[MESOS-8350] Resource provider-capable agents not correctly synchronizing checkpointed agent resources on reregistration

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 1.5.0, 1.6.0
• Component/s: master
• Target Version/s: 1.5.1
• Epic Link: Resource Provider and CSI Tech Debt
• Sprint: Mesosphere Sprint 71
• Story Points: 2
• Assignee: Benjamin Bannier
• Reporter: Benjamin Bannier
• Shepherd: Jie Yu
• Created: 20/Dec/17 14:19
• Updated: 14/Jun/18 20:06
• Resolved: 04/Jan/18 12:25

Description

For resource provider-capable agents the master does not re-send checkpointed resources on agent reregistration; instead the checkpointed resources sent as part of the ReregisterSlaveMessage should be used.

This is not what happens in reality. If e.g., checkpointing of an offer operation fails and the agent fails over the checkpointed resources would, as expected, not be reflected in the agent, but would still be assumed in the master.

A workaround is to fail over the master which would lead to the newly elected master bootstrapping agent state from ReregisterSlaveMessage.