+1 on converting AbstractAWSProcessor to using AWSCredentialsProvider, and
+1 to adding a controller service for AWS creds. I would like to help.
The current AbstractAWSProcessor.onScheduled method already contains code
to probe for the REGION PropertyDescriptor and modify the AWS client based
on that. Could we use a similar strategy for the controller service and the
upgrade to using createClient with AWSCredentialProvider?
if (getSupportedPropertyDescriptors().contains(REGION)) { ...
My mental sketch of how this might work:
1.) AbstractAWSProcessor declares a PropertyDescriptor for the controller
service, AWS_CREDENTIAL_CONTROLLER_SERVICE, or an equally creative name
2.) The AbstractAWSProcessor.onScheduled method would probe for the
existence of AWS_CREDENTIAL_CONTROLLER_SERVICE.
2.a.) If found, onScheduled would use the new createClient(...,
AWSCredentialsProvider, ...) method, pretty much as Mans Singh describes it
2.b.) If not found, the existing createClient(..., AWSCredentials, ...)
would be used as it is now
3.) As part of the roll out of the controller service, we would modify the
various bundled AWS processors to opt-in via the
AWS_CREDENTIAL_CONTROLLER_SERVICE property.
I believe this would allow unknown custom subclasses of
AbstractAWSProcessor existing in the wild to enjoy the current contract,
while allowing us to roll out a controller service for the bundled
processors.
On the down side, this may be an unintuitive side-effect behavior. There
are very sound reasons to refactor AbstractAWSProcessor to use
AWSCredentialProvider that are not connected to the controller service as
Man Singh has discovered. We could also add a method for subclasses to
explicitly declare their credential strategy, and provide a concrete
default in AbstractAWSProcessor maintaining the existing createClient
behavior.
On Sat, Jan 9, 2016 at 11:29 AM, Aldrin Piri <[email protected]> wrote:
> Mans,
>
> Fair points concerning the duplication. I was thinking that in conjunction
> with marking that method deprecated we would also drop the abstract
> classifier and require implementers subclassing the original class to
> provide the override explicitly. It's not ideal, but does alleviate the
> issues concerning excess methods in the interface. Sorry for omission of
> what is certainly a very valid issue.
>
> Outside of that, the items you are establishing sounds like the right
> path. I hashed this out a little more fully to better express my ideas
> [1].
>
> [1] https://gist.github.com/apiri/6a17b71e261f457daecc
>
> On Sat, Jan 9, 2016 at 1:17 PM, M Singh <[email protected]>
> wrote:
>
> > Hi Aldrin:
> > Even if we subclass AbstractAWSProcessor and overwrite the onScheduled
> > method, we still have to add (rather then replace createClient with aws
> > creds argument) a createClient method that would take the credential
> > provider argument rather than the aws credentials argument (the current
> > implementation).
> > current nifi aws createClient (with aws credentials)
> > protected abstract ClientType createClient(final ProcessContext
> > context, final AWSCredentials credentials, final ClientConfiguration
> > config);
> > new nifi aws createClient (with aws credentials provider)
> > protected abstract ClientType createClient(final ProcessContext
> > context, final AWSCredentialsProvider credentialsProvider, final
> > ClientConfiguration config);
> > Regarding overwriting onScheduled method. Here is snippet of current
> > AbstractAWSProcessor.onScheduled method (using aws credentials):
> > @OnScheduled public void onScheduled(final ProcessContext context)
> > { final ClientType awsClient = createClient(context,
> > getCredentials(context), createConfiguration(context));
> > this.client = awsClient;...
> > So, in the subclass AbstractAWSProcessor.onScheduled method, we will
> check
> > if controller is available and if so, call the create client with the
> > credentials provider method. In this case each of the nifi aws processors
> > (currently s3, sns, and sqs) will have to provide two implementation of
> > create client (one with aws creds - the current one, and one with aws
> creds
> > provider).
> > I might be missing something, but it looks like there will duplication of
> > amazon client creation (one using creds and one using creds provider from
> > the controller) along with two createClient method in Nifi's
> > AbstractAWSProcessor which might causing confusion. But that is just my
> > thought.
> > Let me know what you think.
> >
> > Thanks again.
> > Mans
> >
> >
> >
> > On Saturday, January 9, 2016 9:20 AM, Aldrin Piri <
> > [email protected]> wrote:
> >
> >
> > Mans,
> >
> > I think the ControllerService is definitely the right play moving
> forward.
> > What I think we can do is subclass the current AWSAbstractProcessor and
> > override onScheduled to provide a way to interact with the
> > ControllerService and, should one not be configured, defer to the parent
> > implementation.
> >
> > We can mark AWSAbstractProcessor as deprecated and maintain backward
> > compatibility while adding some new functionality in for the accompanying
> > processors.
> >
> >
> > On Sat, Jan 9, 2016 at 8:01 AM, M Singh <[email protected]>
> > wrote:
> >
> > > Hi:
> > > I have started working on implementing a controller for creating creds
> > > provider and changing the createClient method to use the controller.
