Add support for storage controllers

[tintoy]

The Q release (v2.5) of the CloudControl API has introduced support for explicitly controlling the storage controllers in a deployed server.

Because the behaviour of disk management when using custom storage controllers diverges from the existing behaviour, storage controllers and their disks will be managed via a separate ddcloud_storage_controller resource type:

resource "ddcloud_server" {
  name = "My server"

  // Disks can either be specified here (order now matters)
  disk {
    // SCSI unit Id can no longer be specified here
    size_gb = 10
    speed   = "STANDARD"
  }
}

resource "ddcloud_storage_controller" {
  server = "${ddcloud_server.my_server.id}"
  adapter_type = "LSI_LOGIC_PARALLEL"
  bus_number = 1

  // Or disks can be specified here (order does not matter)
  disk {
    scsi_unit_id = 0 // SCSI unit Id can now be specified here
    size_gb       = 10
    speed         = "STANDARD"
  }
}

For a given ddcloud_server, you can either specify disks the old way (inline, but without a SCSI unit Id) or with an explicit storage controller (but not both). Since disks specified the old way no longer have an explicit SCSI unit Id, their order now matters (i.e. you cannot have 3 disks and remove disk 2). Disks specified via the new way (with explicit storage controller) have an explicit SCSI unit Id and their order there does not matter.

These breaking API changes are expected to also require refactoring of the existing ddcloud_server resource implementation due to changes in the format of information returned by v2.5 of the CloudControl API.

Note that it's theoretically possible to to support specifying disks and controllers in the same ddcloud_server resource but this would greatly complicate the provider implementation for this resource type (disks with a value for controller_id would have to be processed after any controller entries, and Terraform may no longer be able to automatically warn the user about changes that force a new resource or are otherwise invalid).

See below for an alternative design proposal.

See also:

• Add support for storage controllers go-dd-cloud-compute#12
• Additional NIC support for servers #33
• Specify network adapter type for primary NIC and also additional NICs #52
• Merge ddcloud_server_nic into ddcloud_server #56

[tintoy]

CC: @kumarappanc @mingsheng36 @wninobla @jeffstoner @johnamurray @alexbacchin @manishkp @BenObermayer

Your input would be appreciated, both regarding this design and the trade-offs it involves.

[tintoy]

A possible alternative for the design might be to rely on the fact that disks which don't explicitly specify a controller Id are always created against controller 0 (the default controller).

resource "ddcloud_server" "my_server" {
	name = "My Server"

	// There's always an implicit storage_controller with bus = 0

	storage_controller {
		bus          = 1
		adapter_type = "BUS_LOGIC"
	}

	// Old-style disk (implicitly targets controller 0)
	disk {
		scsi_unit_id = 0
		size_gb      = 10
		speed		 = "STANDARD"
	}

	// New-style disk (explicitly targets controller 1)
	disk {
		bus          = 1
		scsi_unit_id = 0
		size_gb      = 10
		speed		 = "STANDARD"
	}
}

In this model, the provider always uses the CloudControl APIs for disks to target a specific controller (rather than the old-style APIs), but disk entries that don't specify a controller are assumed to target controller 0 (i.e. the bus property has a default value of 0).

Note that this approach may add some complexity to the implementation because disk management logic for updating a ddcloud_server resource would then have to be done in several passes:

1. Examine storage_controller entries and add controllers if required.
2. Examine disk entries, and add / remove / modify if required.
3. Examine storage controller entries and remove controllers if required.

In keeping with our design ethos, the provider will not attempt to detect an attempt to remove a controller with disks still attached (instead, responsibility for rejecting this change will be delegated to the CloudControl API).

[jeffstoner]

Given what's coming in the R release (support for IDE/SATA,) there may be another jumbling of the storage subsystem metadata. The design phase of this is currently on-going, so I can't comment on exactly what changes are happening (because they're very much in-flux) but we should be careful with the modeling - might want to consider using "disk type" structures that cater to the bus they're attached to (IDE disks won't have a scsi_unit_id, etc.)

