One thing that stands out from the logs is that there are multiple concurrent requests hitting the same revisions. For example, there are 107 requests from the same rev_id: 41731385. Each of these requests fetching from MW API. A caching mechanism would be a great solution for this issue to reduce redundant API calls and lower the risk of timeouts under load.

In T403088#11125140, @kevinbazira wrote:

One thing that stands out from the logs is that there are multiple concurrent requests hitting the same revisions. For example, there are 107 requests from the same rev_id: 41731385. Each of these requests fetching from MW API. A caching mechanism would be a great solution for this issue to reduce redundant API calls and lower the risk of timeouts under load.

Hey @kevinbazira thnx for reporting this issue.
If I understand correctly you suggest to implement a caching logic when we are hitting the MWAPI from lifting right ?
So, the caching logic will be over the get_extractor() ?
Or you meant something different ?

In T403088#11245417, @gkyziridis wrote:

In T403088#11125140, @kevinbazira wrote:

One thing that stands out from the logs is that there are multiple concurrent requests hitting the same revisions. For example, there are 107 requests from the same rev_id: 41731385. Each of these requests fetching from MW API. A caching mechanism would be a great solution for this issue to reduce redundant API calls and lower the risk of timeouts under load.

Hey @kevinbazira thnx for reporting this issue.
If I understand correctly you suggest to implement a caching logic when we are hitting the MWAPI from lifting right ?
So, the caching logic will be over the get_extractor() ?
Or you meant something different ?

@gkyziridis thank you for looking into this. The caching idea is to avoid redundant preprocessing: if a rev_id already has a cached prediction (with the same model version), we would return this prediction directly without re-fetching features from MW API. Since the errors are intermittent and revscoring model-servers are planned for decommissioning, it is probably best to discuss with the ML team before prioritizing this change.