Skip to main content

Server Modules

The server entry point is pipeline/server_app.py. server_fn(context) reads Flower run configuration and creates FederatedGWASStrategy.

Output directory selection​

pipeline/server_app.py requires the server output directory to be declared explicitly. It resolves the directory in this order:

  1. server_output_dir from Flower run_config.
  2. configs/server/config.yaml under the configured config_path.

For the default tiny correctness experiment, the server config is:

experiments/correctness/tiny_even/configs/server/config.yaml

It declares:

output:
  intermediate_dir: "experiments/correctness/tiny_even/results_2/server/intermediate"
  log_dir: "experiments/correctness/tiny_even/results_2/server/logs"

Strategy responsibilities​

FederatedGWASStrategy extends Flower Strategy and implements:

  • initialize_parameters: returns empty parameters because GWAS does not use a conventional global model.
  • configure_fit: samples all available clients and sends the current stage configuration.
  • aggregate_fit: parses client responses, calls aggregators, and advances the stage machine.
  • configure_evaluate, aggregate_evaluate, and evaluate: return empty evaluation results because traditional model evaluation is not used.

Aggregators​

FilePurpose
aggregator_king.pyReconstructs client chunks, merges PLINK data, runs PLINK --het, and runs PLINK2 --make-king-table.
aggregator_lr.pyReconstructs LR chunks, merges PLINK data, runs --logistic, and parses p-values.
prg_masking.pyTracks public keys and provides curve parameters for encrypted client-to-client stages.

Strict stage behavior​

strategy_strict.py keeps stage configs Scalar-only and does not advance through encrypted relay stages when no valid client results arrive. It retries sync, global_qc, KING, and LR drain stages instead of using the older simulation workaround strategy.