What I intent to do is to request resource overrides if the dataset size is too big.

Would you mind letting me know the reason behind returning

Resources

? You should request for resources in the task decorator: https://docs.flyte.org/projects/cookbook/en/latest/auto/deployment/customizing_resources.html.

Hi @tall-lock-23197, Because I extends a class from PythonInstanceTask.

You can request for resources in a PythonInstanceTask by sending

requests

as an argument: https://github.com/flyteorg/flytekit/blob/07299d0b0e8d172e33ad48d719f9014effb5d96b/flytekit/core/python_auto_container.py#L37.

That’s exactly what I am doing. But first I need to check the input dataset size to determine the resource memory usage dynamically by adding a task that returns the Resource object.

@task
def get_xgb_task_resource_request(df: pd.DataFrame) -> Resources:
    _Gibibyte = 1073741824
    data_size_gb = df.memory_usage(index=True, deep=True).sum() // _Gibibyte
    # request more memory for the task if the dataset size is >= 2 GB, because the default task mem alloc might not be enough.
    resource_requests = Resources(mem=f'{data_size_gb}Gi') if data_size_gb >= 2 else None
    return resource_requests

Good to hear that you were able to solve the problem. Have you created a new

PythonInstanceTask

and requested resources in the

execute()

method? Would you mind sharing the coding snippet?

@tall-lock-23197 Sure. Will this work in the execute() method? It runs without error though.

_Gibibyte = 1073741824
        data_size_gb = (train_size + validation_size) / _Gibibyte
        """
            Dynamically request task memory in GB + 0.1 GB buffer for the task (to 3 decimals), 
            because the default task memory allocation might not be enough for big dataset
        """
        self.resources.requests = Resources(mem=f'{round(data_size_gb + 0.1, 3)}Gi')

Yeah, I think this should work. But I’m not sure if these resources will be allocated. Have you checked the task spec?

@tall-lock-23197, where and how to check task spec memory?

On the UI?

It doesn’t work when setting resource.requests in execute() method.

Yeah! You’ll need to set it in the init method.

Then it won’t take calculated / dynamic values for resources at runtime which defeats the purpose.

@glamorous-carpet-83516 @high-accountant-32689, how can the task resources be set dynamically?

@tall-lock-23197, That’s when you define task with the decorator. I don’t think with_overrides() is available for tasks defined by PythonInstanceTask Class. Plus with_overrides() doesn’t dynamically allocate resource based on realtime resource calculation either.

I think

with_overrides()

should work with PythonInstanceTask class. Also regarding real-time resource calculation, I’m not sure how that can be done. I’ll defer to Kevin/Eduardo to answer.

@salmon-refrigerator-32115, you can't set the resources for a task after the task is running, but you could use a dynamic task to achieve this suggestion of setting task resources dynamically. For example:

@task
def t(a: int):
    ...

@dynamic
def dyn(a: int):
    t(a=a).with_overrides(
        requests=Resources(mem="432Mi"),
        limits=Resources(mem="543Mi"),
    )