A New Architecture for Plaid Link: Server-Driven UI with Directed Graphs

Would be interested in:

1. How you implemented this on the backend? Did you consider using something like Temporal, AWS Simple Workflow, or AWS Step Functions?

2. Interested in how you're avoiding (waiting) on network calls on each action. You mentioned it was coming in next post.

3. How are you managing multiple workflows? Does a common child graph make sense for a common workflow?

> 1. How you implemented this on the backend? Did you consider using something like Temporal, AWS Simple Workflow, or AWS Step Functions?

Given this already a major change in our clientserver interaction model and that we didn't want to _fully_ rewrite the entire world, we didn't consider additional major changes to our existing backend services strategy (Go services + RPC) as part of this project. We had a bunch of backend logic and common code we wanted to reuse across Flex Link and the legacy endpoints, which gives us a stepwise mechanism for future refactors once the old APIs are turned down.

Once this rolls out to 100%, the new API boundaries do allow us to refactor the backend into something more like you suggest — leveraging step functions, lambdas etc if desired.

> 2. Interested in how you're avoiding (waiting) on network calls on each action. You mentioned it was coming in next post.

Yes! Full details of that would take a dedicated post. However, if you look at the "Client Rendering Loop" diagram in the post ( https://images.ctfassets.net/zucqsg1ttqqy/5kp6LibAllW1YW69jT... ) you'll see there is a "render neighbor panes" loop in the flow. At the point where we've walked the graph and encountered a pane, we can do a quick breadth-first-search to render out neighboring panes (that don't have a processor in between) and serve those all to the SDK. We all them "additional panes". The response includes a map from "action" to "additional pane" so the SDK can know whether to hit the API or use an already-provided additional pane. (this is in the diagram as well)

The nice part is, clients don't _have_ to implement this functionality. They can always hit the server each time they want the next pane. But as a performance optimization they can leverage the additional panes to render many things locally.

> 3. How are you managing multiple workflows? Does a common child graph make sense for a common workflow?

We use the expanded semantic versioning in the post for teams to have their own versioned workflow namespace. On the `workflow/start` call, we have a dispatcher that can pick which workflow to serve based on the provided configuration, SDK version, platform, experiments, etc.

We don't yet support subgraphs for common chunks of a workflow but it is definitely in our roadmap and would allow teams to share common pieces of the graph.

That is absolutely not true. only thing that's self-registration is development access which is limited to 100 customer connections. Going to production requires submitting a ticket which is picked up a person that transmits requirements from compliance team.

Basically you need to fill out an exhausting 8 part form drilling down on every aspect of your business -

Security - policies, architecture, breach procedures, external review, GDPR compliance, hosting structure, cryptography.

Compliance - proove registration with different compliance bodies, proof of insurance (professional indemnity AND cyber), AML measures, KYC measures, proof of purchase of monitoring systems..

Identity - provide information on ALL directors and shareholdeing structure, provide proof of address and identity of 2 of them, name a company signatory and provide proof of identity and address for him as well.

And more and more... took us 2 months just to compile all the information in endless back-and-forth email conversations.

You should be able to to put traffic through via the development environment without that, but you're right that to go to full production you would need to answer some compliance questions, but it should not take months. I understand your frustration if it's taking that long.

You have my email, and if you reach out I can see what I can do to help. Best.