Oracle outlines MySQL’s vector-search work, with native indexes still in development
The development update targets similarity searches alongside ordinary database filters. An earlier architecture proposal shows two storage approaches—and no final decision.
MySQL Community Architect Henrik Ingo’s November 16 update describes vector search as work in progress: cosine distance is planned, while combining vector retrieval with SQL filters remains a target rather than a released capability. A separate architecture issue, opened August 4, leaves unresolved how vector indexes should be integrated into InnoDB so they work with updates, transactions and crash recovery. The choice matters to developers evaluating MySQL for AI retrieval, but the update provides no performance or memory benchmarks and does not settle the storage design.
01
The architecture issue compares an InnoDB-managed vector page layout used in Oracle’s proof of concept with a hidden InnoDB sub-table approach associated with Google.
02
The project discusses HNSW for approximate nearest-neighbor retrieval and quantization to reduce index storage precision; neither is presented as a measured MySQL improvement.
03
The architecture issue identifies IVF as an initial reference algorithm but leaves the design decision and rationale marked TBD.
Oracle is outlining how MySQL could retrieve similar information for AI applications without dropping the filters used in ordinary database queries. But the November 16 development update from MySQL Community Architect Henrik Ingo describes work in progress, not an available release: cosine distance is planned, and combining vector search with SQL conditions remains a goal.
The update explains both the intended search behavior and the engineering choices behind it. An earlier MySQL architecture issue supplies a useful distinction: developers are evaluating how to build vector indexes into InnoDB, MySQL’s storage engine, rather than treating the choice of search algorithm as the whole project.
Similarity search with ordinary SQL conditions
A vector is a list of numbers. In AI applications, those numbers can represent content, allowing a database to look for nearby vectors as a way to retrieve similar material. Ingo describes the basic retrieval task as sorting records by their distance from an input vector and returning a requested number of nearby results.
The project plans to implement cosine distance, one way to measure how close vectors are. Its more demanding target is hybrid search. Here, that means combining a vector-search parameter with other SQL conditions—not simply returning the closest vectors in isolation. SQL is the language used to query database tables.
Ingo sketches several possible depths of integration. The query optimizer, which chooses how to execute a request, might weigh a vector index against other indexes. A more ambitious structure could combine vectors, numeric values and words within the index itself. These are possible designs, not announced capabilities.
Faster retrieval need not mean an exact ranking
The update also explains why vector search often accepts approximate answers. Finding the exact nearest matches can require scanning the entire dataset. Approximate nearest-neighbor methods instead seek nearby candidates without guaranteeing the precise ranking. Ingo argues that, for many AI uses, a small difference near the end of a result set is less important than retrieving useful nearby material.
Editorial illustration for Oracle outlines MySQL’s vector-search work, with native indexes still in development.
One method discussed is HNSW, short for Hierarchical Navigable Small World. It connects vectors in a layered graph. A search starts among a small sample of points, then moves through increasingly dense layers toward nearby candidates. The hierarchy provides a route through the data rather than comparing every stored vector with the query.
Ingo also describes quantization: storing index values at lower numerical precision to reduce memory use. That is another tradeoff between compact storage and accuracy. The discussion explains available techniques; it does not establish a measured speed or memory improvement for the MySQL project.
The storage decision is still open
The architecture issue, opened August 4, gives the project’s earlier design context. In the secondary-index proposal, vector pages use InnoDB’s normal memory buffering, disk flushing and recovery logging. Row changes become localized index changes, rather than requiring the complete index to be loaded or rewritten for each change.
That design returns candidate row references for MySQL’s normal visibility checks and result processing. The sub-table proposal takes a different storage route but also seeks to reuse existing database guarantees. Its design describes transaction consistency, the ability to read one’s own writes, rollback support and recovery after a crash.
The issue explicitly separates architectural integration from algorithm selection, using IVF as an initial reference algorithm. Its decision and rationale remain marked TBD. For developers, the unresolved work is therefore not just which search method wins: it is how that method stays integrated with database updates, transactions and recovery.
Reader comments
Newest comments first. Replies stay oldest first.