> For the complete documentation index, see [llms.txt](https://docs.roboflow.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.roboflow.com/deployment/self-hosted/inference-server/configuration/video-configuration.md).

# Video Configuration

Set limits for managed video workers and configure server-side video sources in self-hosted Inference.

Configure video workers on the server that runs them. These settings cover the integrated stream manager; they do not set limits on the managed Serverless Video Streaming API or a local Python InferencePipeline.

See [Security Configuration Migration](/deployment/self-hosted/inference-server/configuration/security-migration.md#release-status) for version scope.

## Managed pipeline limits

`STREAM_MANAGER_MAX_ACTIVE_PIPELINES` defaults to `8` and caps the total managed pipeline processes, including busy and idle workers. The effective cap is the larger of the configured value and `STREAM_API_PRELOADED_PROCESSES`, so it cannot fall below the number of preloaded workers.

<table data-search="false"><thead><tr><th>Variable</th><th>Default</th><th>Behavior</th></tr></thead><tbody><tr><td>STREAM_MANAGER_MAX_ACTIVE_PIPELINES</td><td>8</td><td>Limits managed pipeline processes, raised to the preload count when needed.</td></tr><tr><td>STREAM_API_PRELOADED_PROCESSES</td><td>0 in the core library; images may override it</td><td>Sets the target number of warm idle workers.</td></tr><tr><td>STREAM_MANAGER_MAX_RAM_MB</td><td>Unset</td><td>Optionally rejects new pipeline allocation based on measured and estimated RAM use.</td></tr></tbody></table>

At the process cap, the manager can reuse an idle worker but refuses to spawn another process. The background warm pool also respects the cap; a busy worker consumes a slot even when the target idle count has not been reached.

### Set capacity

Pass capacity settings through an environment file when starting the server:

```dotenv
STREAM_MANAGER_MAX_ACTIVE_PIPELINES=16
STREAM_API_PRELOADED_PROCESSES=2
```

```bash
inference server start --env-file inference.env
```

Choose a positive cap that fits your model and hardware, and terminate unused pipelines before increasing it. This is a process limit, not a per-user quota, a request rate limit, or a hard RAM or GPU memory limit.

## Server-side video references

The rules in this section apply to the pending runtime-hardening build in the [migration guide](/deployment/self-hosted/inference-server/configuration/security-migration.md#video-source-validation). Standard stream manager and enterprise stream request schemas validate the supplied media reference before it reaches the video backend.

Supported inputs are:

* Integer camera indices, such as `0`, and numeric CSI references, such as `csi://0`.
* Absolute or explicit relative paths, such as `/video.mp4`, `./video.mp4`, `../video.mp4`, or Windows drive paths.
* Encoded URLs using `http`, `https`, `rtsp`, `rtsps`, `rtspt`, `rtspst`, `rtmp`, `rtmps`, `udp`, `srt`, `rtp`, `tcp`, or `file`.
* Lists of supported references where the request schema permits them.

The chosen image must provide a decoder for the transport, and paths must be readable inside the server container. This validation does not restrict camera network destinations; image URL controls are a separate policy.

### Update existing references

Write `video.mp4` as `./video.mp4`. Encode whitespace and quotes in URLs, and rename local paths containing `!`, `=`, or quotes; unsupported schemes and raw GStreamer descriptions are rejected by default.

### Use raw GStreamer pipelines

For trusted workloads that require raw launch syntax, set:

```dotenv
ALLOW_UNSAFE_GSTREAMER_PIPELINES=True
```

This bypasses the media-reference validation and restores raw pipeline interpretation, including the capabilities of installed plugins. Supported RTSP and RTMP URLs do not need this opt-in, and it does not install missing GStreamer plugins.

### Client-side video sources

These checks apply to references sent to the server's stream request schemas. They do not rename files opened locally by SDK `VideoFileSource` or change direct local Python media use.

See [Video processing with Workflows](https://docs.roboflow.com/workflows/deploy/video-processing) for choosing where video is opened, and [server security](/deployment/self-hosted/inference-server/configuration/security.md) for who can manage pipelines.