Service migration and state management

[ramwan]

Issue to track research into service migration and cases that will arise when services are moved. Initial research into Mobile IP and link/network layer hand-off protocols don't solve the fundamental problem of maintaining state in one service instance and transferring that across to a new instance of the service.

Some cases to consider

• 'stateless' transfers
  • application itself may have no state which means a new service instance could be spun up without issue
  • 'stateless' application could run over TCP which requires state
• Mitigating loss in unexpected failures (eg. power outage)
• transfers across physical machines or spinning up a new instance on the same physical machine may mean state can be maintained

[ramwan]

Initial state migration experiment writeup @ https://github.com/MatrixAI/Relay/blob/master/experiments/stateless_container_ping_migrations/README.txt

[CMCDragonkai]

Something to study: VM Snapshotting, Orthorgonal Persistence and CRIU. Leave it to later though, (we need Vivian and potentially the new hire).

[ramwan]

Stateless connection migration is simple and can be done with iptables nat rules to swap out both src and dst addresses in a packet with the conntrack module being able to keep track of the connection. Migrating TCP however doesn't seem possible for now and further research is being done to see how we can migrate TCP.