[jeffstoner]

There are (currently) 4 storage adapters supported by the platform:

There are four adapters supported by vSphere:

• BusLogic Parallel - This was one of the first two emulated vSCSI controllers made available in the VMware platform, and remains commonly used on older version of Windows as the driver is available by default.
• LSI Logic Parallel - This was the other emulated vSCSI controller made available in the VMware platform, and remains commonly used on UNIX Operating Systems 
• LSI Logic SAS - This is a newer evolution of the parallel driver supported (and in some cases required) for newer versions of Microsoft Windows. It was designed to replace the BusLogic Parallel adapter and provides better performance than the original BusLogic Parallel adapter.
• VMware Paravirtual - This is a VMware-specific driver that is virtualization-aware and designed to support very high throughput with minimal processing cost and is therefore the most efficient driver. However, this driver requires that VM Tools be installed and running in order to function properly. There are some other restrictions, particularly on older operating systems - review the appropriate VMware documentation for more details. 

(source: https://docs.mcp-services.net/pages/viewpage.action;jsessionid=E61080748C1D3F34342D36D225289A95?pageId=3015251)

[jeffstoner]

Your statement

disks which don't explicitly specify a controller Id are always created against controller 0 (the default controller)

conflicts with the defined MCP behavior

If you do not manually select the SCSI Controller and position of the disk, it will be added sequentially to the next available SCSI controller and position.

(source: https://docs.mcp-services.net/display/CCD/How+to+Add+Additional+Local+Storage+%28Disk%29+to+a+Cloud+Server - Prerequisite 5.)

[tintoy]

Thanks for the feedback @jeffstoner - this is all useful information :)

Unfortunately it sounds like these changes may make the provider quite painful to use, so I may hold off on building anything until we can find a design compromise that fits. I'll keep thinking about it.

[tintoy]

conflicts with the defined MCP behavior

On the other hand, this is Terraform behaviour we're talking about, not CloudControl. As long as the behaviour is well-documented, I think I'd still be comfortable with it.

As for SATA controllers, we could add a sata_disk type:

resource "ddcloud_server" "my_server" {
  name = "My Server"

  // SCSI disk, bus 0
  disk {
    scsi_unit_id = 0
    size_gb      = 10
  }

  // SCSI disk, bus 1
  disk {
    bus          = 1
    scsi_unit_id = 0
    size_gb      = 100
  }

  // SATA disk
  sata_disk {
    size_gb = 50
  }
}

[wninobla]

Just my two cents worth… Since we treat the no guest OS customization differently in terms of deployment vs. on that does guest OS customization maybe we just keep it separate. In other words, maybe you have a different method that handles ddcloud_server vs. say one that’s dedicated to storage say ddcloud_storage? Not sure if that makes it more convoluted but we kind of already do that for NIC’s and anti-affinity. Deleting/changing around post deployment may be easier although I’m not sure what we have to adhere to API wise (e.g. VM’s currently require to always have one disk connected). Again just some thoughts here. If I’m off-base please ignore as it’s Monday… =)

[tintoy]

(I'm trying to be mindful of backward-compatibility here)

[tintoy]

Thanks @wninobla - that does make sense. I'm favouring option 2 at the moment as it's a little more user-friendly, but I'm open to suggestions :)

[wninobla]

Works for me. It’ll keep the existing server.tf files consistent if they break them out vs. doing it as one big monolithic build. And yeah I agree on the behavior piece. We’ll just need to flag it as one that we can do during a build vs. after the fact.

[wninobla]

I’d say deleting some snips vs. adding new one’s is 50/50. Just depends on how complex someone has gotten with their TF files. Then again , most of those folks code for a living so it should be more of an annoyance than a blocker so…

[alexbacchin]

Same here, backwards compatibility is a must. These storage adapters changes makes practically no difference for 90% of the users as the underlying physical storage infrastructure still the same.