Skip to content

Latest commit

 

History

History
 
 

aries-backchannels

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Aries Backchannels

This folder contains the Aries backchannels that have been added to the Aries Agent Test Harness, each in their own folder, plus some shared files that may be useful to specific backchannel implementations. As noted in the main repo readme, backchannels receive requests from the test harness and convert those requests into instructions for the component under test (CUT). Within the component backchannel folders there may be more than one Dockerfile to build a different Test Agents sharing a single backchannel, perhaps for different versions of the CUT or different configurations.

Writing a new Backchannel

If you are writing a backchannel using Python, you're in luck! Just use either the ACA-Py or VCX backchannels as a model. They sub-class from a common base class (in the python folder), which implements the common backchannel features. The Python implementation is data driven, using the txt file in the data folder.

What's Needed

If you are implementing from scratch, you need to implement a backchannel which:

  • implements the standard backchannel web service interface (which is very simple) and all of the test step protocols (called topics), of which there are a fair number
  • starts an instance of the component under test (CUT), and
  • converts the requests from test harness into directives understood by the CUT

Once you have the backchannel, you need to define one or more docker files to create docker images of Test Agents to deploy in an AATH run. To do that, you must create a Dockerfile that builds a Docker image for the Test Agent (TA), including the backchannel, the CUT and anything else needed to operate the TA. The resulting docker image must be able to be launched by the common ./manage bash script so the new TA can be included in the standard test scenarios.

Standard Backchannel API

The test harness interacts with each backchannel using a small set of standard set of web services. Endpoints are here:

  • POST /agent/command/{topic}/{operation}
  • GET /agent/command/{topic}/
  • GET /agent/command/{topic}/{id}

That's all of the endpoints your agent has to handle. Of course, your backchannel also has to be able to communicate with the CUT (the agent or agent framework being tested). Likely that means being able to generate requests to the CUT (mostly based on requests from the endpoints above) and monitor events from the CUT.

See the OpenAPI definition located here for an overview of all current topics and operations.

Standard Backchannel Topics and Operations

Although the number of endpoints is small, the number of topic and operation parameters is much larger. That list of operations drives the effort in building and maintaining the backchannel. The list of operations to be supported can be found in this OpenAPI spec. It lists all of the possible topic values, the related operations and information about each one, including such things as:

  • related RFC and protocol
  • a description
  • the HTTP method
  • data associated with the operation

A rendered version of the OpenAPI spec can be found We recommend that in writing a backchannel, any Not Implemented commands and operations return an HTTP 501 result code ("Not Implemented").

Support for testing new protocols will extend the OpenAPI spec with additional topics and related operations, adding to the workload of the backchannel maintainer.

Backchannel/Agent Interaction

The test harness interacts with each published backchannel API using the following common Python functions. Pretty simple, eh?

Docker Build Script

Each backchannel should provide one or more Docker scripts, each of which build a self-contained Docker image for the backchannel, the CUT and anything else needed to run the TA.

The following lists the requirements for building AATH compatible docker images:

  • the Dockerfile for each Test Agent (TA) must be called Dockerfile.<TA>. For example Dockerfile.acapy, Dockerfile.vcx.
    • The ./manage script uses the <TA> to validate command line arguments, to tag the agent, and for invoking docker build and run operations.
    • A backchannel may have multiple TA configurations and versions, each based on a different Dockerfile.
      • See the acapy backchannel where there the Dockerfile.acapy builds the latest released version of ACA-Py, where Dockerfile.acapy-main builds from the ACA-Py main branch.
    • TAs docker images are tagged with their <TA> name.
  • The Dockerfiles must be located in a predefined location in this repo, currently in folders in the aries-backchannels folder.
    • The ./manage script looks for the TA Dockerfiles in those folders.
  • A TA docker image must include all of the processes needed to run the backchannel and the CUT. For example:
    • the VCX TA includes its backchannel, which embeds the VCX agent framework and the code to operate the VCX "dummy" agency.
    • the ACA-Py framework TA includes its backchannel and an appropriate release of ACA-Py.
  • On startup, each TA has a range of ten public (to the docker network) ports available to it.
    • The lowest port number is passed to the TA on startup and is used by the test harness to send HTTP requests to the running TA.
    • The next nine higher ports are exposed across the docker network and can be used as needed by the TA.

