bd46075acd
We've seen some issues related to the fact that on CNI DEL we wait for pod annotation to appear in CNI registry: 1. Overloaded kuryr-daemon HTTP server, because many stale DEL requests from CRI will saturate the number of free connections, causing CNI to answer requests extremely slowly. 2. Some CRI's will not delete network namespace before getting a successful CNI DEL. If we'll assign the same IP (or vlan) to a next pod, we might end up with an IP or vlan conflict and NetlinkError. This commit makes sure we only wait for VIFs up to 5 seconds on DEL requests, enough time for watcher to populate registry on restart. Also some code got moved around to make implementating the above simpler and some logs got clarified to make debugging of such issues easier later on. Change-Id: I9221b6bc9166597837a4b53382862aa6c6f3e94c Closes-Bug: 1882083 Related-Bug: 1854928 |
||
---|---|---|
.. | ||
__init__.py | ||
service.py |