Manticore memory usage and performance comparisons
Manticore memory usage grows linearly with data imports, such as when a 3.3G storage size causes pod memory to rise to 2.8G. The system manages high-cardinality datasets using distinctprecisionthreshold and offers hybrid search through RRF fusion methods.
Manticore memory usage grows linearly as data continues to import. In a test of a neuralmap_columar table with knn_type='hnsw', a 1.7G storage size caused 600M of pod memory usage. When the user increased the data to 3.3G, the pod memory rose to 2.8G. Restarting the pod dropped memory usage to 1.1G. The max_matches parameter controls the number of matches the server retains in RAM for each table. If you increase the max_matches setting to 1,000,000, the searchd process must allocate and initialize a massive matches buffer for every query, which significantly impacts the RAM and CPU usage of the server. The default value for this setting is 1,000. Meilisearch users report high memory consumption and difficulty tweaking ranking with BucketSort. While Manticore provides an rt_mem_limit for tables, some users find that the indexer does not always honor the mem_limit. One user noted that min_infix_len for a dictionary of keywords does not follow the mem_limit and can overflow box RAM. The neuralmap_columar table specifically used id bigint, file_path text, camera text, label_unit_id string indexed attribute, city text, labels text, scenarios text, embedding float_vector knn_type='hnsw' knn_dims='1280' hnsw_similarity='COSINE', bbox float_vector, and created_at timestamp with a limit of 5368709120. When Manticore cannot determine the exact count of matching documents, the total_relation field in the query meta information shows gte instead of eq. How do developers manage the memory spike during large-scale ingestion?
Deployment and installation methods
Manticore 25.0.0 bundles Manticore Buddy, the Manticore Columnar Library, backup, load, and Galera into a single package. This bundle removes the need to install separate manticore-* packages. The software supports CentOS 8, RHEL 8, Oracle Linux 8, AlmaLinux 9, and AlmaLinux 10. The one-line installer detects if the system uses APT, YUM/DNF, or Homebrew to automate the setup. This installer checks that the CPU architecture is x86-64 or ARM64. For those of you already managing containerized workloads, you know that the manticoresearch/manticore Docker image provides a reliable environment. The image provides default listeners on port 9306 for MySQL connections, port 9308 for HTTP, and port 9312 for binary protocols. To run a two-node cluster, you can use a docker-compose.yml file to define the instances. The image includes libraries to index data from MySQL, PostgreSQL, XML, and CSV files. You can use docker exec -it manticore mysql to connect to the database. To build all plain tables on startup, you set the CREATE_PLAIN_TABLES environment variable to 1. The installer for Debian or Ubuntu adds the APT repository, updates metadata, installs the package, and starts the service. On macOS, the installer uses Homebrew to perform the installation and starts the service via brew services.
| Protocol | Default Port |
|---|---|
| MySQL | 9306 |
| HTTP | 9308 |
| Binary | 9312 |
Search accuracy and scaling parameters
High-cardinality datasets in Manticore require adjustment of the distinct_precision_threshold setting. Values between 500 and 15500 are accepted. When the setting exceeds 0, the system uses a hash table to load counts, but it switches to a HyperLogLog data structure if the hash table grows too large. This probabilistic algorithm maintains a fixed maximum memory usage per group but introduces a tradeoff in count accuracy. The accurate_aggregation setting ensures group count and aggregate accuracy in RT tables by forcing groupby searches to run in a single thread. This single-threaded approach prevents accuracy loss when the number of unique values in a groupby attribute is high, but it makes searches slower. Manticore also supports hybrid search result fusion through the fusion_method='rrf' option. This option runs MATCH and KNN as independent sub-queries and fuses the results by rank. Users can use boolean_simplify=1 to speed up queries or expand_keywords=1 to expand keywords with exact forms. The max_matches_increase_threshold setting helps Manticore adjust max_matches during groupby queries. For IDF computation, idf=plain avoids keyword penalization, while idf=tfidf_unnormalized prevents IDF drift across queries. The system also supports the jieba_mode setting for Chinese segmentation. The low_priority=1 option reschedules jobs 10 times less frequently than normal priority queries.