mariadb-core-storage-engines

mariadb-core-storage-engines

Deterministic guidance for picking and operating MariaDB storage engines.

Quick Reference

InnoDB is the right default 99% of the time. It is ACID, row-locked, MVCC-backed, foreign-key-aware, crash-safe, and the only engine to pick for new transactional workloads.

MyISAM lacks transactions, foreign keys, and crash-safety. NEVER use MyISAM for new tables. Use Aria as the MyISAM replacement when you need crash-safety on system-like data.

ColumnStore is analytical-only. It cannot serve OLTP, runs as a separate distribution, and is column-oriented. NEVER point an OLTP application at ColumnStore.

MEMORY engine loses all rows on server restart. Use it only for ephemeral lookup tables that can be re-populated from a source of truth.

Spider federates across MariaDB backends. Spider's high-availability features were removed in 10.7.5 ; combine Spider with Galera or replication for HA.

CONNECT federates external data sources (CSV, XML, JSON files, ODBC, JDBC, MongoDB). NEVER use CONNECT for high-throughput writes ; it is a federation layer, not a storage layer.

Decision Tree : Which Engine

Need ACID + foreign keys + concurrent writes ?
  YES -> InnoDB (default, 99% of cases)
  NO  -> continue

Internal system table or on-disk temporary table ?
  YES -> Aria (managed by MariaDB itself, do NOT pick Aria for user tables)
  NO  -> continue

Analytical OLAP scans over billions of rows ?
  YES -> ColumnStore (separate distribution, never OLTP)
  NO  -> continue

Horizontally shard a single logical table across many MariaDB nodes ?
  YES -> Spider (combine with Galera or replication for HA)
  NO  -> continue

Read external data files or other DBMS as a SQL table ?
  YES -> CONNECT (CSV / XML / JSON / ODBC / JDBC / MongoDB)
  NO  -> continue

Ephemeral lookup table that can be lost on restart ?
  YES -> MEMORY (HEAP) ; no TEXT / BLOB columns ; default HASH index
  NO  -> InnoDB

Legacy MyISAM table found in your schema ?
  YES -> ALTER TABLE x ENGINE=InnoDB; (NEVER leave it on MyISAM)

ALWAYS state the engine explicitly in CREATE TABLE so the choice is visible in code review.

Core Patterns

Pattern 1 : InnoDB as the default

-- 10.6+ : explicit engine, charset, row format
CREATE TABLE orders (
  id BIGINT UNSIGNED NOT NULL AUTO_INCREMENT PRIMARY KEY,
  customer_id BIGINT UNSIGNED NOT NULL,
  total_cents BIGINT NOT NULL,
  created_at TIMESTAMP(6) DEFAULT CURRENT_TIMESTAMP(6),
  FOREIGN KEY (customer_id) REFERENCES customer (id)
) ENGINE=InnoDB
  ROW_FORMAT=DYNAMIC
  DEFAULT CHARSET=utf8mb4
  COLLATE=utf8mb4_uca1400_ai_ci;

ALWAYS specify ENGINE=InnoDB even though it is the default. ALWAYS specify ROW_FORMAT=DYNAMIC so off-page storage of long variable-length columns is unambiguous. ALWAYS specify DEFAULT CHARSET=utf8mb4 ; stock 10.6 and 10.11 still ship with latin1 as the global default.

Pattern 2 : Engine swap from MyISAM to InnoDB

-- Audit a legacy schema
SELECT table_schema, table_name, engine
FROM information_schema.tables
WHERE table_schema = DATABASE()
  AND engine NOT IN ('InnoDB', 'Aria')
ORDER BY table_schema, table_name;

-- Convert one table, version-safe
ALTER TABLE legacy_orders ENGINE=InnoDB,
  ALGORITHM=COPY, LOCK=SHARED;

NEVER use ALGORITHM=INSTANT for an engine swap ; engine change always requires a full table rewrite. NEVER swap an engine on a busy table during peak hours ; the table is fully locked for SHARED access during the COPY rewrite.

Pattern 3 : MEMORY for ephemeral lookup

-- 10.6+ : ephemeral session-state, lost on restart
CREATE TABLE session_lookup (
  session_token CHAR(64) NOT NULL PRIMARY KEY,
  user_id BIGINT UNSIGNED NOT NULL,
  expires_at INT UNSIGNED NOT NULL,
  INDEX ix_user (user_id) USING HASH
) ENGINE=MEMORY
  MAX_ROWS=100000;

NEVER store any value of record in MEMORY tables ; a restart wipes them. NEVER use TEXT or BLOB columns ; MEMORY rejects them. ALWAYS size MAX_ROWS to bound memory usage.

Pattern 4 : Aria for system-like crash-safe storage

-- 10.6+ : crash-safe but non-transactional ; use only for system-like tables
CREATE TABLE audit_archive (
  id BIGINT UNSIGNED NOT NULL AUTO_INCREMENT PRIMARY KEY,
  event_at TIMESTAMP(6) NOT NULL,
  payload TEXT NOT NULL
) ENGINE=Aria
  ROW_FORMAT=PAGE
  TRANSACTIONAL=1;

ALWAYS know that TRANSACTIONAL=1 on Aria means "crash-safe", NOT "supports BEGIN / COMMIT". Aria has no MVCC and uses table-level locking. NEVER pick Aria for concurrent user data.

