VersionControl:PRs/maybelle-config/297: Difference between revisions
From PickiPedia: A knowledge base of bluegrass, old time psychedelic jams, and other public domain music
Jump to navigationJump to search
cryptograss/maybelle-config#297, as GitHub has it |
cryptograss/maybelle-config#297, as GitHub has it |
||
| (One intermediate revision by the same user not shown) | |||
| Line 5: | Line 5: | ||
|url=https://github.com/cryptograss/maybelle-config/pull/297 | |url=https://github.com/cryptograss/maybelle-config/pull/297 | ||
|author=magent-cryptograss | |author=magent-cryptograss | ||
|state= | |state=merged | ||
|base=production | |base=production | ||
|branch=magent/strip-empty-sdtp | |branch=magent/strip-empty-sdtp | ||
|opened=2026-11-09 11:51:10 | |opened=2026-11-09 11:51:10 | ||
|pushed=2026-11-09 11:50:56 | |pushed=2026-11-09 11:50:56 | ||
|merged= | |merged=2026-11-09 17:39:05 | ||
|closed= | |closed= | ||
|worked_on_by=magent, skyler (testing on an iphone 17), with justin | |worked_on_by=magent, skyler (testing on an iphone 17), with justin | ||
| Line 16: | Line 16: | ||
|deploy= | |deploy= | ||
}} | }} | ||
<pre class="pr-body">## Why | |||
Skyler's iPhone 17 shows nothing (no picture, no sound) when Safari's own player plays a release's stream. Over five rounds of test streams ([test page](https://magent-cryptograss.github.io/av1-iphone-test/)), tapped on the phone: | |||
- AV1 plays, as a plain file and as a stream, with Opus or AAC sound, in full colour range too. | |||
- The IPFS gateway isn't it: the release's stream copied byte for byte to GitHub also fails. | |||
- The release's exact stream, with **only** the 12-byte empty `sdtp` box removed from the video init segment, plays (test 18). | |||
ffmpeg 7.1 (Lavf 61.7, from `python:3.12-slim`'s Debian) writes that empty `sdtp` into every video init segment. ffmpeg 8 doesn't. The audio-only rendition has none, and it is the one that always played. | |||
## What | |||
- `_drop_empty_sdtp(init_segment)` in `transcode.py`: walks moov → trak → mdia → minf → stbl, removes any `sdtp` that is exactly a 12-byte empty full box, and shrinks every parent's size to match. It leaves alone an `sdtp` with entries, and any file with a 64-bit or malformed box. | |||
- Run on each `stream_*/init*.mp4` right after ffmpeg succeeds, before the poster, probe and pin. | |||
## Checked | |||
- Applied to the real release's `stream_360p/init_2.mp4` (`QmbYnW…`), the output is byte-identical to test 18's init segment, which Skyler's iPhone played. | |||
- 5 new tests (`tests/test_drop_empty_sdtp.py`); the full delivery-kid suite passes, 144 tests. | |||
- Not checked: a full encode on the delivery-kid image (ffmpeg 7.1) end to end. | |||
## Not in this PR: existing releases | |||
26 of the 44 HLS Releases on PickiPedia carry the empty `sdtp` (all encoded by Lavf 61.7.103); the other 18 have none. Fixing them means rewriting their init segments, which gives each a new CID, so a new Release page superseding the old via `subsequent_to`, as `scripts/reladder.py` already describes. That's a call for a person. | |||
Redeploy delivery-kid. | |||
Worked on by: magent and Skyler (testing on an iPhone 17), with Justin, in #uploads-and-embeds</pre> | |||
Latest revision as of 17:39, 9 November 2026
| State | 🟣 merged 2026-11-09 17:39:05 |
|---|---|
| Worked on by | magent, skyler (testing on an iphone 17), with justin |
| Moods | #uploads-and-embeds |
| Deploy | not said |
| Branch | magent/strip-empty-sdtp → production, by magent-cryptograss
|
| Opened · last pushed | 2026-11-09 11:51:10 · 2026-11-09 11:50:56 |
## Why Skyler's iPhone 17 shows nothing (no picture, no sound) when Safari's own player plays a release's stream. Over five rounds of test streams ([test page](https://magent-cryptograss.github.io/av1-iphone-test/)), tapped on the phone: - AV1 plays, as a plain file and as a stream, with Opus or AAC sound, in full colour range too. - The IPFS gateway isn't it: the release's stream copied byte for byte to GitHub also fails. - The release's exact stream, with **only** the 12-byte empty `sdtp` box removed from the video init segment, plays (test 18). ffmpeg 7.1 (Lavf 61.7, from `python:3.12-slim`'s Debian) writes that empty `sdtp` into every video init segment. ffmpeg 8 doesn't. The audio-only rendition has none, and it is the one that always played. ## What - `_drop_empty_sdtp(init_segment)` in `transcode.py`: walks moov → trak → mdia → minf → stbl, removes any `sdtp` that is exactly a 12-byte empty full box, and shrinks every parent's size to match. It leaves alone an `sdtp` with entries, and any file with a 64-bit or malformed box. - Run on each `stream_*/init*.mp4` right after ffmpeg succeeds, before the poster, probe and pin. ## Checked - Applied to the real release's `stream_360p/init_2.mp4` (`QmbYnW…`), the output is byte-identical to test 18's init segment, which Skyler's iPhone played. - 5 new tests (`tests/test_drop_empty_sdtp.py`); the full delivery-kid suite passes, 144 tests. - Not checked: a full encode on the delivery-kid image (ffmpeg 7.1) end to end. ## Not in this PR: existing releases 26 of the 44 HLS Releases on PickiPedia carry the empty `sdtp` (all encoded by Lavf 61.7.103); the other 18 have none. Fixing them means rewriting their init segments, which gives each a new CID, so a new Release page superseding the old via `subsequent_to`, as `scripts/reladder.py` already describes. That's a call for a person. Redeploy delivery-kid. Worked on by: magent and Skyler (testing on an iPhone 17), with Justin, in #uploads-and-embeds