`"FastFile"` for Processing Job Input

[lorenzwalthert]

Describe the feature you'd like

"FastFile" to be an available option for s3_input_mode in sagemaker.Processing.ProcessingInput, in addition to "File" and "Pipe". The s3 input mode is already available for TrainingInput since 2021 and greatly improves speed (-82%) according to an AWS Blog post.

How would this feature be used? Please describe.

To speed up processing jobs compared to donwloading all data and allow complex filtering of files before accessing them.

Describe alternatives you've considered

Other methods like

• downloading relevant files as part of training job with sagemaker.s3.S3Downloader(). Problem: I can't shard by s3 key and have to build my own sharding logic.
• Using S3 prefix as the s3_data_type in sagemaker.Processing.ProcessingInput to filter out by prefix: Problem: Some data can't be easily filtered by prefix and you need more complex pattern matching.
• Using a ManifestFile.

Additional context

I know it's not an SDK topic as long as the underlaying APIs don't provide that functionality but I don't know where I can put the feature request otherwise.

[lorenzwalthert]

I think this is actually low prio, since I was able to use ManifestFile to achieve my goal as well.

[martinRenou]

As far as I can see, there is no validation in the code that would prevent from using "FastFile" instead of "File" for s3_input_mode. Have you seen an error when trying?

The only thing that seems to be missing is mention to "FastFile" in the docstring (e.g. https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/processing.py#L1245). I will open a PR for that.

[martinRenou]

Closing as answered and the dosctrings was fixed. Feel free to reopen if you think the issue still stands.

[lorenzwalthert]

Thanks @martinRenou.

[julestalloen]

It seems this doesn't actually work as the underlying API does not support it:

S3InputMode
Whether to use File or Pipe input mode. In File mode, Amazon SageMaker copies the data from the input source onto the local ML storage volume before starting your processing container. This is the most commonly used input mode. In Pipe mode, Amazon SageMaker streams input data from the source directly to your processing container into named pipes without using the ML storage volume.

Type: String
Valid Values: Pipe | File
Required: No

https://docs.aws.amazon.com/sagemaker/latest/APIReference/API_ProcessingS3Input.html

I also get the following error when trying to use FastFile for my ProcessingInput:

botocore.exceptions.ClientError: An error occurred (ValidationException) when calling the CreatePipeline operation: Unable to parse pipeline definition. Model Validation failed: Value 'FastFile' for 'ProcessingS3InputMode' failed to satisfy enum value set: [Pipe, File]

[lorenzwalthert]

To clarify, I did not try the suggested solution by Martin.

[julestalloen]

@lorenzwalthert Thanks for the clarification. Could you reopen this issue? If not, I'll create a new one.

[lorenzwalthert]

I can't reopen it seems.

[ShwetaSingh801]

Hi @lorenzwalthert,

Thanks for using SageMaker and taking the time to suggest ways to improve SageMaker Python SDK. We have added your feature request it to our backlog of feature requests and may consider putting it into future SDK versions. I will go ahead and close the issue now, please let me know if you have any more feedback. Let me know if you have any other questions.

Best,
Shweta