Would anyone be able to point me in the direction ...
# ask-the-community
a
Would anyone be able to point me in the direction of where Flyte’s DAG model lives in the source code repository?
d
So the CompiledWorkflowClosure is the protobuf definition of a workflow. It is the result of "compiling" a Workflow proto. Then FlyteAdmin calls the BuildFlyteWorkflow function to create a FlyteWorkflow CRD based on a
CompiledWorkflowClosure
and the input values to the workflow. The CRD is what FlytePropeller uses to execute a workflow.
Maybe a little more information than what you wanted - but depending on what you're looking for it can be in one of a few places.
a
@Dan Rammer (hamersaw) I appreciate all the info I can get, so thank you! I actually wrote my own DAG workflow framework before discovering Flyte so I am very interested in the internals. Since I don’t have the support of a team for my project, it’s on hold for now 😅
So in short, i am interested in how Flyte treats workflows from a graph theoretic perspective
@Dan Rammer (hamersaw) Thanks for the excellent links to Flyte’s underlying DAG model. Is there a place you could point me to for more code providing context on the actual execution heuristics for the DAG model?
My team and I are curious about this and if there is any existing explicit message passing algorithm for nodes in a Flyte workflow. Wondering if Flyte is stateful or stateless, and if the Chandy–Lamport algorithm might be a useful feature to contribute
d
Oh fantastic! So right now things are a little embedded. The entrypoint is the RecursiveNodeHandler function which is initially called on the DAGs
start-node
. It then traverses throughout the DAG in a depth first search scheduling and executing nodes along the way. Admittedly this may not be the most efficient algorithm for many workflows - it's something of a passion project of mine to refactor this so I'm personally pretty invested in your vision here. There are a few other community members who has expressed interest in allowing plugable schedulers into Flyte and allow for different workflows to execute under different schedulers.
As far as the stateful vs stateless debate - the components of Flyte are stateless. So flytepropeller handles workflow executions. It stores execution state in a k8s CRD, aptly named FlyteWorkflow and then operates as a k8s controller. So it periodically processes the workflow, checking node status' and scheduling new nodes (along with many other operations).
and then updates the CRD with the current execution state.
If, after you look into the code a little bit, you (and / or your team) are interested in diving into abstracting the node scheduling algorithms away I would be very interested in hopping on a call to discuss this. We could potentially setup some collaboration with other community members.
a
Absolutely! There is a lot I’d also love to discuss. Right now we’re still looking to get the green light from leadership after we demonstrate incremental proof-of-concepts, but assuming that happens I’d want to look into getting an additional green light to disclose info on my DAG framework to share details about that with you.
d
That sounds great! Please keep us posted.
a
Will do!
So not all DAGs are Trees, but all Trees are DAGs, and Flyte’s
executor
/DAG model just so happens to be a Tree? 🙂
102 Views