Mariadb Core Replication Model

mariadb-core-replication-model

Architectural reference for MariaDB replication. Covers the three replication shapes, the MariaDB-specific GTID format, binary log formats, the parallel-applier model, and multi-source replication. Topology setup details and recovery procedures live in references/.

Quick Reference

• Async is the default. Primary writes the binary log, replica I/O thread fetches events into the relay log, replica SQL thread (or parallel workers) applies them. The primary does NOT wait for replicas. RPO depends on replication lag.
• Semi-sync is built into the server from MariaDB 10.3+. NEVER install a separate plug-in on 10.3 or later. Enable via rpl_semi_sync_master_enabled=ON on the primary plus rpl_semi_sync_slave_enabled=ON on the replica. Default rpl_semi_sync_master_timeout is 10000 ms ; on timeout the primary reverts to async until a replica re-ack.
• Parallel-applier default mode is optimistic since MariaDB 10.5.1+. Older 10.5.0 and earlier used conservative. Set worker pool with slave_parallel_threads. Both primary and replica MUST be MariaDB 10.0.5 or later.
• MariaDB GTID format is domain-server-sequence (for example 0-1-1234). MySQL GTID format is uuid:seqno. The schemes are NOT interconvertible. Per the KB : "MariaDB can be a replica for a MySQL primary, but MySQL cannot be a replica for a MariaDB primary." Migration between the two with GTID is a one-shot dump-and-load, NEVER continuous GTID-tracked replication.
• Binary log default format is MIXED since MariaDB 10.2.4 (was STATEMENT until 10.2.3). MIXED auto-switches to row-encoding when a statement is unsafe (non-deterministic functions, LIMIT without ORDER BY, certain stored routines). Use ROW explicitly for Galera and for any cluster that cannot tolerate replication drift.
• binlog_row_image default is FULL. Other values : MINIMAL, NOBLOB ; MariaDB 11.4+ adds FULL_NODUP. NEVER set MINIMAL in environments that need point-in-time recovery with full row reconstruction.
• Multi-source replication uses NAMED CONNECTIONS, not FOR CHANNEL. Syntax is CHANGE MASTER 'connection_name' TO .... The FOR CHANNEL syntax is MySQL-only. Connection names are case-insensitive. Use SHOW ALL SLAVES STATUS to see all channels at once.
• Seconds_Behind_Master is unreliable. It measures the timestamp on the SQL thread relative to the event being applied, NOT actual wall-clock lag. Use pt-heartbeat (or the equivalent token-based heartbeat) for accurate lag measurement.

Decision Tree : Which Replication Model

START : you need replication
├── data-loss tolerance is non-zero (analytics, reporting)
│   └── async (default)
│       └── tune : slave_parallel_threads, slave_parallel_mode=optimistic
├── data-loss tolerance is small but non-zero (most OLTP)
│   └── semi-sync (10.3+ built-in)
│       ├── set rpl_semi_sync_master_enabled=ON on primary
│       ├── set rpl_semi_sync_slave_enabled=ON on at least one replica
│       └── understand : on timeout (default 10s) the primary reverts to async
├── ZERO data-loss tolerance + multi-master writes
│   └── Galera Cluster (see mariadb-core-galera-cluster skill)
│       └── NEVER conflate Galera with semi-sync ; different consistency model
├── one replica must pull from multiple primaries
│   └── multi-source (10.0+, named connections)
└── you are migrating off MySQL and want continuous replication
    ├── MySQL primary -> MariaDB replica : OK with binlog-position or MySQL-GTID
    └── MariaDB primary -> MySQL replica : NOT SUPPORTED, dump-and-load only

Decision Tree : Binary Log Format

START : choose binlog_format
├── Galera Cluster node
│   └── ROW (required, no exceptions)
├── async or semi-sync primary with non-deterministic functions (UUID(), NOW() in some contexts, RAND())
│   └── ROW (safest) OR MIXED (auto-switches per statement)
├── async or semi-sync primary with mostly deterministic SQL + tight binlog-volume budget
│   └── MIXED (default since 10.2.4)
├── primary needs point-in-time recovery with full row state
│   └── ROW + binlog_row_image=FULL (default)
└── you think STATEMENT will save space
    └── STOP. Use MIXED. STATEMENT corrupts on unsafe statements ; the savings are an illusion.

