Skip to content

architecture.md

Latest commit

 

History

History
94 lines (63 loc) · 3.76 KB

architecture.md

File metadata and controls

94 lines (63 loc) · 3.76 KB

GiGL Architecture

GiGL now has two execution models:

  • The current, recommended path uses in-memory subgraph sampling, where the graph is loaded into memory and sampled live during training and inference. If you are looking for the older tabularized pipeline, see Deprecated tabularized docs.
  • The older tabularized path materializes sampled subgraphs ahead of time through Subgraph Sampler and Split Generator. NOTE: The tabularized version of GiGL will be removed in a future release.

This page focuses on the current in-memory subgraph sampling architecture and points to the legacy docs separately.

Primary Pipeline Flow

The primary GiGL flow is:

Config Populator -> Data Preprocessor -> Trainer? -> Inferencer -> Post Processor

Trainer is optional. Inference-only pipelines skip training and run inference against a graph using a pre-trained model.

For the shared runtime behavior behind the current path, see In-Memory Subgraph Sampling.

Components

Config Populator: Freezes the template task config into a runnable GbmlConfig.

Data Preprocessor: Builds graph metadata, transforms features, and enumerates node IDs into compact integer IDs.

Trainer: Launches either legacy training or in-memory subgraph sampling training.

Inferencer: Launches either legacy inference or in-memory subgraph sampling inference.

Post Processor: Restores original node IDs for outputs produced by in-memory subgraph sampling and runs optional user-defined post-processing logic.

Component Diagram

Below is a high-level system overview. Note that both training and inference are backed by the same in-memory graph sampling engine.

System overview

Source Entry Points

Component Source Code
Config Populator {py:class}gigl.src.config_populator.config_populator.ConfigPopulator
Data Preprocessor {py:class}gigl.src.data_preprocessor.data_preprocessor.DataPreprocessor
Trainer {py:class}gigl.src.training.trainer.Trainer
Inferencer {py:class}gigl.src.inference.inferencer.Inferencer
Post Processor {py:class}gigl.src.post_process.post_processor.PostProcessor

Related Guides

  • For the shared in-memory runtime, deployment modes, example entry points, and cost discussion, see In-Memory Subgraph Sampling.
  • For stage-specific behavior and configuration, use the component guides linked above.

Deprecated Tabularized Architecture

If you are maintaining an older deployment that still relies on precomputed sampled subgraphs, see Deprecated tabularized docs.

That flow is:

Config Populator -> Data Preprocessor -> Subgraph Sampler -> Split Generator -> Trainer -> Inferencer

:maxdepth: 2
:hidden:

components/config_populator
components/data_preprocessor
components/trainer
components/inferencer
components/post_processor
in_memory_subgraph_sampling
deprecated_tabularized/index