Authentication, sharded tables, conversational search and faster vector search

Manticore Search 27.1.5 has been released. This release brings built-in authentication and authorization, sharded tables, conversational search, faster HNSW builds, better faceting and aggregations, and a long list of fixes across KNN, replication, protocol compatibility and other areas.

This post is a catch-up for everything shipped from 25.0.1 through 27.1.5.

Upgrade Notes

Please review these before upgrading:

• 27.0.0 adds built-in auth/authz, and enabling it changes access assumptions. Auth is not enabled by default, but once you enable it, anonymous access no longer works. Roll it out in stages: upgrade remote agents and replication peers first, then upgrade the masters that query or manage them, and enable auth only after the whole topology is on the new version. Distributed remote-agent and replication-related operations also need matching stored auth data across the participating daemons. A successful JOIN CLUSTER replaces the joining node’s local auth data with the donor cluster’s auth data. (Issue #2833 , PR #3648 )
• 26.0.0 changed replication storage layout. Incoming replicated tables now live under the normal data_dir/<table> layout instead of the cluster path. If you run replication clusters with a custom path, you may need to move or re-synchronize replicated tables after upgrade. Downgrade is only safe before the new layout is adopted. (Issue #4431 , PR #4598 )
• If you manage MCL separately from the daemon, upgrade it together with Manticore. This release line moves through several MCL updates, from vector-performance work to multithreaded HNSW builds and later stability fixes. Mixing an older library with a newer daemon is not recommended. (25.2.0 , 25.15.0 , 26.0.3 , 26.3.2 , 27.1.0 )

Highlights

Built-in authentication and authorization

Manticore now supports users, passwords, bearer tokens, and fine-grained permissions across MySQL, HTTP/HTTPS, distributed remote agents, and replication-related operations. This makes access control a first-class part of the product instead of something that always has to be handled outside the database.

Sharded tables

Manticore can now create and manage sharded tables , distribute inserts across shards, and handle more of the surrounding lifecycle in one place. That makes larger write-heavy deployments easier to operate and reduces the amount of sharding-specific logic that has to live outside the engine.

Conversational search

This release adds conversational search to Manticore Search. It is exposed through CREATE CHAT MODEL and CALL CHAT , so you can ask questions over an existing vectorized table instead of building a separate retrieval layer around the same data.

Under the hood, Manticore Search runs KNN on a FLOAT_VECTOR field, builds LLM context from that field’s from='...' source columns, keeps conversation history by conversation_uuid, and returns both the answer and the supporting sources. If you already keep embeddings in Manticore, this makes document Q&A and support-style assistants much easier to wire up.

Faster vector builds and KNN improvements

Vector search kept improving throughout this cycle.

Manticore improved KNN performance, added local ONNX embeddings support, sped up ONNX inference, and then made HNSW build and rebuild work much faster with multithreaded index construction.

A few important steps in that work:

• 25.1.0 improved KNN distance calculation and AVX-512 loading.
• 25.2.0 added local ONNX embeddings support in MCL and improved vector-search performance further.
• 25.14.0 and 25.15.0 added multithreaded HNSW builds together with the required library support.

The biggest practical improvement here is a much faster auto-embedding and shorter build and rebuild time for large vector tables. Initial KNN builds, chunk merges, and ALTER TABLE ... REBUILD KNN are all affected.

Better faceting and aggregations

Faceting and aggregations also became more useful.

facet_filter_mode makes it easier to build e-commerce-style filters that preserve selected, available, and unavailable buckets under active filtering.

On the analytics side:

• date_histogram() gained time_zone and offset
• Opensearch dashboards support
• Manticore added statistical aggregations such as percentiles, percentile_ranks, and mad

Other Notable Improvements

This release line also includes several smaller but useful additions:

• searchd --check validates configuration before startup without side effects.
• EXIT CLUSTER lets a node leave a replication cluster online without restarting.
• dict=keywords_32k makes it possible to index very long machine-generated tokens such as hashes and message IDs without silent truncation.
• The built-in Ukrainian lemmatizer expands native morphology support for Ukrainian text search.
• Systemd Type=notify improves startup and shutdown supervision.
• searchd process under systemd management now logs to systemd journal
• JOIN queries now support explicit left-table column prefixes.
• OpenSearch Dashboards support.
• manticore-load gained multi-query support.

Bug Fixes

This release line also includes 65 changelog-listed fixes. The latest follow-up releases added a few more worth calling out:

• 27.1.5 fixed a crash when fetching columnar float_vector attributes.
• 27.1.4 fixed ALTER TABLE ... RECONFIGURE and SHOW CREATE TABLE for one-way upgrades from dict="keywords" to dict=keywords_32k.
• 27.1.3 updated Buddy to 4.0.1 and tightened Queue-plugin mutation permission handling under auth.
• KNN-by-doc_id queries now preserve offset and max_matches correctly.
• KNN rescoring order was fixed, so explicit ORDER BY tie-breakers work again.
• Hybrid fused queries with GROUP BY on columnar tables stopped crashing.
• Replication and node-rejoin crash paths were cleaned up further.
• Binary MySQL protocol behavior was fixed in 25.12.1, which matters for integrations that expect real client compatibility.
• Fluent Bit bulk-ingest interoperability was fixed, preventing successful responses from being replayed as duplicate inserts.
• 27.1.2 fixed sql_attr_multi handling for plain indexes built from multiple source blocks.

For the complete list, see the changelog .

Need help or want to connect?