cloudstack-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <>
Subject [jira] [Commented] (CLOUDSTACK-9003) Make VM naming services injectable and in their own module
Date Thu, 14 Apr 2016 10:47:25 GMT


ASF GitHub Bot commented on CLOUDSTACK-9003:

Github user ProjectMoon commented on the pull request:
    > @ProjectMoon correct resource naming is critical to the proper operation of the management
server. We have had a significant bugs and production issues caused by subtle changes to resource
naming strategies between releases where CloudStack suddenly can't find a resource on the
device it is attempting to control. 
    > Have you performed any upgrade testing for this PR? If so, what tests have you performed
in which configurations?
    We have not specifically performed any upgrade testing. Our current stable version is
based on 4.7.1, and essentially our "upgrade testing" has consisted of deploying 4.7.1 before
and after our development of this feature is complete. The configuration has been tested with
KVM, VmWare, and the simulator. 
    > Also, could you please add an FS to the wiki and start a conversation on dev@? Given
the importance of resource naming, it would be extremely helpful to have an explanation of
the design and its operation.
    As far as I know, I don't have access to the wiki. Otherwise I would add to it. Do I need
to register a separate account or can I use my Apache JIRA account?
    > @ProjectMoon we'll need more information on why you're doing this, why we should
have it, what is fixes and will it guarantee backward resource-name compatibility (for example,
vmware vms have this strictly tie up with internal ACS vm name, such config are set in each
vmware's VM's annotations) and upgrade paths
    * **Why we are doing this**: we implement our own naming scheme for the supported resources.
On our 4.2 branch we hacked this in, but now we want to present a framework that we can extend,
and open the possibilities to other.
    * **Why should mainline have this**: More flexibility for developers, easier testing (static
classes notoriously cause problems), a unified way to generate names (DRY principle).
    * **What it fixes**: It doesn't _fix_ anything per se, but the refactoring helps move
us towards a cleaner codebase.
    * **Backwards compatibility**: The default plugin generates names and UUIDs in the exact
same way as before, so yes.
    VmWare is an interesting case though. Can you describe in more detail/point me to where
this stuff happens so I can verify that custom naming plugins will not break it?

> Make VM naming services injectable and in their own module
> ----------------------------------------------------------
>                 Key: CLOUDSTACK-9003
>                 URL:
>             Project: CloudStack
>          Issue Type: Improvement
>      Security Level: Public(Anyone can view this level - this is the default.) 
>            Reporter: Jeffrey Hair
>            Priority: Minor
> Proposal: Make the various classes/code that give VMs and other resources hostnames,
hypervisor guest names, and UUIDs into their classes as injectable dependencies in their own
module under the core or backend module.
> This proposal originally only concerned the VirtualMachineName class and can be broken
down into several parts:
> * Make the VirtualMachineName class an injectable dependency instead of being full of
static methods.
> * Refactor the generateHostName method in UserVmManagerImpl to be backed by an injectable
service which generates host names.
> * Move the UUIDManagerImpl from the core module to a new module (grouped with the other
2 ideally).
> Rationale:
> * VirtualMachineName is one of the few remaining classes that has static methods tangled
like spaghetti throughout the code. This change will put it in line with the rest of the management
server codebase and opens the door to extensibility. Which brings us to...
> * Extensibility: The ultimate goal of this feature is to provide 3rd party developers
the option of changing default instance/resource naming policies. Currently this is possible
in a very limited fashion with the global setting, but this proposal makes it
much more extensible.
> By moving the naming-related services (VirtualMachineName, UUIDManager, and more as added/discovered)
to their own module, the module can be excluded by and different ones substituted
in. Alternatively, it could use the adapter model that other classes use, and the user can
configure which adapters are active and also provide custom ones.
> A good use case for this functionality is using a different style naming to emulate other
cloud providers such as AWS (i-abc123) or GCE. 

This message was sent by Atlassian JIRA

View raw message