Re: [pve-devel] [PATCH manager] ui: file-restore: start the file-restore on the selected node

[Stefan Reiter <s.reiter@proxmox.com>]

On 6/18/21 5:17 PM, Thomas Lamprecht wrote:
> On 10.06.21 13:28, Dominik Csapak wrote:
>> On 6/10/21 11:29, Stefan Reiter wrote:
>>> On 6/10/21 10:40 AM, Dominik Csapak wrote:
>>>> On 6/10/21 10:23, Stefan Reiter wrote:
>>>>> On 6/10/21 9:37 AM, Dominik Csapak wrote:
>>>>>> and not the node where the browser connects.
>>>>>> there are at least two good reasons for this:
>>>>>> * it is confusing, since the user would expect it to start where
>>>>>>     the ui is pointint to
>>>>>> * the storage may not be available on the node the browser connects
>>>>>>     to, but it must be available on the node selected in the ui
>>>>>>
>>>>>> Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
>>>>>> ---
>>>>>
>>>>> Does this work with the proxying code for downloading files? I believe my original reasoning behind this was because my implementation for quickly forwarding data between the unprivileged and privileged daemons uses a local Unix socket. Thus forwarding between nodes (as I understand this will do?) would be subject to caching the data at the node being contacted, before sending it to the browser.
>>>>
>>>> well it does work, since i sucessfully downloaded some files from a backup
>>>>
>>>> what exactly do you mean with
>>>>   > would be subject to caching the data at the node being
>>>>   > contacted, before sending it to the browser.
>>>>
>>>
>>> Try downloading a larger file (multiple GB) - is it streamed directly to the browser or first transferred to node1 (from your example below), cached in memory there, and only once it's fully in RAM transferred to the client? I believe with this patch, the second variant will happen.
>>>
>>> This whole idea was the reason for these shenanigans:
>>> https://git.proxmox.com/?p=pve-http-server.git;a=commitdiff;h=51841e98fa5d4ad4d5b5250523c45f88769c577f
>>>
>>
>> from a short test it seem you're right
>> (have a pveproxy worker that sits at 2GiB RES atm)
>>
>> but the solution here is probably to drop the 'proxyto'
>> in the api call, and do the same as we do for
>> vnc proxying? ssh tunnel and starting the file
>> restore on the target node
> 
> @Stefan Any feedback regarding that proposal (or alternative fix)?
> 
> I mean, in the gui it may still make sense to use the local node if the storage is
> available there, as it normally means less total traffic produced, but if it's not
> configured as available there that's just not an option.
> 

I talked with Dominik and we agreed that his idea is the right one. We left it at
"one of us will implement it", since we deemed it not that big of an issue for right
now, but I can take a look sooner as well if needed.