> 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/workflows/developer-guide/developer-guide/profiling.md).

# Workflow profiling

Measure Workflow execution timing and find bottlenecks with Chrome-compatible profiling traces.

You can enable profiling to measure execution timing of your Workflow and identify performance bottlenecks. Profiling traces are compatible with Chrome's built-in tracing tool for visualization.

## Example Workflow

The screenshots in this guide use the following workflow as an example; it includes 2x object detection, dynamic cropping, a Google Gemini call, and several visualization steps:

<figure><img src="https://1126624165-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoO0Ceoy9R0lHDZATcPyh%2Fuploads%2Fgit-blob-1d6acedcd335ef7c260cef8595ad06f80273b716%2Fworkflow-profiling-example.webp?alt=media" alt="Example workflow in the Roboflow editor"><figcaption><p>Example Workflow used for the profiling screenshots below</p></figcaption></figure>

## Profiling with the Inference SDK

To enable profiling when running workflows via the HTTP client, you first need to run self-hosted [Inference Server](https://docs.roboflow.com/deployment/self-hosted/inference-server) with `ENABLE_WORKFLOWS_PROFILING=True` env variable. Then you can enable profiling when running workflows via the HTTP client by setting `enable_profiling=True`:

```python
from inference_sdk import InferenceHTTPClient, InferenceConfiguration

client = InferenceHTTPClient(
    api_url="http://localhost:9001", # Serverless Cloud API does not support profiling
    api_key="API_KEY"
).configure(InferenceConfiguration(api_key_transport="header"))

result = client.run_workflow(
    workspace_name="workspace-name",
    workflow_id="workflow-id",
    images={"image": "YOUR_IMAGE.jpg"},
    enable_profiling=True
)
```

When profiling is enabled, trace files are automatically saved to `./inference_profiling/` by default. Each trace file is named `workflow_execution_tack_{timestamp}.json`.

## Viewing Profiling Traces

Profiling traces are saved in Chrome Tracing Format. To visualize them:

1. Open Chrome and navigate to `chrome://tracing/`.
2. Click **Load** and select your trace file (e.g., `workflow_execution_tack_2026_03_30_10_46_04.json`).
3. Use the timeline view to inspect execution timing for each workflow step.

Clicking on any span shows its duration, category, and arguments - such as which block's `step_execution` you are looking at:

<figure><img src="https://1126624165-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoO0Ceoy9R0lHDZATcPyh%2Fuploads%2Fgit-blob-137d29561237d184a31937396bde007609180006%2Fworkflow-profiling-trace.webp?alt=media" alt="Chrome tracing timeline with a selected step"><figcaption></figcaption></figure>

Most of the time was spent on 2x object detection (process 57) and Gemini call (process 78). If you zoom-in, you can see other steps (detection offset, dynamic cropping, visualization, etc.) between the object detection steps:

<figure><img src="https://1126624165-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoO0Ceoy9R0lHDZATcPyh%2Fuploads%2Fgit-blob-d9dfe6a02d28e5bcbb72fc4b796e793ed75bb454%2Fworkflow-profiling-trace-2.webp?alt=media" alt="Chrome tracing aggregated statistics view"><figcaption></figcaption></figure>

## Key Details

| Detail              | Value                                                   |
| ------------------- | ------------------------------------------------------- |
| **Trace format**    | JSON, compatible with Chrome's `chrome://tracing/` tool |
| **Output location** | `./inference_profiling/` by default (configurable)      |
| **File naming**     | `workflow_execution_tack_{YYYY_MM_DD_HH_MM_SS}.json`    |

## Limitations

* **Compilation overhead**: Workflow compilation adds 10-25ms latency, which is significant for small/fast models where GPU inference time is comparable. This overhead only occurs once at the start for video processing.
* **Execution engine latency**: The Workflows Execution Engine adds \~1-2ms nominal latency per frame.
* **Dynamic blocks**: Self-hosted deployments using dynamic blocks may experience +100-140ms latency per request due to pydantic model rebuilding. This is not applicable to the hosted platform or video processing where compilation happens once.
* **Model block dominance**: Typically \~95% of execution time is spent in model inference itself, which includes additional data transformations and metadata management.

## Performance Tips

* Workflow definitions are cached for 15 minutes by default. Use `use_cache=False` in `run_workflow()` to bypass caching if needed.
* For video processing, compilation overhead is negligible since it only occurs once at the beginning.
* Profile your workflows to identify whether bottlenecks are in model inference, data transformations, or workflow orchestration.