> > > If there is any advice/feedback on this, please let me know.
> > > Thanks again.
> > > Mans
> > >
> > >
> > >
> > > On Friday, January 8, 2016 12:48 PM, M Singh
> > > <[email protected]> wrote:
> > >
> > >
> > > Hi Aldrin:
> > > The unfortunate things is that AWSCredentialsProvider does not inherit
> > > from AWSCredentials interface.
> > > As far as I can see, the provider interface is much more flexible and
> > > provides everything with we/anyone can need. As we can see, the creds
> > > based constructors (AmazonS3/SQS/SNSClients) internally create a static
> > > creds provider instance. If we support both the creds and creds
> provider
> > > based arguments, it could also confusing and error prone for developers
> > > extending the class.
> > > Even if we have an adapter how would the subclass of
> AbstractAWSProcessor
> > > call createClient allow the two arguments (creds and creds provider) to
> > > work seemlessly. Let me know if you have any other thoughts/paths I can
> > > investigate.
> > > Thanks for the feedback and I am learning a lot with this experience.
> > > Mans
> > >
> > > On Friday, January 8, 2016 11:49 AM, Aldrin Piri <
> > [email protected]>
> > > wrote:
> > >
> > >
> > > Mans,
> > >
> > > Thanks for sticking with this and continuing to see things through, the
> > > community certainly appreciates it as these are very popular processors
> > and
> > > this functionality will help a wide base of users.
> > >
> > > I am poking around a bit more and thinking we might be able to work
> > > something out with a class that adapts an AWSCredentialsProvider to
> > > AWSCredentials. The AWSProvider interface is just composition of
> > > AWSCredentials with a refresh method. Need to mull things over a bit,
> > and
> > > dig through the associated libraries to understand how these are
> > typically
> > > used, but this feels like it could be another avenue to consider and
> > where
> > > I am directing my attention at the moment.
> > >
> > > --aldrin
> > >
> > > On Fri, Jan 8, 2016 at 9:58 AM, M Singh <[email protected]>
> > > wrote:
> > >
> > > > Just one more note Joe (as mentioned in the Jira Ticket)
> > > > From what I can see, we cannot just deprecate createClient method in
> > the
> > > > AbstractAWSProcessor which uses the AWSCredentials argument, since
> the
> > > > subclasses AbstractS3/SNS/SQSProcessor call that to create the
> > respective
> > > > clients. We will have to change the argument to
> AWSCredentialProvider.
> > > > If I can assist with any other investigation, please let me know.
> > > > Thanks again.
> > > >
> > > > On Friday, January 8, 2016 5:31 AM, M Singh
> > > > <[email protected]> wrote:
> > > >
> > > >
> > > > Thanks Joe.
> > > > If you think that we can accept the change to creds provider, I will
> > work
> > > > on making all the components in nifi aws processors to be consistent.
> > I
> > > > think using the creds provider interface is the way to go since it is
> > > more
> > > > flexible and at this moment we just have 3 aws processors to migrate.
> > > > Looking forward to hearing from you/anyone else for advice/feedback.
> > > > Mans
> > > >
> > > > On Friday, January 8, 2016 5:18 AM, Joe Witt <[email protected]>
> > > > wrote:
> > > >
> > > >
> > > > Mans,
> > > >
> > > > I am working to put out a proposed roadmap and then probably won't be
> > > > very responsive until later tonight. Will try to help then if no one
> > > > else has had a chance to.
> > > >
> > > > That said I see what you mean in terms of a breaking change in the
> > > > processor implementation as far as anyone else that has extended it.
> > > > There have been some discussions recently about this and I think the
> > > > plan is to start annotating everything with the audience and
> stability
> > > > of a given bit of code. Processors are not meant to be locked down
> > > > APIs. So, for now, given that it has been ambiguous to the community
> > > > the best course is to probably just deprecate a given method if it
> > > > cannot be safely repurposed and then use a new one which does meet
> the
> > > > need in the event the controller service is supplied. This last
> > > > statement though is not based on me having looked at the code in any
> > > > detail yet.
> > > >
> > > > Thanks
> > > > Joe
> > > >
> > > > On Fri, Jan 8, 2016 at 8:08 AM, M Singh <[email protected]
> >
> > > > wrote:
> > > > > Hi Joe:
> > > > > I have not worked with the controller interface and aws processors
> so
> > > > perhaps you can help me understand it .
> > > > > From what I can see (as mentioned in
> > > > https://issues.apache.org/jira/browse/NIFI-1325): currently, the
> Nifi
> > > > AbstractAWSProcessor has a method
> > > > > protected abstract ClientType createClient(final ProcessContext
> > > context,
> > > > final AWSCredentials credentials, final ClientConfiguration config);
> > > > > This method is overridden in the AbstractS3/SNS/SQSProcesors to
> > provide
> > > > the respective amazon s3/sns/sqs client using AWSCredentials
> argument.
