Flyte enables production-grade orchestration for machine learning workflows and data processing created to accelerate local workflows to production.

Flyte

has anyone thought about a resource model on flyte where instead of specifying resources via `requests` and `limits`, we can just specify an instance type for a task? maybe also a proportion to select like half of an instance for example

we currently do this via a class wrapper:
```@dataclass
class Resources(DataClassJSONMixin):
    name: str
    cpu: int
    mem: int
    gpu: Optional[int] = None
    ephemeral_storage: int = field(default=1000)```
with some presets e.g.
```M5D_XLARGE: Resources = Resources(
    name="m5d.xlarge", cpu=3900, mem=14162, gpu=None, ephemeral_storage=1000
)```
and the class implements `__truediv__` so can be divided

but then to use this, we have to do `@task(requests=DEFAULT.flyte_resource)` which is okay but it would be nicer if this was baked into flytekit to abstract out `requests`, `limits` and maybe even `pod_template` - then, we can specify some instance type resources to come with a node selector/toleration like flyte natively does for gpus

that's an interesting approach with some challenges of course, including how to manage the differences between instance types throughout cloud providers and how to maintain it up to date with the time.

Feel free to start a discussion in the <https://github.com/flyteorg/flyte/discussions/new?category=rfc-incubator|RFC Incubator> so we can discuss it with other contributors

yeah definitely makes sense - I was thinking naively that Flyte could provide the interface like the `class Resources` and it could be up to users to define the list of instance types they want to support

a big motivation being to help ensure that users don't specify invalid resources that don't match the available instances in the cluster and to ensure their workloads don't schedule on unexpected instances

it _might_ not be that huge of a change either - it could be optional to use the new resources class in the `@task`, and essentially will fill out the `requests`, `limits` and `pod_template` for the user

but of course there's `with_overrides` to think about

I'll definitely start a thread in the incubator!

created! <https://github.com/flyteorg/flyte/discussions/6376>