Problem

We started storing image detection results to history raw storage (Avro) and re-using it on history collections. Such a solution allows to optimize usage of IR resources.

But it implies that we might provide image detection of sub-optimal quality on history collections:

• If a client buys a logo and collects history, he might not get logo detections if we were not able to detect that logo back when the image was analyzed.
• If a model is updated significantly, e.g. set of possible values for a certain image attribute is changed, some mentions collected via history will not reflect that update.

This was different before we introduced re-using of detections- history collections to topics with VI enabled got results from the latest IR version.

Ideas

1. Store date of on when detection was stored - such way we will be able to introduce logic later to determine if the result is too stale to be reused.
2. If we go to IR API for mention collected from history (i.e. detection was missing in Avro or stale), we store it to History
3. Have endpoint on IR API to just detect logos to mitigate scenario with missing logos (?? Should we bother? What are the computational resources required compared to the full detection?)