Pattern 5 : Spider sharding shell

-- 10.6+ : on the shard-router node
INSTALL SONAME 'ha_spider';

CREATE SERVER backend1
  FOREIGN DATA WRAPPER mysql
  OPTIONS (HOST '10.0.0.11', DATABASE 'shop', USER 'spider', PASSWORD '...', PORT 3306);

CREATE TABLE order_shard (
  id BIGINT UNSIGNED NOT NULL,
  customer_id BIGINT UNSIGNED NOT NULL,
  total_cents BIGINT NOT NULL,
  PRIMARY KEY (id, customer_id)
) ENGINE=Spider
  PARTITION BY HASH(customer_id) (
    PARTITION p0 COMMENT = 'srv "backend1", table "orders_p0"',
    PARTITION p1 COMMENT = 'srv "backend2", table "orders_p1"'
  );

ALWAYS pair Spider with Galera or replication on each backend for HA ; Spider's own HA was removed in 10.7.5. NEVER deploy Spider over a slow WAN without spider_connect_timeout and spider_net_read_timeout tuned.

Pattern 6 : CONNECT to an external CSV

-- 10.6+ : federate a CSV file as a SQL table
CREATE TABLE product_feed (
  sku VARCHAR(64) NOT NULL,
  name VARCHAR(255) NOT NULL,
  price_cents BIGINT NOT NULL
) ENGINE=CONNECT
  TABLE_TYPE=CSV
  FILE_NAME='/var/feeds/products.csv'
  SEP_CHAR=','
  QUOTED=1
  HEADER=1;

NEVER write heavy transactional traffic to a CONNECT table ; it has no real storage engine of its own. NEVER let untrusted user input control FILE_NAME ; that is a file-system read primitive.

Pattern 7 : ColumnStore as an analytical sidecar

-- ColumnStore Enterprise / Community : analytical-only
CREATE TABLE sales_fact (
  sale_date DATE NOT NULL,
  store_id INT NOT NULL,
  sku VARCHAR(64) NOT NULL,
  qty INT NOT NULL,
  total_cents BIGINT NOT NULL
) ENGINE=ColumnStore;

ALWAYS treat ColumnStore as a separate distribution that is loaded from InnoDB via batch ETL or change-data-capture. NEVER point an OLTP application at ColumnStore for row-by-row UPDATE / DELETE ; performance is by design slow for that pattern.

Engine Feature Matrix (Quick Reference)

Engine	ACID	Locking	FK	MVCC	Crash-safe	Full-text	Spatial	Default
InnoDB	yes	row	yes	yes	yes (redo log)	yes (10.0.5+)	yes	YES (default since 10.2)
Aria	no	table	no	no	yes (PAGE row format)	yes	no	no
MyISAM	no	table	no	no	no	yes	yes	no (legacy)
ColumnStore	no	block	no	no	yes (snapshot)	no	no	no (separate dist)
Spider	depends on backend	depends	depends	depends	depends	depends	depends	no
CONNECT	no	none (read-mostly)	no	no	no	no	no	no
MEMORY	no	table	no	no	no (lost on restart)	no	no	no

Full method-level matrix in references/methods.md.

Cross-Engine Query Limits

ALWAYS know that a multi-statement transaction touching both InnoDB and MyISAM rows commits the MyISAM writes immediately and only the InnoDB writes participate in the transaction. NEVER mix engines inside a transactional boundary that needs all-or-nothing semantics.

ALWAYS know that foreign keys are forbidden in partitioned tables and forbidden across engines ; a FK can only point from one InnoDB table to another InnoDB table.

ALWAYS know that JOINing an InnoDB table to a CONNECT or Spider table forces a row-by-row pull through the federation layer ; the optimiser has no statistics on the remote side.

Anti-Patterns (See references/anti-patterns.md)

1. MyISAM for orders / users (no transactions, crash-loss)
2. ColumnStore for OLTP (write amplification, latency)
3. MEMORY for "cache" without considering restart-loss
4. Mixing engines in a JOIN that crosses transaction boundaries
5. Spider over slow WAN without connection pooling
6. Aria for high-concurrency writes (still table-locking semantics)
7. InnoDB with innodb_file_per_table=OFF (shared tablespace bloats)
8. ENGINE= without ROW_FORMAT specification on tables with long columns

Version Notes

• MariaDB 10.2 : InnoDB became the global default ; DYNAMIC became the default ROW_FORMAT.
• MariaDB 10.3 : InnoDB instant ADD COLUMN at the end of the table.
• MariaDB 10.4 : InnoDB instant ADD COLUMN at any position, instant DROP COLUMN.
• MariaDB 10.6 : atomic DDL via binary log ; ignored indexes ; Aria continues as system-table storage.
• MariaDB 10.7.5 : Spider's built-in high-availability features removed (MDEV-28479).
• MariaDB 11.6 : default charset migrated from latin1 to utf8mb4 ; affects new-table defaults across all engines.

Reference Links

• InnoDB overview : https://mariadb.com/kb/en/innodb/
• InnoDB row formats : https://mariadb.com/kb/en/innodb-row-formats/
• Aria overview : https://mariadb.com/kb/en/aria-storage-engine/
• MyISAM overview : https://mariadb.com/kb/en/myisam-storage-engine/
• ColumnStore : https://mariadb.com/docs/columnstore/
• Spider overview : https://mariadb.com/kb/en/spider-storage-engine-overview/
• CONNECT overview : https://mariadb.com/kb/en/connect/
• MEMORY overview : https://mariadb.com/kb/en/memory-storage-engine/

See references/methods.md for the complete per-engine feature matrix, references/examples.md for full working examples per engine, and references/anti-patterns.md for in-depth anti-pattern walkthroughs.