Infino Blog: Engineering notes and company news
Engineering
- Billions of ids per second, no coordinator needed How Infino’s 128-bit Snowflake-style _id lets writers mint billions of ids per second without coordinating, keeps them ordered by time, and stores each copy in a little over a byte.
- The hidden index: why search can’t just be exact match anymore An agent asking for “spicy noodles” has to find a dish called Korean Ramyun. Why search has to match on meaning, what an embedding is, and why Infino keeps its vector index hidden inside the same Parquet files as the text.
- Infino Cloud adds bring-your-own-bucket and customer-managed keys Two enterprise-grade capabilities for Infino Cloud: host a database in your own S3, GCS or Azure bucket, and encrypt it with a key only you control in your own cloud KMS. Your data stays in your account, and you can revoke access independently of Infino.
- Querying a billion vectors from object storage, on one box A one-billion-vector deep1b index that lives in object storage, queried from a single 16 vCPU / 64 GB machine at about 74 ms and 0.95 recall@10. The latency, the recall, and how the engine gets there.
- What changes when your search index lives on S3? On object storage every random read is a billed GET. How superfiles, the manifest, the term index and byte-range reads keep a search to as few round trips as possible.
- Cost-efficient alternatives for large-scale search and vector databases What sets the cost of search at scale, the alternatives to Elasticsearch and dedicated vector databases, and dedicated engines versus SQL with search extensions.
- What’s the SQLite equivalent for full-text and vector search? Embedded engines that run inside your app with no server, and how five options compare on keyword, vector, hybrid and SQL search.
- Open VDBBench: a fair and open benchmark for vector databases The Zilliz-hosted leaderboard stopped re-running competitors in June 2025. We forked the harness and ran 19 engines on current versions, one Azure VM size, every row dated.
- Choosing a database for AI agents: memory, retrieval and state compared What agent memory needs from its store, and how seven common options compare on hybrid search, SQL, deployment and storage format.
- You don’t need two databases for hybrid search A vector database next to a search cluster is two copies of the corpus and a sync loop. One table, both indexes, one ranked result set.
- Where keyword search and vector search each fail The same queries through BM25 and vector search over 10,003 real banking questions. Each retriever fails where the other is strong.
- One vector engine from a million vectors to a billion Infino serves vectors from an in-RAM graph while the corpus is small and from an object-storage index as it grows, in one engine.
- Comparing three approaches to vector quantization: FAISS, turbovec and Infino Three 4-bit indexes at the same 768 bytes per vector: recall within two points, latency 30× apart.
- What is hybrid search? Two retrievers that miss different things, and the rank fusion that merges their lists.
- Why object-storage vectors use OPANN + Sq16, while RAM uses HNSW One durable vector corpus, with a graph in RAM and bounded range reads on object storage.
- Infino's self-driving vector engine You write a table; the engine inverts it in the background and stamps the ANN dials from your data: 0.99 recall@10 on Cohere-10M, with zero knobs.
- Inside a Parquet superfile Why the indexes live inside Parquet, how the footer addresses them, and where the format still hurts.
- mmap behind the scenes in a database What the kernel does when a database mmaps a file, how page faults and reclaim work, why the classic case against mmap mostly misses append-only storage, and how S3, a disk cache and the page cache stack up in Infino.
- Cutting a ten-term BM25 OR from 74.6 ms to 20.0 ms Common terms defeated MaxScore skipping; a 4096-document window removed the per-document cursor merge.
- The 3× to 10× regression from running decode on tokio’s I/O pool Cold vector search went from 1.1 s to 3.7-11 s when CPU-heavy decode starved object-store I/O.
- The three costs of a cold query on S3 File discovery, connection setup, and cache residency on the first query from a new worker.
- A commit on S3 is one compare-and-swap Stage immutable files in parallel; publish the complete table with one conditional pointer write.