Logging has several problems:

• We have no logs before a VM is created and the agent is running
• We cannot associate the log parts we send with the work item they are for
  • if we did, we could raise an alarm when nothing happens for a work item for a long time
• We have mangled and mixed logs (due to the retrying and messed up gzipping)
• We should log all network errors to find patterns (send them to influxdb)

I think this warrants re-visiting the topic.