> > > > > Here is a snippet from AmazonS3Client:
> > > > > public AmazonS3Client(AWSCredentials awsCredentials,
> > > > ClientConfiguration clientConfiguration) {
> > > > super(clientConfiguration); this.awsCredentialsProvider = new
> > > > StaticCredentialsProvider(awsCredentials); init(); }
> > > > > So, AmazonS3/SNS/SQSClient created with AWSCredentials use
> > > > StaticCredentialsProvider in their implementation.
> > > > > All the AWSCredentials impls are static creds (Anonymous/Properties
> > > > Credentials) except for the STSSessionCredentials which has a refresh
> > > > method but is deprecated in favor of the
> STSSessionCredentialsProvider
> > > > interface. AWSCredentials is the interface being used in nifi aws
> > > > processors.
> > > > > The AWSCredentialsProvider interface has a fresh method which all
> > it's
> > > > subclasses implement appropriately - the static ones (like
> > > > PropertyFileCredentialsProvider/StaticCredentialsProvider have a no
> op
> > > for
> > > > refresh method) as follows:
> > > > > public void refresh() {}
> > > > > From what I can see, there is no common interface available for
> > > > AWSCredentials and AWSCredentialsProvider that Nifi's
> > > > AbstractAWSProcessor.createClient can support. So if we need to use
> > the
> > > > controller interface with creds providers, will will have to change
> > > > AbstractAWSProcessor.createClient to the following
> > > > >
> > > > > protected abstract ClientType createClient(final ProcessContext
> > > context,
> > > > final AWSCredentialsProvider credentialsProvider,
> > > > > final ClientConfiguration config);
> > > > > This appears to be a breaking change for the clients who have
> > extended
> > > > the AbstractAWSProcessor.createClient with the AWSCredentials
> argument
> > > > rather that the AWSCredentialsProvider.
> > > > > So, can you please elaborate on how the AbstractAWSProcessor will
> be
> > > > able to support both the current impl (ie, invoking aws components
> with
> > > > creds) and the proposed credentials provider interface ?
> > > > > Thanks
> > > > > Mans
> > > > >
> > > > > On Thursday, January 7, 2016 9:00 PM, Joe Witt <
> > [email protected]>
> > > > wrote:
> > > > >
> > > > >
> > > > > Mans,
> > > > >
> > > > > It appears to me that there is a path for this not to be a breaking
> > > > > change for the flow. By creating a controller service to handle
> the
> > > > > credential provider piece you should be able to just update the
> > > > > processor to support that controller service interface. If the
> user
> > > > > sets that controller service then you use that and if they don't
> then
> > > > > you revert to using the older properties. We can mark those
> > > > > properties as no longer the preferred model and deprecate them in
> the
> > > > > codebase then when we reach a 1.0 milestone we can remove them.
> > > > >
> > > > > Thanks
> > > > > Joe
> > > > >
> > > > > On Thu, Jan 7, 2016 at 9:07 PM, M Singh
> <[email protected]
> > >
> > > > wrote:
> > > > >> Hi:
> > > > >> Just wanted to mention that if we go with the creds provider
> > interface
> > > > it will be breaking change for nifi aws components as mentioned in
> > > > https://issues.apache.org/jira/browse/NIFI-1325.
> > > > >> Also, I am considering creating one aws creds provider controller
> > > which
> > > > will provide creds provider based on property file, basic, anonymous
> or
> > > > assume role session params.
> > > > >> Please let me know if there is any additional feedback for me.
> > > > >> Thanks again.
> > > > >> Mans
> > > > >>
> > > > >> On Thursday, January 7, 2016 2:56 PM, M Singh
> > > > <[email protected]> wrote:
> > > > >>
> > > > >>
> > > > >> Hi Joe:
> > > > >> Based on your feedback I will try to explore the controller
> > interface
> > > > for aws creds provider.
> > > > >> Thanks for your advice.
> > > > >> Mans
> > > > >>
> > > > >> On Thursday, January 7, 2016 4:15 AM, Joe Witt <
> > [email protected]
> > > >
> > > > wrote:
> > > > >>
> > > > >>
> > > > >> Mans
> > > > >>
> > > > >> Appreciate you pushing this forward. There is a related idea to
> > > better
> > > > >> handle aws credentials for all the aws procs. Will look more and
> > > > respond.
> > > > >>
> > > > >> Thanks
> > > > >> Joe
> > > > >> On Jan 7, 2016 6:52 AM, "M Singh" <[email protected]>
> > > wrote:
> > > > >>
> > > > >>> Hi:
> > > > >>> Just wanted to follow-up and see if anyone has any feedback on .
> > > > >>> https://issues.apache.org/jira/browse/NIFI-1325.
> > > > >>> Thanks
> > > > >>> Mans
> > > > >>
> > > > >>
> > > > >>
> > > > >>
> > > > >>
> > > > >
> > > > >
> > > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > >
> > >
> > >
> > >
> > >
> > >
> >
> >
> >
> >
>
