[MESOS-2546] Mesos 0.20.1 causes framework starvation on single node clusters when using Chronos and Marathon

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 0.21.1
• Component/s: framework
• Assignee: Unassigned
• Reporter: Sunil Shah
• Created: 25/Mar/15 17:44
• Updated: 09/May/18 00:56
• Resolved: 09/May/18 00:56

Description

Tracking an issue raised by Chronos users that appears to be a regression in Mesos: https://github.com/mesos/chronos/issues/381#issuecomment-83647539

1) Chronos's interval between refusing offers and receiving the next one is at 0.1 seconds to allow finer grained scheduling of jobs.

2) On single node clusters running Mesos 0.20.1 with both Chronos and Marathon installed, Marathon did not receive any offers. On multi-node clusters, this behaviour was not observed. This behaviour was not observed when using previous versions of Mesos.

3) Changing this interval back to the default value (i.e., by not setting it) fixed this problem. (See commit.)

This can be replicated using an installation of playa-mesos and running both the latest Mesosphere packages of Chronos and Marathon.

Issue Links

• duplicates: MESOS-3202 Avoid role/framework offer starvation in DRF allocator.