[
https://issues.apache.org/jira/browse/MESOS-7607?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16047025#comment-16047025
]
Neil Conway commented on MESOS-7607:
------------------------------------
[~drcrallen]: Hi Charles. The argument for using a first-class message rather
than agent attributes is as follows:
* We want to assign special semantics to fault domains. For example, the
allocator will treat agents in remote regions differently from agents in the
local region. Similarly, we might allow agents to transition from the unset
domain to a configured domain without requiring an agent drain. If we used
agent attributes, we'd need to special-case attributes with certain names and
treat them differently. This is doable but seems against the spirit of
user-defined key-value attributes.
* In the future, we might want to support more kinds of domains (e.g., latency
domains, fault domains, upgrade domains). The upgrade / forward-compatibility
story is more flexible if we allow fault domains to be first-class protobuf
messages, rather than just a collection of key-value pairs.
An alternative design would have been to introduce a notion of "system
attributes" with special semantics, and then make fault domains a kind of
system attribute. We considered this, and it would work, but in the end we
thought first-class domains was more flexible.
> Support for first-class fault domains.
> --------------------------------------
>
> Key: MESOS-7607
> URL: https://issues.apache.org/jira/browse/MESOS-7607
> Project: Mesos
> Issue Type: Epic
> Reporter: Neil Conway
> Assignee: Neil Conway
> Labels: mesosphere
>
> Mesos should support a first-class notion of "fault domains", which
> effectively provide a common vocabulary for describing the region and zone
> where a node (either master or agent) is located.
> Design doc:
> https://drive.google.com/open?id=1gEugdkLRbBsqsiFv3urRPRNrHwUC-i1HwfFfHR_MvC8
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)