After the latest load testing on the article-descriptions model server we discovered a discrepancy between in the preprocessing phase when the service is run on Cloud VPS vs when it runs on Lift Wing.
After further investigation done by @kevinbazira we found out that the model server behaves differently when ran from localhost or ml-sandbox. The difference is in the way that we access Rest API. Local runs use RESTBase while Lift Wing is uses REST Gateway.
Taking into consideration that RESTBase is in the process of deprecation, the correct way to access the API is through the Gateway.
In this task we'll investigate the issue to find the root cause of this and answer the following questions:
- Check whether the issue is caused by misconfiguration on the istio side
- Check whether the issue is caused by a wrong request (e.g. wrong header etc)
- The model server makes multiple requests to the rest endpoint. Given that RESTBase has a cache while the Gateway doesn't, is the total increased latency validated because of the cumulative latency of the underlying requests? If so, how could we overcome this?
- If none of the above yield a solution, investigate what is the deprecation timeline for RESTBase and check if we could use it on the LW service.