The single major source of integration flake tests at Loculus is that LAPIS has a habit of terminating/dropping/aborting connections mid-response, after it initially returns HTTP 200.
This happens much more in integration tests than in stable production environments - which makes me think that it might be related to data version switches - which happen frequently in a test setting but not so often in production.
Ideally, LAPIS would successfully complete existing requests with the old data version and not cut connections off when new data arrives.
The single major source of integration flake tests at Loculus is that LAPIS has a habit of terminating/dropping/aborting connections mid-response, after it initially returns HTTP 200.
This happens much more in integration tests than in stable production environments - which makes me think that it might be related to data version switches - which happen frequently in a test setting but not so often in production.
Ideally, LAPIS would successfully complete existing requests with the old data version and not cut connections off when new data arrives.