Problem
Canva runs several search experiences: audio, fonts, video and media, with templates next. A backend that mixes query cleanup, retrieval, feature lookup and ranking in one code path is hard to explain, reuse and instrument. Canva's search and recommendations group rebuilt the stack around one pipeline model whose phases match the team specializations.
Example from the article: the query "Lune Bleue" is normalised to "lune bleue", translated to "blue moon" and split into two tokens. Annotators tag the tokens with a color and content-type scores, Elasticsearch returns candidates, and rerankers handle diversity, personalisation and free/pro ratios.
Solution in brief
Treat every request as a fixed sequence of steps (validation, rewriting, tokenization, annotation, candidate generation, feature extraction, re-ranking) holding isolated, stateless components. One specification object describes a pipeline, so it can be cloned, cached, A/B tested or interleaved, and switched to a cheaper variant for deep result pages.
Questions this case answers
- How to compose a multi-stage ranking system from isolated, ordered components so that each team can change one stage without touching the others
- How to trade ranking precision against compute cost depending on how deep into the result list the user is
- How to structure a search backend so that query rewriting, retrieval, feature lookup and re-ranking stay separate, reusable and testable?
- How to reuse the same search components across several search experiences (audio, fonts, video, media, templates)?
Tools and methods
- Phased pipeline of isolated components — shared vocabulary, parallelism inside a step, no shared mutable state.
- Transparent query rewriting — rewrites the query and keeps the original.
- Query and token annotation — adds entities, language, sensitive-term class, synonyms.
- Several candidate generators — query OpenSearch, Solr or SageMaker with different recall trade-offs.
- Feature extraction — Redis-backed feature store (DynamoDB meanwhile).
- Interleaving — compares two pipelines in one result list.
Google or a one-time email code. No password.