brooklyn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aled Sage <>
Subject Re: Proposal: Add appId optional paramater to deploy api
Date Tue, 25 Jul 2017 16:44:35 GMT
Hi all,

I've been exploring adding support for `&deploymentUid=...` - please see 
my work-in-progress PR [1].

Do people think that is a better or worse direction than supporting 
`&appId=...` (which would likely be simpler code, but exposes the 
Brooklyn internals more).

For `&appId=...`, we could either revert [2] (so we could set the id in 
the EntitySpec), or we could inject it via a different (i.e. add a new) 
internal way so that it isn't exposedon our Java api classes.





On 07/07/2017 18:28, Aled Sage wrote:
> Hi,
> Taking a step back to justify why this kind of thing is really 
> important...
> This has come up because we want to call Brooklyn in a robust way from 
> another system, and to handle a whole load of failure scenarios (e.g. 
> that Brooklyn is temporarily down, connection fails at some point 
> during the communication, the client in the other system goes down and 
> another instance tries to pick up where it left off, etc).
> Those kind of thing becomes much easier if you can make certain 
> assumptions such as an API call being idempotent, or it guaranteeing 
> to fail with a given error if that exact request has already been 
> accepted.
> ---
> I much prefer the semantics of the call failing (with a meaningful 
> error) if the app already exists - that will make retry a lot easier 
> to do safely.
> As for config keys on the app, in Duncan's use-case he'd much prefer 
> to not mess with the user's YAML (e.g. to inject another config key 
> before passing it to Brooklyn). It would be simpler in his case to 
> supply in the url `?appId=...` or `?deploymentId=...`.
> For using `deploymentId`, we could but that feels like more work. We'd 
> want create a lookup of applications indexed by `deploymentId` as well 
> as `appId`, and to fail if it already exists. Also, what if someone 
> also defines a config key called `deploymentId` - would that be 
> forbidden? Or would we name-space the config key with 
> `org.apache.brooklyn.deploymentId`? Even with those concerns, I could 
> be persuaded of the `org.apache.brooklyn.deploymentId` approach.
> For "/application's ID is not meant to be user-supplied/", that has 
> historically been the case but why should we stick to that? What 
> matters is that the appId is definitely unique. That will be checked 
> when creating the application entity. We could also include a regex 
> check on the supplied id to make sure it looks reasonable (in case 
> someone is already relying on app ids in weird ways like for filename 
> generations, which would lead to a risk of script injection).
> Aled
> On 07/07/2017 17:38, Svetoslav Neykov wrote:
>> Hi Duncan,
>> I've solved this problem before by adding a caller generated config 
>> key on the app (now it's also possible to tag them), then iterating 
>> over the deployed apps, looking for the key.
>> An alternative which I'd like to mention is creating an async deploy 
>> operation which immediately returns an ID generated by Brooklyn. 
>> There's still a window where the client connection could fail though, 
>> however small it is, so it doesn't fully solve your use case.
>> Your use case sounds reasonable so agree a solution to it would be 
>> nice to have.
>> Svet.
>>> On 7.07.2017 г., at 18:33, Duncan Grant 
>>> <> wrote:
>>> I'd like to propose adding an appId parameter to the deploy 
>>> endpoint.  This
>>> would be optional and would presumably reject any attempt to start a 
>>> second
>>> app with the same id.  If set the appId would obviously be used in 
>>> place of
>>> the generated id.
>>> This proposal would be of use in scripting deployments in a distributed
>>> environment where deployment is not the first step in a number of
>>> asynchronous jobs and would give us a way of "connecting" those jobs 
>>> up.
>>> Hopefully it will help a lot in making things more robust for 
>>> end-users.
>>> Currently, if the client’s connection to the Brooklyn server fails 
>>> while
>>> waiting for a response, it’s impossible to tell if the app was 
>>> provisioned
>>> (e.g. how can you tell the difference between a likely-looking app, and
>>> another one deployed with an identical blueprint?). This would make 
>>> it safe
>>> to either retry the deploy request, or to query for the app with the
>>> expected id to see if it exists.
>>> Initially I'm hoping to use this in a downstream project but I think 
>>> this
>>> would be useful to others.
>>> If no one has objections I'll aim to implement this over the next 
>>> couple of
>>> weeks.  On the other hand I'm totally open to suggestions of a better
>>> approach.
>>> Thanks
>>> Duncan Grant

View raw message