---
title: "What Changed in HeliosDB Full Perf: Vectorized Comparisons, HNSW for Hybrid, RowDescription Cache"
description: "v7.7–v7.8 was a 12-release perf wave: vectorized Bool/String/Timestamp + mixed-numeric comparisons, HNSW for `/api/v1/vectors/search/hybrid`, RowDescription cache, parse_cache fast hash, compaction throttle. Numbers and the deltas."
date: 2026-06-05
updated: 2026-06-05
tag: "Benchmarks"
canonical: https://www.heliosdb.com/blog/full-perf-vectorized-comparisons-hnsw-hybrid.html
---

# What Changed in HeliosDB Full Perf: Vectorized Comparisons, HNSW for Hybrid, RowDescription Cache

Twelve releases in five weeks. v7.7 through v8.0 was a sustained performance wave; this post is the joined-up view of what shipped, what it bought, and which workloads benefit. Numbers are measured on representative fixtures — reproduce them on your own and report variance.

## The biggest wins

**Vectorized comparisons.** Bool, String, and Timestamp filter predicates were the last hold-outs running scalar — numeric had already been vectorized. Mixed-numeric comparisons (e.g. `i32 = i64`) now go through a fast path that promotes once and SIMDs the loop. Filter throughput on string-heavy workloads roughly doubled.

**HNSW for hybrid search.** The HTTP `/api/v1/vectors/search/hybrid` endpoint was running brute-force scan for the vector half of the fusion. Switching to the HNSW index dropped p99 latency on a 1M-row corpus from ~340 ms to ~18 ms.

**RowDescription cache.** The Postgres wire's RowDescription message is sent once per result-set and was being constructed per-query. Caching it by query fingerprint dropped latency on small result-sets by 8–12% — on small responses the per-row encode now dominates the cost.

**WAL flush dedup.** The HTAP storage engine has its own durable, direct-I/O WAL; a redundant secondary fsync on the same write path was retired. Single-row insert throughput up ~25% on NVMe.

## The unsexy but compounding ones

A handful of mechanical refactors compound into noticeable wins on contended workloads: tightening hot-path locking primitives, bounding previously-unbounded caches, and swapping the parse cache's hash function for one tuned to short keys. Individually each is a percent or two; together they cut several percent off contended workloads.

## The composed picture

On a representative OLTP-plus-vector mix at 1 000 RPS sustained, v8.0.3 is roughly 38% faster than v7.6.x at the same load. The middle releases each contributed 2–8%, compounding. v7.7.0 itself was a small dip relative to v7.6.x — the price of the lock-primitive transition that downstream releases then amortised.

## What this means for your workload

- **Vector + text retrieval at scale** — the HNSW switch is the single largest improvement. If you're using `/api/v1/vectors/search/hybrid`, expect order-of-magnitude p99 improvements on million-row corpora.
- **OLTP at sustained RPS** — the WAL dedup + lock-primitive work compound on multi-tenant servers under contention.
- **String-heavy filters** — reports, search, log queries with WHERE clauses on text columns benefit from the vectorized path.

HeliosDB Full is source-available under SSPL-1.0. Repository: [github.com/dimensigon/HDB-HeliosDB-Full](https://github.com/dimensigon/HDB-HeliosDB-Full).