When a namespace is deleted all its associated resources (pods, svc, rc, ...) are deleted too. This creates a race when pooling is enabled as when pods are deleted its associated ports are marked to be reused and put into the recyclable_ports dict.
The problem arises as this dict of ports to be reused is checked periodically (15 second interval by default) and the function to remove all the ports associated to a given network can be executed at the same time, therefore calling function trigger_return_to_pool at the same time, which may lead to skipping released ports. Thus making network deletion to fail and consequently the associated kuryrnet CRD.
In addition, if several namespaces are deleted concurrently, the handler will call delete_network_pools for each namespace subnet, which leads to several calls to _trigger_return_to_pool. The second call will end up reading ports from the copy of recyclable_ports to available_ports, leading to duplicates port_ids.
Fix proposed to branch: master
Review: https:/