First day in India: HEVC with no picture, RTMP instead of SRT, and four recordings, 25.6 GB
The first feed from the venue arrived as HEVC. MistServer's HLS playlist had audio and no video, and the encoder reported 0 fps while the real rate was 30. The stream existed, the viewer saw nothing, and no check noticed. I rank that worse than an outage, because an outage at least gets noticed.
Before the match we moved the encoder to H.264. On 15 November Glydecast streamed and recorded two matches of the CAVA men's volleyball league from KIIT, court 1. Kickoffs by the log: 08:30 and 14:00 UTC. Viewers got H.264 1080p50 over HLS.
I tested the camera over SRT the day before and dropped it
On 14 November the camera picture over SRT was damaged: a Chrome preview died after about 10 seconds, and a 40-second sample had seven decode errors. Mist showed about 1.6% packet loss and no recovered packets. The latency reading looked wrong, but I never confirmed its units or the cause, so I don't treat the diagnosis as closed.
We moved the camera to RTMP over TCP. A 35-second live test played 1,715 frames with none dropped, and a separate 45-second sample had no decode errors. A match runs two hours; the test ran for half a minute. The venue's own TV feed stayed on SRT and held through both matches, so nothing about SRT follows from this day.
Pakistan in July looked the same: SRT broke up and we moved the venue feed to RTMP/TCP. The incident record puts the blame on UDP handling at the venue and on an international leg. In India I have the same symptoms and no evidence of the same cause. The rule I took into match day is short: test the actual route and keep the transport that gives a clean picture.
Recording on ingest, one copy and four hashes
The old recorder read the viewer-facing HLS. For India we record the incoming feeds directly on the MistServer ingest, so a playback fault can't take the recording with it. It is still one copy.
The two matches produced four MKV files: a camera file and a TV-feed file for each match. All four are H.264 1080p50 with AAC, the camera at about 2.6 Mbps and the TV feed at about 8.2 Mbps, 25.6 GB in total. The first match was 4.3 GB of camera and 13.5 GB of TV feed. The last 30 seconds of each file decoded without errors.
After the matches I uploaded all four files to a private Cloudflare R2 bucket. The first attempt returned 403 on every file, because rclone tried to create the bucket first. Then it turned out that rclone 1.60 prints a 501 error even when an upload succeeded. The log can't be trusted, so I check only the SHA-256, at the source and again after a full read-back from R2. Sizes and hashes matched on all four.
The upload ran at 20 MB/s. I had copied that cap from the live-match settings and left it on for an idle machine. (The 20 was mine.) Without it the upload ran at about 106 MB/s. The local copies are still on the ingest box.
What else went wrong
The auth server hit its per-minute limit on session checks, and Ops answered "Authentication service unavailable" twice before the first kickoff. I raised the limit and will watch it at the next match. An hour of clean log is not a check.
At 15:26 UTC an alert of my own fired: the loss rule took a feed with nothing to compare against for 100% loss. That is the likely cause; I haven't read the check's output from that minute.
One measurement changed the setup before the first whistle. A cold, on-demand playback path took 16 to 24 seconds to return its first HLS playlist, so prewarming stays on, and it ran at 08:00 UTC. That is the playlist time; I haven't measured time to the first decoded frame.
What I haven't measured
→ Decoding of the files end to end: only the last 30 seconds of each were checked.
→ The recordings' duration: MKV files here have no duration header.
→ Whether the SRT failure in India and the July one in Pakistan share a cause.
→ Whether the auth limit survives a full day of matches, and which client made the extra calls.
→ Why the alert fired at 15:26: I haven't read the check's output.
→ Time to first frame for a viewer.
→ The rest of the tournament: so far there are two matches.
I was wrong three times: the rclone command, the speed cap and the alert rule. At the next match I start with the login log, then check the files.
Related reading
- 5 min read
Gemma 4 26B caught up with Qwen 3.8 27B on my tasks, and answers in 0.3 seconds instead of 5–7
- 7 min read
Qwen3.8-Flash-Next, 125B parameters at home: 16 GB of VRAM, 40 GiB of experts in RAM, 32–40 tokens/s
- 2 min read
I asked for an AI micro-drama experiment for 900 RUB of GPU. The edit lost half the dialogue