A human centric machine learning framework
Showing 2 of 2 projects. Click any project card for scope, mentors, and proposal studio.
Mentors: Student: Aryan Dhawan
Metaflow's metadata service returns unbounded JSON for every list endpoint. A single GET /runs on a production deployment returns megabytes of data the caller immediately discards. For AI agents, this is worse: a failure-diagnosis workflow generates 2,021 HTTP calls to check 1,000 tasks. This proposal adds cursor-based pagination and server-side tag filtering to the metadata service (prototyped in PRs #9 and #10, 65 tests passing), integrates pagination into the Python client SDK, and builds agent-friendly query utilities that, across five common agent workflows, collapse 11,828 API calls into 16. Server-side work is already prototyped. GSoC starts at hardening, not greenfield.
Mentors: Student: EunhyeJeong
metadata_service returns raw arrays with no pagination or filtering. ui_backend_service in the same repository already supports both via an envelope format. Meanwhile, the DB layer (find_records, DBPagination, runs_v3_idx_epoch_ts_desc) already has everything needed — the gap is that the API layer does not expose it. This project brings metadata_service to parity: pagination via LIMIT+1, tag filtering with four operators (_tags:any/all/likeany/likeall), unified response schema matching ui_backend_service, and GIN index migration — without breaking existing clients. The key design: without _limit, behavior is identical to current. With _limit, the envelope format activates. Beyond pagination and filtering, the project also connects tag cache refresh to notify.py's existing LISTEN/NOTIFY infrastructure (debounce or incremental update strategy confirmed with mentor before implementation), and adds Content-Type header to the format_response decorator. Deliverables: metaflow-service PR covering all 9 GET endpoints, GIN index migration, feature-gating, benchmark results with EXPLAIN ANALYZE verification, backward compatibility test suite, and a pagination implementation guide for future contributors.