MinT supports machine translation for different products. In the initial integrations of MinT to support Content Translation and the Localization infrastructure, the requested contents to translate are quite unique. That is, an editor creating a new translation for a Wikipedia article will generate requests for MinT to translate the different sentences of such article. which will be very different from those translated by another editor working on a different article. In this context, there was no need for caching.

However, as MinT is applied in contexts where multiple users may be reading the same translated content, there may be a bigger benefit in applying caching to reduce waiting times and the general workload for the service. This may be relevant for context such as the translation of wishes (T363306) for the new Wishlist process and MinT for Wikipedia readers (T359072).

This ticket proposes to consider ways in which the MinT service could use caching to provide a better performance in the described contexts. Given a certain request text, language pairs and models, the result should come from the caching system if the same exact request was made previously/recently.
We need to consider that (a) language models are not updated frequently, but we also plan to allow users to provide community-verified translations (T351748) and the caching system should not get in the way of users being able to correct a bad translation (i.e. taking time to see changes reflected).