Re: [pve-devel] applied: [PATCH manager] Allow setting device class on osd create

[Thomas Lamprecht <t.lamprecht@proxmox.com>]

Am 7/24/20 um 2:24 PM schrieb Alwin Antreich:
> On Fri, Jul 24, 2020 at 11:54:10AM +0200, Thomas Lamprecht wrote:
>> Am 7/24/20 um 11:46 AM schrieb Alwin Antreich:
>>> On Fri, Jul 24, 2020 at 11:34:33AM +0200, Thomas Lamprecht wrote:
>>>> Am 7/23/20 um 3:25 PM schrieb Alwin Antreich:
>>>>> In some situations Ceph's auto-detection doesn't recognize the device
>>>>> class correctly. The option allows to set it directly on osd create,
>>>>> instead of altering it afterwards. This way the cluster doesn't need to
>>>>> shift data back and forth unnecessarily.
>>>>>
>>>>> Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
>>>>> ---
>>>>>  PVE/API2/Ceph/OSD.pm | 7 +++++++
>>>>>  1 file changed, 7 insertions(+)
>>>>>
>>>>
>>>> applied, thanks - comments still inline
>>>>
>>>>> diff --git a/PVE/API2/Ceph/OSD.pm b/PVE/API2/Ceph/OSD.pm
>>>>> index ceaed129..f1f39bf9 100644
>>>>> --- a/PVE/API2/Ceph/OSD.pm
>>>>> +++ b/PVE/API2/Ceph/OSD.pm
>>>>> @@ -260,6 +260,11 @@ __PACKAGE__->register_method ({
>>>>>  		default => 0,
>>>>>  		description => "Enables encryption of the OSD."
>>>>>  	    },
>>>>> +	    'crush-device-class' => {
>>>>> +		optional => 1,
>>>>> +		type => 'string',
>>>>> +		description => "Set the device class of the OSD in crush."
>>>>> +	    },
>>>>
>>>> why not having an enum with 'nvme', 'ssd', and 'hdd' here?
>>> Ceph allows the class to be an arbitrary string, eg. my-very-fast-disk.
>>>
>>
>> Is it then "auto-generated" or has ceph an index of known ones floating
>> around?
> AFAICT, Ceph check /sys/block and gathers the information from there and
> then decides if it is hdd/ssd/nvme.

that's not what I meant, talking about existing classes ;-)

> 
>>
>> We could also add this to the UI, to advanced as editable KVCombobox which is
>> emptyText "auto", and has "hdd", "nvme" and "ssd" as convenience selectors.
>> If we can get all currently available ones also (relatively) cheaply we could
>> add them too.

> I think only the active ones. But I guess it wouldn't hurt if we just
> hard-code the usual suspects + auto. :)
 
As said, if it's editable one can actually enter all information, the provided
ones are just for convenience, check the replication schedule field for what I
mean.