Engineering deep dives, product updates, and insights from the HeliosDB team.
From a new account to a filtered, sorted result in five minutes: pick a table, tick columns, add WHERE conditions, set ORDER BY and LIMIT, and the console writes the SQL for you. Then take the same query to REST with bound parameters, or to an AI agent over MCP.
Read more →Two releases since 1.7. 1.9 records what the backend actually committed and replays it in commit order, from disk. 2.0 drives query-cache invalidation from that record on both wire protocols and every tier, makes Patroni a write authority, and fails closed on conflicting primaries. With the cache on, write throughput is back at the cache-off level (+23–45% against 1.9).
Read more →Replay now digests every response as the client saw it and re-verifies it on the replacement backend — divergence rolls back with 40001. Snapshot-pinned transactions are refused, read re-execution is a real side-effect policy, and one deadline bounds the whole recovery. In-session failover is core; no cargo feature required.
Read more →Sessions survive a primary failure: re-home and restore state, re-run interrupted reads, opt-in replay of uncommitted transactions. A COMMIT with unknown outcome is never retried.
Read more →The WAL is now fsync-durable on acknowledgement, with real group commit so durable throughput scales with concurrency. Plus request-path panics eliminated, a startup-handshake timeout and connection cap, PRIMARY KEY uniqueness enforced across restarts, and ~17× read / ~36× write gains at eight clients on representative fixtures.
Read more →Three structural ceilings of pgvector at corpus scale: RAM cost, deletes that rot recall, licensing if you want to host it. v3.32 closes all three — over the Postgres wire, under Apache 2.0, with existing clients and ORMs unchanged.
Read more →Two clean signals. The new persistent PQ-HNSW path hits 16× compression at sub-0.5% recall drop and is restart-survivable; the default-features binary is byte-identical to v3.31.2 with measured cargo bench deltas (−13 % … +16 %) that scatter bidirectionally — the signature of machine noise, not code.
v7.7–v7.8 was a 12-release perf wave: vectorized Bool/String/Timestamp + mixed-numeric comparisons, HNSW for hybrid search (`/api/v1/vectors/stores/{store}/search/hybrid`), RowDescription cache, a faster parse-cache hash, compaction throttle. Numbers and the deltas.
Read more →hnsw_rs, No faiss, No BSLEight phases (P0–P7) of the in-house, from-scratch persistent PQ-HNSW: Malkov & Yashunin graph, RocksDB persistence under __vidx:, PQ + ADC + two-stage rerank, online deletes with neighbour repair, filtered KNN evaluated during traversal, F32/F16/I8 rerank dial. Apache 2.0 end-to-end, no third-party graph crate.
HeliosDB Full v7.5/v7.6 ships chat proxy, vector store, embeddings, transcribe, SSE, agents, documents, and per-tenant quotas — all behind one Postgres-wire instance. The AI surface that collapses Pinecone + Qdrant + an inference proxy into the engine.
Read more →End-to-end Rust walkthrough: fastembed BGE-Small embeddings, PersistentVectorIndex::create_with_pq, exact two-stage rerank, online deletes with remove() + compact(), restart proof, and the updated Nano vector skill for Claude Code and Codex.
`heliosdb-proxy install skills` unpacks an embedded skill bundle into `~/.claude/skills` and `~/.codex/skills` — 22 operator playbooks covering pool tuning, failover rehearsal, time-travel replay, plugin signing, and PG-12→17 upgrade orchestration.
Read more →v3.32.0 lands the persistent PQ-HNSW vector index (opt-in via vector-persist); v3.32.1 fixes the PQ default that controls recall on every callsite, including SQL CREATE INDEX … USING hnsw. Crash-recoverable, ~16× less resident RAM at < 0.5% recall drop, online deletes with neighbour repair, filtered KNN. Apache 2.0, in-house, no third-party graph crate.
Design notes on scoping each statement's scratch memory to the request and watching per-query arena bytes. Editor's note: in Lite 3.6.0 and Full 8.3.0, per-request arenas and the helios_query_stats view are planned.
Design notes on PREPARE COMPILED, a fingerprint-keyed LRU plan cache shared across connections. Editor's note: PREPARE COMPILED is planned for HeliosDB Lite and Full; Lite 3.6.0 and Full 8.3.0 serve standard PostgreSQL wire clients.
v7.9.0→v8.0.0 retired RocksDB: native LSM, custom WAL, content-addressed CAS. Editor's note: Full 8.3.0 opens 7 database wire protocols plus an HTTP gateway, and branching and time travel for Full are planned.
Read more →Eight Nano releases closed the gaps most evaluators hit in the first half-hour: SCRAM-SHA-256 with proper GS2 parsing, CREATE DATABASE + StartupMessage validation, information_schema completion, real \d / \l / \dt, keyset row-constructor pagination, and FK enforcement on the index path. The unglamorous wire-protocol work that turns 'interesting' into 'drop-in'.
May 2026 design notes on graph traversal, BM25 ranking, hybrid search and RRF/MMR fusion as SQL functions for Lite. Editor's note: in Lite 3.6.0 these are planned, as are PREPARE COMPILED and an MCP server; Lite 3.6.0 serves PostgreSQL wire and REST.
