ParadeDB’s benchmark shows a three‑table Top‑K query dropping from two minutes to 100 ms.
*ParadeDB says its new search‑join engine can turn a three‑table Top‑K query from two minutes into 100 ms. The claim challenges the long‑standing practice of denormalizing data into Elasticsearch.*
ParadeDB announced a breakthrough for relational search on October 27 2026. The company claims its new search‑join engine can execute a three‑table Top‑K query in 100 ms, a thousandfold speedup over vanilla Postgres. The benchmark uses a realistic StackOverflow schema—2.3 M users, 20 M posts, 29 M comments, and 29.4 M badges—showing that a single SQL statement can now span multiple normalized tables without the usual latency. If true, the result upends the entrenched practice of denormalizing data into Elasticsearch or other external engines, promising faster results with a single Postgres‑compatible stack.
Search engines like Elasticsearch force users to flatten relational data. The ParadeDB blog notes that traditional pipelines push users to either denormalize via fragile ETL processes or accept Postgres’s slow, row‑by‑row scoring of millions of rows. Denormalization inflates storage up to three times after replication and introduces write amplification, especially when power‑law skew hits popular rows. The result is a brittle architecture that must keep two databases in sync.
ParadeDB’s engineers ran a StackOverflow benchmark with users (2.3 M rows), posts (20 M), comments (29 M) and badges (29.4 M). Their search‑join query dropped from a natural three‑table Top‑K run time of roughly two minutes (118,000 ms) to 100 ms on a single core – a speedup of up to 1,000×. The same test on vanilla Postgres with a B‑Tree + GIN schema and 8‑way parallelism still took about two minutes. The authors stress that the improvement comes without adding external infrastructure.
Postgres must join all 28.5 M rows before it can compute a combined relevance score. Because no index pre‑orders by the sum of scores, the executor fetches every tuple from the heap, pulling all columns needed for the final projection. Row‑by‑row execution, lack of vectorization, and the need to materialize full rows drive the two‑minute latency. Even aggressive rewrites—turning the OR into an 84‑line UNION ALL CTE—only shave the time to 1.2 s for rare terms, while common terms still balloon to 14 s.
ParadeDB eliminates the need for a separate search cluster, but the blog warns that denormalized copies still consume up to three times the storage after replication. The system also respects Postgres’s transactional consistency, meaning write paths still bear the cost of maintaining additional indexes. Users must weigh the storage and write amplification against the dramatic query‑time gains.
ParadeDB’s benchmark shows a three‑table Top‑K query dropping from about two minutes (118,000 ms) to 100 ms, a speedup of up to 1,000×.
The benchmark used a StackOverflow‑style schema with 2.3 M users, 20 M posts, 29 M comments and 29.4 M badges.
Postgres must join all rows before computing combined relevance scores, fetching full tuples from the heap and lacking vectorized execution, which leads to two‑minute runtimes.
The ParadeDB announcement forces engineers to reconsider the default trade‑off between speed and schema purity. By delivering sub‑second relevance across normalized tables, the platform challenges the notion that external search clusters are mandatory for modern applications. The next test will be real‑world adoption: whether teams can tolerate the extra storage and write costs while reaping the latency gains. Until then, the debate over denormalization versus native search joins is far from settled.
Sources: Hacker News (ParadeDB blog post “Do We Still Have to Denormalize? Have Your Schema and Search It Too”)