Examples are provided for aca-py (Dockerfile.acapy), VCX (Dockerfile.vcx) and .NET (Dockerfile.dotnet).

./manage Script Integration

The ./manage script builds images and runs those images as containers in test runs. This integration applies some constraints on the docker images used. Most of those constraints are documented in the previous section, but the following provides some additional context.

An image for each backchannel using the following command:

      echo "Building ${agent}-agent-backchannel ..."
      docker build \
        ${args} \
        $(initDockerBuildArgs) \
        -t "${agent}-agent-backchannel" \
        -f "${BACKCHANNEL_FOLDER}/Dockerfile.${agent}" "aries-backchannels/"

where:

  • ${agent} is the name of the component under test (CUT)
  • $(initDockerBuildArgs) picks up any HTTP_PROXY environment variables, and
  • ${args} are any extra arguments on the command line after standard options processing.
  • Note that the docker build context is the aries-backchannel folder—the folder above the backchannel folders

Once built, the selected TAs for the run are started for the test roles (currently Acme, Bob and Mallory) using the following commands:

  echo "Starting Acme Agent ..."
  docker run -d --rm --name acme_agent --expose 9020-9029 -p 9020-9029:9020-9029 -e "DOCKERHOST=${DOCKERHOST}" -e "LEDGER_URL=http://${DOCKERHOST}:9000" ${ACME_AGENT} -p 9020 -i false >/dev/null
  echo "Starting Bob Agent ..."
  docker run -d --rm --name bob_agent --expose 9030-9039 -p 9030-9039:9030-9039 -e "DOCKERHOST=${DOCKERHOST}" -e "LEDGER_URL=http://${DOCKERHOST}:9000" ${BOB_AGENT} -p 9030 -i false >/dev/null
  echo "Starting Mallory Agent ..."
  docker run -d --rm --name mallory_agent --expose 9040-9049 -p 9040-9049:9040-9049 -e "DOCKERHOST=${DOCKERHOST}" -e "LEDGER_URL=http://${DOCKERHOST}:9000" ${MALLORY_AGENT} -p 9040  -i false >/dev/null

Important things to note from the script snippet:

  • each backchannel is provided a range of public ports (the -expose parameter), which are mapped to localhost
    • the backchannel must allocate the ports at runtime based on the -p parameter, and not hard code them in the container.
  • the binding of the TA (e.g. acapy or vcx, etc.) is done earlier in the script by setting the ${ACME_AGENT} etc. environment variables
  • environment variables provide the Docker host IP (DOCKERHOST) and a url to the ledger genesis transactions (LEDGER_URL)
    • the variables are defaulted if not already set, with the LEDGER_URL assumed to be for a locally running instance of von-network
  • parameters passed to the backchannel specify the base port number (-p port) and to use non-interactive mode (-i false)

The ACA-Py and Indy Influence

Many of the BDD feature steps (and hence, backchannel requests) in the initial test cases map very closely to the ACA-Py "admin" API used by a controller to control an instance of an ACA-Py agent. This makes sense because both the ACA-Py admin API and the AATH test cases were defined based on the Aries RFCs. However, we are aware the alignment between the two might be too close and welcome recommendations for making the backchannel API more agnostic, easier for other CUTs. Likewise, as the test suite becomes ledger- and verifiable credential format-agnostic, we anticipate abstracting away the Indy-isms that are in the current test cases, making them test parameters versus explicit steps.

The Google Sheet list of operations has that same influence, referencing things like connection_id, cred_exchange_id and so on. As new backchannels are developed, we welcome feedback on how to make the list of operations easier to maintain backchannels.