> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tuturuuu.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Meet volunteer load test checklist

> Prepare, run and record a live Meet session with about 50 volunteers.

Use this checklist for a hosted Meet session with real people, devices and
networks. For how layers, thumbnail mode and audio-only fallback behave, see
[Meet media quality](https://docs.tuturuuu.com/platform/features/meet-media-quality.md). The hosted
rehearsal steps are in [Meet calls](https://docs.tuturuuu.com/platform/features/meet-calls.md).

## Before the session

* [ ] Record the deployed Meet and `meet-realtime` commit SHAs. Confirm that the
  room server's `ready` message lists both `ascii-simulcast` and
  `layer-update` in `mediaCapabilities`:
  * Without `ascii-simulcast`, clients publish legacy `h`/`m`/`l` layers and
    request layers only from legacy publishers. A subscription to a publisher
    advertising current layers carries no layer preference.
  * Without `layer-update`, clients still choose a layer when they subscribe
    but never change it in place. Receive-side camera telemetry is also off.
* [ ] Confirm the room's effective participant cap admits the planned head
  count, plus about ten for rejoining devices.
* [ ] Confirm the publisher cap admits every admitted participant who is not a
  viewer, including muted participants and those with cameras off.
  Admission enforces publishers separately. The room server keeps media
  health for at most 150 devices.
* [ ] Confirm the host can open **Settings → Connection → Room media health**.
  Only participants who can manage the room can open it. A classroom
  co-host without participant management cannot.
* [ ] Ask volunteers to use a current browser and to close other video calls.
  Include at least one low-end laptop, one phone on mobile data and one iOS
  Safari device.
* [ ] Decide on any A/B cohort in advance (see below). Write down which
  volunteers run which flags.

## Flags and kill switches

Every behaviour is on by default. Turn one off on a single device through the
browser storage key `tumeet:media-flags`, then reload the page:

```js theme={null}
localStorage.setItem('tumeet:media-flags', JSON.stringify({ hiddenTabSaver: false }))
```

| Flag | Turns off |
| - | - |
| `layerSwitching` | Per-tile simulcast layer switching |
| `opusDtx` | Microphone silence suppression |
| `cpuAdaptation` | Dropping layers and frame rate under CPU pressure |
| `hiddenTabSaver` | Lowering, then closing, remote video in a hidden tab |

Thumbnail mode has its own per-device setting under **Settings → Media**, also
stored as `tumeet:small-mode` (`auto`, `on` or `off`). Automatic mode turns on
at 25 or more participants. It also turns on for a single receiver whose
congestion ceiling has reached the lowest layer, which the weak-link scene can
trigger in a smaller room.

To restore defaults on a device, run
`localStorage.removeItem('tumeet:media-flags')` and reload.

If one behaviour misbehaves for everyone, ask volunteers to set its flag to
`false` and reload. That takes effect immediately, without a deploy.

## During the session

Run these scenes for at least five minutes each:

1. **Gallery:** everyone's camera on, no spotlight.
2. **Spotlight:** one presenter spotlit.
3. **Screen share:** a presenter shares a screen with small text.
4. **Muted majority:** everyone muted except two speakers. This includes at
   least one silent join and one unmute after more than 30 seconds of silence.
5. **Hidden tab:** five volunteers switch to another tab for 30 seconds, then
   return.
6. **Weak link:** one volunteer throttles their network or moves to a weak
   connection.

Every 30 seconds, the host records from **Room media health**:

* Reporting devices, devices measuring cameras and devices measuring freezes.
* Send and receive rate, p50 and p95.
* Packet loss p95 and round-trip time p95.
* Freezes per device-minute.
* Layer counts.
* Counts of CPU-limited, audio-only, hidden and thumbnail devices.

Volunteers note anything perceptible, with the time and their device:

* Speech that cuts out.
* Video that freezes.
* A screen share that blurs.
* A spotlight that stays blurry.
* Video that does not return within about two seconds of refocusing the tab.

Devices that run scripts can read `globalThis.__meetMediaStats` every two
seconds. It holds rates and counters only, and readings expire six seconds
after their last measurement.

## Pass criteria

* Speech stays intelligible throughout, including right after unmuting.
* Freezes stay under one per device-minute in the gallery and spotlight scenes.
* In a room of 40 or more, the receive rate p95 stays at or under about
  2.5 Mbps per device. The receive rate covers simulcast camera video only;
  the send rate also includes audio and screen share.
* Screen text stays legible.
* A hidden tab's video returns within about two seconds.
* The low-end laptop is not CPU-limited for the whole session. Sample
  `globalThis.__meetMediaStats?.sender?.cpuLimited` on that laptop, with
  `cpuAdaptation` on, in a Chromium or WebKit browser that reports encoder
  limitation reasons. In any other setup, `false` is missing evidence rather
  than a pass, and so is a missing reading.

The freeze and receive-rate criteria need fresh camera readings. Room media
health computes receive rates only over devices that measured cameras in their
last report interval, shown as **Devices measuring cameras**. It computes the
freeze rate only over those whose browser also reported freeze counters, shown
as **Devices measuring freezes**. It divides by the time each device actually
measured, so a device that measured for only part of an interval cannot dilute
the rate. A device's measured freeze time still counts when its final receive
reading in that interval was missing. A device
measures nothing in these cases:

* `layerSwitching` is off.
* The room server lacks `layer-update`.
* The device is hidden or audio-only.
* The device has no camera feed.

Score the receive rate only when **Devices measuring cameras** covers every
device expected to be watching cameras in that scene. Score the freeze rate
only when **Devices measuring freezes** covers them. Record uncovered devices as
missing evidence. Do not sum polled
`globalThis.__meetMediaStats` values: it is a latest-reading view, not a
sample log.

## After the session

* [ ] Record the SHAs, head count, scene timings, health readings and volunteer
  notes in the coordination handoff. Keep names and personal details out of
  it.
* [ ] Open follow-ups for every failed criterion, with the scene and time.
* [ ] Record any flag the session needed to turn off, and why.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.