The recommendation-api service provides endpoints that we use for getting candidates for Suggested Edits - i.e. articles with missing descriptions, and Commons images with missing captions.
The thing is, though, the endpoints provided by this service are merely proxies for regular MW API calls that we might as well make ourselves from inside our client. The original intention of the service was to leave open the possibility of changing/improving the algorithm for suggestions, transparently to the consumer. However, this never really happened, and now this hypothetical benefit is being outweighed by the maintenance cost of this service.
This task is to measure the cost of moving these API calls into the client, with the potential goal of sunsetting this service:
- Is there a significant increase in latency when the client makes the MW calls directly, instead of this service call?
- Is there prohibitive complexity in moving and maintaining these MW calls locally in the client?