Search you can trust on a live-commerce platform

A multi-brand live-shopping platform: brands publish shoppable video content on their own storefronts and run live sessions with chat and an in-session cart. It ships a Shopify app and an embeddable SDK on top of a shared platform API.

Search results drifted from the source data. The index carried copies of fields that changed elsewhere, such as display names, so renames left stale results behind. At the same time the Shopify app still ran on legacy serverless functions that duplicated platform logic.

• Rank in the index, hydrate from the database

  The search index answers ranking, facets and totals; the item bodies come from the database. The index stays lean and holds nothing that can go stale.

  Trade-off Every search does a second read to hydrate results.

• Index sync driven by domain events

  I rewrote the sync from the database to the index around domain events and dropped every field that couldn't be kept consistent.

  Trade-off More event traffic: every change that affects search triggers an index update.

• Shopify app moved onto the platform API

  Install-time catalog sync and product webhooks now go through the platform's own API, with a mapped legacy-versus-current flow so old code could be deleted safely.

  Trade-off A short period of running two backends side by side.

• A backfill without a new index

  One streamed query with filtering in code, instead of adding a database index for a one-off job.

  Trade-off Some logic is duplicated, guarded by an assert self-check.

• Delete fields only with proof

  Unused fields were removed only after scanning all production data and finding no writers and no values.

  Trade-off A slower clean-up than deleting on sight.

1. Two stores, one truth

   Keeping a database and a search index consistent under renames, deletes and re-links, without a nightly rebuild.

2. Shopify edge cases

   Shops with several domains and URL variants, and a billing change shipped as a one-shot migration with a runbook.

3. A safe backfill

   Bringing every existing record into a new state model in one pass, without downtime or a temporary index.

My own agent setup around the repo: an index-only CLAUDE.md that routes to per-area rules, a domain glossary, a legacy-versus-current flow map, custom builder and reviewer subagents, and written specs and plans with session handoffs. I also integrated the platform SDK, live sessions and cart sync into partner storefronts built with Next.js and React Native.