Custom s3 path for uploading spark configuration files

[nebur395]

Describe the feature you'd like In order to be able to organise our S3 bucket in which the input and the outputs are being uploaded for our Sagemaker pipelines, we want to be able to specify a custom spark configuration path. To improve our data lineage we are trying to store these files by pipeline name and execution ids.

Currently there is no possibility to do so and those spark config files are being uploaded automatically following this convention: https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/spark/processing.py#L391-L394.

So Ideally we would like to be able to specify a custom s3 path or at least a prefix for the uploaded files in S3. It's important to be able to use ExecutionVariables in that s3 path. Something like this:

serialized_configuration = BytesIO(json.dumps(configuration["content"]).encode("utf-8"))
if configuration["s3_uri"] != None:
    s3_uri = (
        f"{configuration["s3_uri"]}/{self._conf_file_name}"
    )
else:
    s3_uri = (
        f"s3://{self.sagemaker_session.default_bucket()}/{self._current_job_name}/"
        f"input/{self._conf_container_input_name}/{self._conf_file_name}"
    )

How would this feature be used? Please describe.

import boto3
import sagemaker
from sagemaker.workflow.pipeline import Pipeline
from sagemaker.processing import ProcessingInput, ProcessingOutput
from sagemaker.workflow.steps import ProcessingStep
from sagemaker.spark.processing import PySparkProcessor
from sagemaker.workflow.execution_variables import ExecutionVariables
from sagemaker.workflow.functions import Join

boto_session = boto3.Session(region_name="eu-west-1")

sagemaker_client = boto_session.client("sagemaker")

default_bucket = "XYZ"

sagemaker_session = sagemaker.session.Session(
    boto_session=boto_session,
    sagemaker_client=sagemaker_client,
    default_bucket=default_bucket
)

bucket_pipeline_execution_prefix = [
    "s3:/", default_bucket, ExecutionVariables.PIPELINE_NAME, ExecutionVariables.PIPELINE_EXECUTION_ID
]

spark_processor = PySparkProcessor(
    base_job_name="sm-spark",
    role="XYZ",
    instance_count=1,
    instance_type="ml.m5.xlarge",
    max_runtime_in_seconds=1200,
    image_uri="XYZ",
    sagemaker_session=sagemaker_session
)

spark_config = {}

run_args = spark_processor.get_run_args(
    submit_app="./processing.py",
    configuration={ 
        content: spark_config,
        s3_uri: Join(on="/", values=[
            *bucket_pipeline_execution_prefix, "spark-config"
        ])
    },
    inputs=[],
    outputs=[
        ProcessingOutput(
            source='/opt/ml/processing/output',
            output_name='demo-output',
            destination=Join(on="/", values=[
                *bucket_pipeline_execution_prefix, "preprocessed"
            ])
        )
    ],
    spark_event_logs_s3_uri=Join(on="/", values=[
                *bucket_pipeline_execution_prefix, "spark-event-logs"
            ]) # work in progress: https://github.com/aws/sagemaker-python-sdk/pull/3167
)

step_process = ProcessingStep(
    name="demo-processing",
    processor=spark_processor,
    inputs=run_args.inputs,
    outputs=run_args.outputs,
    code=run_args.code,
)

pipeline = Pipeline(
    name='demo-pipeline',
    parameters=[],
    sagemaker_session=sagemaker_session,
    steps=[step_process],
)

pipeline.upsert(role_arn="XYZ")

execution = pipeline.start()

In the end the s3 bucket should look like something similar to this:

|- XYZ 
|--- demo-pipeline
|----- execution-1
|------- spark-config
|--------- configuration.json
|------- spark-event-logs
|--------- application_0001
|------- preprocessed
|--------- *.* # processing output files

Therefore now you have everything related with that demo-processing (spark app logs, spark app config files, spark outputs...) under s3://XYZ/demo-pipeline/execution-1/, increasing data lineage, and reproducibility

[jmahlik]

I also ran in to this when using the spark processors for a normal job where certain prefixes in a bucket are locked down. There's no way to write the files to a different prefix other than the root of the bucket.

I ended up patching around it with the hack below. Ideally, this could either be a direct input to the spark processors or specified in the conf like @nebur395 was thinking. It seems like specifying the url in the config might fail _validate_configuration though.

from sagemaker.spark.processing import _SparkProcessorBase, PySparkProcessor
from io import BytesIO
from sagemaker.s3 import S3Uploader

def _stage_configuration(self, configuration):
    """Serializes and uploads the user-provided EMR application configuration to S3.

    This method prepares an input channel.

    Args:
        configuration (Dict): the configuration dict for the EMR application configuration.
    """

    serialized_configuration = BytesIO(json.dumps(configuration).encode("utf-8"))
    s3_uri = (
        # Patch
        f"{my_custom_prefix}"
        f"input/{self._conf_container_input_name}/{self._conf_file_name}"
    )

    S3Uploader.upload_string_as_file_body(
        body=serialized_configuration,
        desired_s3_uri=s3_uri,
        sagemaker_session=self.sagemaker_session,
    )

    conf_input = ProcessingInput(
        source=s3_uri,
        destination=f"{self._conf_container_base_path}{self._conf_container_input_name}",
        input_name=_SparkProcessorBase._conf_container_input_name,
    )
    return conf_input

_SparkProcessorBase._stage_configuration = _stage_configuration

spark_processor = PySparkProcessor(
    base_job_name="my-job",
    role=role,
    instance_count=1,
    instance_type=instance_type,
    image_uri=image,
    sagemaker_session=sagemaker_session,
)

[DmitriyLamzin]

Yes. One more reason to have this feature is that with every pipeline deployment operation a configuration file is uploaded to a new directory, which breaks a step caching functionality.

So I would use this feature in the opposite way. I would set a static path for the configuration file (or based on its hash, whatever). This would allow using the step caching.

[jmahlik]

I believe this is fixed?

[martinRenou]

Closing as fixed by #3486