Current structure

• score_processor(context_map, score_cache, metrics_collect) (/scores/enwiki/reverted/23456789)
• scoring_context (== wiki)
• scorer_model (revscoring)
• score_caches
• metrics_collectors
• wsgi -- (Flask framework)
• utilities -- (CLI utilities)

Problems:

• Cool thing: https://ores.wmflabs.org/v2/scores/enwiki/damaging/642345234/?features
  • It gets messy when working with celery (multiprocessing)
• Performance improvment: https://phabricator.wikimedia.org/T134606
  • Opportunity: Many models use the same features (reverted, damaging, goodfaith) & (wp10)
  • Problem: cache can't be modified in place in celery/multiprocess (pickle and send somewhere)
  • Solution: change _process function to apply to multiple models
• Painful to do features output or cache injection with complex request
  • https://ores.wmflabs.org/v2/scores/enwiki/?models=reverted|damaging&revids=2345678|3456789|987654
• Web nodes load models into memory for no reason
  • https://github.com/wiki-ai/ores/blob/master/ores/score_processors/celery.py#L35
  • Solution: Don't load the models into memory on the web nodes -- celery needs know how to gets us model information

Proposed structure
Philosophy: Make "scoring system" look like ORES API -- wsgi (web nodes) just be a thin wrapper

• scoring_systems(context_map, score_cache, metrics_collector)
  • Methods:
    • score(context, model_names, rev_ids) -- "/score/context?model=<model_names>&revids=<rev_ids>" || "/score/context/<model_name>/?revids=<rev_ids>"
    • model_info(context, model_names) -- "/scores/context?model_info&models=<model_names>" || "/scores/context/<model_name>/?model_info"
  • Notes:
    • variants (simple, multiprocessing_pool, celery_queue, etc.)
• scoring_context