Arvados

Currently, if a container cannot be started due to some infrastructure problem (whether or not it's related to the specific container) it will be retried repeatedly forever.

Proposed solution:

Add a site config knob (analogous to num_retries) that limits the number of times a container can be unlocked (moved from Locked to Queued state) before being automatically cancelled.

• Config key max_container_dispatch_attempts (default 5)
• DB column "lock_count" (do not include in API response)
• Increment lock_count during lock()
• When unlocking a container, if lock_count >= Rails.configuration.max_container_dispatch_attempts, change state to Cancelled instead of Queued (the unlock API should still respond 200 in this case) and update runtime_status[error] with an error message.

Write tests and update documentation.