GraphConstruction¶

GraphStorm gconstruct — build a partitioned DGL heterograph from the node/edge parquets + gconstruct schema emitted by GraphFeatureProcessing. BYO GraphStorm image.


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

GraphStorm gconstruct wrapper (ports Nexus run_gconstruct.sh): rewrites the Step-2 output paths in gconstruct_config.json to the Step-3 input mount, patches a known GraphStorm gconstruct unbound-variable bug (idempotent guard), runs python3 -m graphstorm.gconstruct.construct_graph with multi-process partitioning, then an optional graph-integrity sanity check. Emits a partitioned DGL heterograph (not parquet).

Inputs (dependencies)¶

Input	Type	Required	Compatible producers
`graph_data`	`processing_output`	yes	GraphFeatureProcessing

Outputs¶

Output	Type
`dgl_graph`	`processing_output`

Consumers (downstream steps)¶

Steps that declare this step as a compatible input source:

Framework requirements¶

Package	Version
`graphstorm`	`>=0.3`
`dgl`	`>=1.1`
`pyarrow`	`>=12.0.0`
`pandas`	`>=1.3.0`
`numpy`	`>=1.21.0`
`pyyaml`	`>=6.0`
`tqdm`	`>=4.0`

← Back to the Step Catalog