Media Model Proxy provides a proxy service to the DeepBird Prediction services operated by Media Understanding.
The service intermediates with Blobstore to load the media requested and can issue inference requests to multiple models simultaneously. This is useful to reduce Blobstore and network usage for clients. Clients themselves can pass only lightweight metadata (model flags, media type, Blobstore path, etc.).
Media Model Proxy does not publish/store model scores for consumption, instead it returns a map of the requested models to the observed responses to the caller. The caller is responsible with interpreting and persisting the results.
Current clients are Media Analysis Service. Periscope issues requests through a custom thrift endpoint in Media Analysis Service.
(This service was originally Cortex Media Annotator. Design Documentation and service notes regarding Cortex Media Annotator are potentially also of use.)
- Runbook: the internal runbook
Media Model Proxy source code is at media-understanding/media-model-proxy in Source.
Contents:
.. toctree:: :maxdepth: 3 using service
