Laboratoire-de-Chemoinformatique/SynPlanner v1.4.3

on GitHub

Changed

Parallelization

• Removed Ray dependency entirely — all parallel pipelines now use
  ProcessPoolExecutor via the new process_pool_map_stream utility
• process_pool_map_stream enhanced with ordered mode (submission-order
  yield), per-future timeout, initializer/initargs for non-picklable
  worker state, max_tasks_per_child (Python 3.11+), and on_timeout callback
• New graceful_shutdown() context manager for SIGTERM/SIGINT handling with
  automatic signal handler restoration

Data Pipeline

• Standardization, filtering, rule extraction, and ML preprocessing pipelines
  migrated from Ray to process_pool_map_stream with initializer pattern
• Writer-side CGR dedup: hash(~rxn) (condensed graph of reaction hash) for
  mechanism-level reaction deduplication — 8 bytes per entry in memory
• New shared result types: ProcessResult, ErrorEntry, FilteredEntry,
  PipelineSummary in synplan.chem.data.reaction_result

Compatibility

• Removed from __future__ import annotations from all modules (Dagster
  compatibility)
• Forward references quoted for self-referencing return types

Removed

• ray dependency removed from pyproject.toml
• init_ray_logging() removed from synplan.utils.logging
• DedupActor Ray actor removed

Added

• 10 unit tests for process_pool_map_stream and graceful_shutdown
  (tests/unit/utils/test_parallel.py)
• 8 unit tests for ProcessResult, PipelineSummary, and CGR dedup
  (tests/unit/chem/data/test_pipeline.py)