Regarding inquiry by Marica Antonacci on Wed, 18 Nov 2015 04:43:47 -0800
Failover scenarios are anticipated in the DVDI module and are being tracked as
this issue on github
https://github.com/emccode/mesos-module-dvdi/issues/23
The resolution of this issue requires that the underlying volume driver
supports a "pre-emptive mount" whereby any prior mounts are terminated
automatically by a mount from another node.
Many but not all storage providers are capable of supporting this behavior.
The intention is that DVDI will handle scenarios including:
1. Task crashes on an agent node and is rescheduled on a different node. Note
in this case, if the slave process remains running, it is very likely that the
dvdi module will remove the mount automatically now. However this might not
occur BEFORE the task is rescheduled on another cluster node, so the upcoming
pre-emptive mount feature could be useful in this scenario.
2. Entire Agent node crashes while task is running and does not restart or
return. The agent machine is a smoking piece of molten metal. This will REQUIRE
the new dvdi feature and support from the underlying storage driver.
3. Agent crashes and restarts. Similar to scenario #1 but perhaps the unmounts
takes longer. This scenario should result in removal of the original mount upon
restart, and is implemented now. If the Agent restart takes a long time, the
roadmap pre-emptive mount feature will be of value.
As a related note, some storage providers and some filesystems actually allow
use of concurrent mounts from multiple cluster nodes, although this is
uncommon. The architecture of the dvdi module is intended to support this,
although no volume driver available today enable the functionality.
In summary pre-emptive mount is a roadmap feature for the dvdi module. I expect
it to be delivered shortly, after the feature is enabled in the rexray driver.
My best guess is that this will not occur within the next two weeks, but it may
occur within the next two months.
Steve Wong
Contributor to the Mesos dvdi module project Developer Advocate EMC{code} - the
open source advocacy group within EMC
CODE OPEN, DEPLOY EVERYWHERE
https://github.com/emccode
blog.emccode.com
1.562.417.7048
[email protected]
@cantbewong
