← Experience

metaflow (netflix / anaconda) · google summer of code '26

ml engineering

2026 · waterloo, on

faster, smaller apis for the open-source ml workflow framework.

• — shipped async postgres filters and cursor pagination (#481, #487): payloads 200× smaller and latency 21× lower on 10k runs (v2.6.0).
• — cut inherited-tag reads from 2 sql queries to 1 with opt-in runs_v3 joins on the step, task and artifact apis (#489).
• — cut a 100-run flow.runs listing from 101 http calls to 2 with pagination and a legacy fallback (#3348, merged).
• — wrote the cursor rfc and prototype (#9, #11); found unbounded sql joins caused by a zero limit (#473).