsCount/episodesCount 0 on the movie. Neither the migration nor the search import caused it.
48. - **Fix in the repair path**: `details.hl hasTvLeftovers` treats a movie that has season or episode counts as incomplete.
49. - The details repair asks `/movie/<id>` again and drops the counts, so it runs only once per record.
50. - On the copy: `details repair: start, 1 incomplete titles` → `done 1/1 complete=1`. The record now has the movie's text, status Released and counts null.
51. - It is the only such movie in the real data.
52. - **Shared TMDB ids**: 0 records share a TMDB id with a different title (`tools/count-tmdb-ids.hl`).
54. ## 5. Gates
55. - `TRACKER_GATE_SHOTS=/tmp/w074/shots TRACKER_GATE_PORT=8720 TRACKER_GATE_IDENT_PORT=8721 TRACKER_GATE_CHROME=8722-8726 TRACKER_GATE_TMDB_PORT=8727 node tests/browser.mjs` → `327 passed, 0 failed` (+2: the TV-leftover repair, and the next start has nothing to do).
56. - The same env with `tests/franchises.mjs` → `52 passed, 0 failed` (+2: widget position at 390 and 1280; the timeline name once).
57. - The same env with `tests/kinds.mjs` → `32 passed, 0 failed`.
59. ## Screenshots (on Loreana, I looked at them)
60. - Gate: `/tmp/w074/{shots,fr,kinds}`.
61. - Real data: `/tmp/w074/real/`:
62. - First Contact full page at 390: the widget is under the poster. At 1280 it is under the title.
63. - The A Kid in… timeline: the name appears once.
64. - /my/series, /my/shows, /my/unwatched, /my/movies at 390 and 1280: no page scrolls sideways.
66. ## Docs
67. - README: new section for mission 025, the gate counts, and the Files table.
68. - STATUS: new top entry with the table and "How to repeat" (`tests/realdata-074.sh` and `tests/realdata-074-bench.mjs`).
69. - LOG.md: one line.
71. ## Open points
72. 1. **Duplicate title records**: 68 type+tmdbId pairs are each held by 2–6 records of the SAME title.
73. - Mostly adult filmography titles, but also Archive 81 ×6, Ally McBeal ×2 and The Flight Attendant ×2.
74. - Some of them share one slug.
75. - Some import path creates them. This needs its own ticket.
76. 2. **One process per storage**: the caches assume it, as `watches.hl` already does.
77. 3. **Deploy**: boot is about 0.5 s longer (the cache warm-up). First Contact gets repaired on the first start after the deploy.
79. ## Hybriel
80. No new bugs. Two things the workers' guide should mention:
81. - `let x = map[k]` deep-copies the value: 3.7 ms per 20k-row list. Returning `map[k]` from a function does not copy.
82. - Inside an object literal, a key set earlier shadows a variable of the same name: `{ s = e.seasonNumber e = e.episodeNumber title = e.title }` fails because `e` is the number by then.