GraphFeatureProcessing

Turn per-seed subgraph pickles + labelled seeds into the GraphStorm GConstruct input (per-type node/edge parquets, reverse edges, node-ID-keyed masks, gconstruct_config.json). BYO GraphStorm image.

SageMaker step type	Processing
Node type	internal (consumes upstream, produces downstream)
Container entry point	graph_feature_processing.py
Build-time requirement	none (SAIS SDK — fatal on load if absent)
Interface file	steps/interfaces/graph_feature_processing.step.yaml

Compute

Compute kind	byo_container
Image	BYO container — config.image_uri is passed VERBATIM to AppSpecification.ImageUri / AlgorithmSpecification.TrainingImage (no image_uris.retrieve; the framework deps live in the image’s Dockerfile)
Container entrypoint	['bash', '/opt/ml/processing/input/code/entrypoint.sh'] (ContainerEntrypoint bypass — the image runs its own entrypoint instead of the SageMaker toolkit)

Functionality

Graph feature-processing orchestrator (ports Nexus run_feature_processing.py): runs prepare_graphstorm_format.py (type-aware node/edge feature extraction, reverse-edge generation, node-ID-keyed multi-task masks, gconstruct_config.json) then custom features then sanity_check.py, all from the BYO GraphStorm image’s bundled code. Output is the GraphStorm GConstruct input tree.

Inputs (dependencies)

Input	Type	Required	Compatible producers
SUBGRAPHS	processing_output	yes	GraphSubgraphExtraction
SEEDS	processing_output	yes	GraphSubgraphExtraction, CradleDataLoading

Outputs

Output	Type
graph_data	processing_output

Consumers (downstream steps)

Steps that declare this step as a compatible input source:

• GraphConstruction

• GraphStormGNNTraining

• GraphStormGNNTuning

Framework requirements

Package	Version
polars	>=0.20.0
pandas	>=1.3.0
numpy	>=1.21.0
pyyaml	>=6.0
tqdm	>=4.65.0

---

← Back to the Step Catalog