Problem
Stitch Fix serves personalized search through a chain of microservices. When a client searches for running shoes and gets paisley sweaters, the team must quickly say which service caused it, so that its owners can fix it. Reading logs from every service is slow and does not scale with iterative experiments and several teams working together. The questions are always the same. Did the query parser misread the attributes? Did the core engine match them to the wrong items? Did reranking overweight the client's past taste?
Example from the article: a client searches "plaid shacket" and the team needs to see how the query turns into weighted subqueries, a candidate set and a personalized ranking.
Solution in brief
Build one interactive page over the production search API that draws the whole pipeline for a given query and client. Linked views let an engineer spot the faulty stage, follow one item through all stages, and rerun the search with a changed configuration.
Questions this case answers
- How to localize the faulty component of a multi-stage ranking pipeline when each stage is owned by a different team and logs alone are too slow to read
- How to trace a single expected item through every stage of a ranking pipeline to see where it was lost
- How to find which service in a multi-stage personalized search pipeline produced a bad result?
- How to find out where an expected item was dropped on its way from the catalog to the final ranking?
Tools and methods
- Sankey diagram of the pipeline — shows query, parsed subqueries, core-engine candidates and reranked results in one picture.
- Linked result views — thumbnail cards, beeswarm plots of metrics, parallel coordinates, relevance-score table.
- Pipeline sandbox — editable configuration, saved in a lightweight store and shared by URL.
- React, D3.js, visx — the frontend stack of the tool.
- Elasticsearch — the core search engine whose scores are displayed.
Google or a one-time email code. No password.