FK validator now reads-its-own-writes inside a transaction (PG-compatible). Cross-process `INSERT … ON CONFLICT` no longer duplicates. MCP `result_cache` stats published in `helios/info` and `GET /mcp/info`. Three fixes that close the daily-workflow silent-fail matrix.
Read more →Every HeliosDB product (Nano, Lite, Full, HeliosProxy) ships a `.claude/skills/` catalogue plus an `AGENTS.md` aggregate, installed with one command per product (a script for Nano and Full, `heliosdb-lite skills install` with a target directory, `heliosdb-proxy install skills`). Verb-by-verb tool catalogues for Claude Code, Codex CLI, and any MCP-aware agent.
Read more →Internal tests of HeliosDB Lite builds labelled 3.6.1 to 3.6.3 (Lite has no release tags; its public version is 3.6.0): window functions −25%, INSERT throughput −17 to −31%, CTE −13%, time-travel index 39 000× over linear scan. Four silently-red regression tests now green.
Read more →Stage-driven upgrades with shadow-execution validation, chaos rehearsal, and time-travel replay rollback. The wedge into the PG-12-EOL migration market.
Read more →Real WASM plugin runtime, signed-plugin distribution via OCI, time-travel replay, zero-downtime PG major-version upgrade orchestrator, built-in admin console. Largest release in the project's history.
Read more →Three HTTP transports, 16 core auto-registered tools in builds with the mcp-endpoint feature, streaming progress events, single-shot discovery. Claude Code, Cursor, Continue, Codex, and Aider connect to HeliosDB Nano with no glue server.
Read more →WITH CONTEXT Clause and Why RAG Belongs in SQLWITH CONTEXT expands SQL seed rows over a graph in Nano. Editor's note: the ingest adapters are Rust API methods, and applying RERANK BY in SQL is planned.
Read more →Internal tests of a HeliosDB Lite build labelled 3.6.1 vs 3.6.0 (Lite has no release tags; its public version is 3.6.0): 33 of 35 categories match or improve. DELETE point −36% (1000× on a 10K-row table), IN subquery −47%, DISTINCT −45%, JOINs ~−29%.
Read more →A patch release focused on the request path: 28–37% faster pool checkout, 3.1× faster metrics reads, zero-allocation protocol parsing, and concurrent L1 cache hits.
Read more →The AST is a table and LSP calls are SQL in Nano. Editor's note: the index is CREATE AST INDEX src_ast ON repo (content) USING tree_sitter(lang).
Part 3 of the BaaS series. Three callers, three needs: long-lived servers want wire protocol, serverless workers want REST, agents want MCP. Why HeliosDB exposes all three on the same engine and how to pick per caller.
Read more →Five weeks, eight minor versions, three new tracks. Code-Graph (CREATE EXTENSION hdb_code), Graph-RAG (WITH CONTEXT clause), native MCP server. In-process fastembed, Docling ingestion, Drizzle/Prisma/TypeORM compat.
Part 2 of the BaaS series. A hands-on build: provision, schema with VECTOR(1536), seed with embeddings, per-user scoped keys, branch for a feature, time-travel to debug, self-host with the same SQL. No SDK, no backend.
Read more →Part 1 of the BaaS series. SQL portability isn't database portability — drivers, pools, auth, and provisioning leak through. A RESTful BaaS moves the seam to the HTTP boundary. Bootstrap in one POST, deploy anywhere, scale stateless.
Read more →The January 2026 v7.1.0 announcement, with an editor's note: Full 8.3.0 opens 7 database wire protocols plus an HTTP gateway, clustering is planned, and GPU acceleration (NVIDIA CUDA) is available in builds on request.
Read more →A deep dive into Product Quantization in HeliosDB Nano. Editor's note: the recall-safe default saves about 16x resident RAM (internal tests, Nano 3.32.1 vs an unquantized index, see benchmarks.html); 384x costs most of the recall.
Read more →How zero-cost copy-on-write branching at the storage layer changes the way teams develop, test, and deploy schema changes.
Read more →Memory safety without garbage collection, fearless concurrency, and zero-cost abstractions. Here's why Rust was the only choice.
Read more →November 2025 internal tests comparing a HeliosDB build with PostgreSQL 16 across OLTP, OLAP, vector and multi-tenant workloads. Editor's note: not reproduced on current releases; current figures are on benchmarks.html.
Read more →A step-by-step guide to building Retrieval-Augmented Generation pipelines using HeliosDB's native vector search and SQL.
Read more →HeliosProxy: the connection router whose flagship feature is Transaction Replay, plus multi-tier caching, a GraphQL gateway, WASM plugins and query routing.
Read more →The 2025 design for client-side encryption. Editor's note: HeliosDB Nano ships zero-knowledge encryption; ZKE for Lite is planned.
Read more →Real-world patterns for using AS OF queries in production: audit compliance, debugging, data recovery, and A/B testing analysis.
Read more →Row-Level Security, per-tenant resource quotas, and isolation patterns for building SaaS applications on HeliosDB.
Read more →Deploy a full-featured database at the edge: IoT sensors, mobile apps, and offline-first architectures with one self-contained binary (about 36-44 MB in the 4.41.0 release, depending on platform).
Read more →Why vector search is just the beginning. Our vision for databases that understand, predict, and adapt to your application needs.
Read more →