Manticore Search

Manticore Search is a database designed for search rather than a general store with search bolted on. The project positions it against Elasticsearch, and its published benchmarks report faster completion on several workloads while using a fraction of the CPU cores and RAM.

It is written in C++, starts quickly, and holds around 40 MB of resident memory for an empty instance. SQL is the native syntax and any MySQL client can connect, with an HTTP JSON API and Elasticsearch-compatible writes covering everything else. Query execution is multi-threaded and parallelized across cores.

Retrieval and storage can be matched to the shape of the data.

• Hybrid search: full-text and vector retrieval combine in one query for better relevance.
• Conversational search: vectorized tables answer questions using KNN retrieval, conversation history, and LLM-backed responses.
• Row-wise and columnar storage: row storage suits datasets that fit in memory, columnar handles the ones that do not.
• Automatic secondary indexes: a PGM index maps indexed keys to memory locations, and a cost-based optimizer picks plans from data statistics.
• Replication and sharding: virtually synchronous multi-master replication through Galera, load balancing, and sharded tables across nodes.

Writes are not fully ACID but support isolated transactions and binary logging, and backups run through manticore-backup, the SQL backup command, or S3-compatible storage. Authentication covers MySQL, HTTP, remote agents, and replication with users, bearer tokens, and fine-grained permissions. Official clients ship for several languages, and data synchronizes from MySQL, PostgreSQL, ODBC, XML, and CSV sources.