Pattern : Async Replication Baseline (10.6+)

# primary my.cnf
[mariadb]
server_id=1
log_bin=mariadb-bin
binlog_format=MIXED
sync_binlog=1                  # crash-safe binlog
expire_logs_days=14

# replica my.cnf
[mariadb]
server_id=2
log_bin=mariadb-bin
binlog_format=MIXED
read_only=ON
slave_parallel_threads=4
slave_parallel_mode=optimistic # 10.5.1+ default, set explicitly for clarity

-- replica side, GTID-based attach (10.0.13+)
CHANGE MASTER TO
  MASTER_HOST='primary.example.com',
  MASTER_USER='repl',
  MASTER_PASSWORD='<secret>',
  MASTER_USE_GTID=slave_pos;

START SLAVE;
SHOW SLAVE STATUS\G

Pattern : Enable Semi-Sync (10.3+)

# primary my.cnf (10.3+)
[mariadb]
rpl_semi_sync_master_enabled=ON
rpl_semi_sync_master_timeout=10000   # default, 10 seconds
rpl_semi_sync_master_wait_point=AFTER_SYNC

# replica my.cnf (10.3+)
[mariadb]
rpl_semi_sync_slave_enabled=ON

ALWAYS confirm both sides are active :

-- on primary
SHOW STATUS LIKE 'Rpl_semi_sync_master_status';   -- expect ON
SHOW STATUS LIKE 'Rpl_semi_sync_master_clients';  -- expect >= 1

-- on replica
SHOW STATUS LIKE 'Rpl_semi_sync_slave_status';    -- expect ON

NEVER assume semi-sync is "always synchronous". On rpl_semi_sync_master_timeout expiry the primary silently reverts to async. Watch Rpl_semi_sync_master_status over time, alert if it flips OFF.

Pattern : Parallel-Applier Tuning (10.5.1+)

# replica my.cnf
[mariadb]
slave_parallel_threads=8                # worker pool size
slave_parallel_mode=optimistic          # default in 10.5.1+
slave_parallel_max_queued=262144        # per-worker queue in bytes
slave_domain_parallel_threads=2         # per-domain cap, 0 = no cap

ALWAYS benchmark slave_parallel_threads > 16 against your actual disk IOPS. Beyond the point where workers wait on flushes, more threads ADD latency rather than removing it.

Pattern : Multi-Source Replication (10.0+, Named Connections)

-- primary A
CHANGE MASTER 'analytics' TO
  MASTER_HOST='primary-a.example.com',
  MASTER_USER='repl',
  MASTER_PASSWORD='<secret>',
  MASTER_USE_GTID=slave_pos;

-- primary B
CHANGE MASTER 'reporting' TO
  MASTER_HOST='primary-b.example.com',
  MASTER_USER='repl',
  MASTER_PASSWORD='<secret>',
  MASTER_USE_GTID=slave_pos;

START SLAVE 'analytics';
START SLAVE 'reporting';

SHOW ALL SLAVES STATUS\G   -- all channels in one go
STOP SLAVE 'analytics';
RESET SLAVE 'analytics' ALL;   -- remove channel permanently

NEVER use FOR CHANNEL <name> syntax in MariaDB. That is MySQL 5.7+ syntax and will fail. NEVER reuse auto_increment_increment and auto_increment_offset ranges across primaries fanning into the same replica ; conflicting ranges cause silent PK collisions on aggregation tables.

Pattern : GTID Configuration

# primary my.cnf
[mariadb]
server_id=1
log_slave_updates=ON          # required on intermediate replicas in chains
gtid_domain_id=0              # 32-bit unsigned, identifies logical stream
gtid_strict_mode=ON           # reject divergent GTIDs proactively

-- attach by GTID
CHANGE MASTER TO
  MASTER_USE_GTID=slave_pos;   -- one of : slave_pos | current_pos | no

-- inspect
SELECT @@global.gtid_domain_id, @@global.gtid_current_pos, @@global.gtid_binlog_pos;
SHOW SLAVE STATUS\G            -- look at Gtid_IO_Pos and Using_Gtid

Pattern : Monitoring Replication Health

-- single-channel async or semi-sync
SHOW SLAVE STATUS\G

-- multi-source
SHOW ALL SLAVES STATUS\G

-- 10.5+ also supports the REPLICA aliases
SHOW REPLICA STATUS\G
SHOW ALL REPLICAS STATUS\G

Critical columns :

NEVER alert on Seconds_Behind_Master > 0 alone. It jumps to 0 between transactions and can mislead. Pair with a real heartbeat table.

Anti-Patterns (Top Picks)

ALWAYS read references/anti-patterns.md for the full set. Top-3 reminders :

1. STATEMENT binlog with NOW(), UUID(), RAND() : silent drift. Use MIXED or ROW.
2. Attempting MariaDB-primary -> MySQL-replica with GTID : not supported per the KB. One-way dump-and-load only.
3. slave_parallel_mode=aggressive in production without exhaustive testing : data corruption risk.

Cross-References

• mariadb-core-galera-cluster : synchronous multi-master, different consistency model.
• mariadb-core-architecture : storage-engine + replication interaction (binlog hooks).
• mariadb-impl-migration-mysql-to-mariadb : GTID-incompatibility recovery patterns.
• mariadb-errors-replication-broken : Last_SQL_Error triage flow.
• mariadb-impl-backup-and-pitr : binlog-based point-in-time recovery.

References

• references/methods.md : full variable reference (rpl_semi_sync_*, slave_parallel_*, gtid_*, binlog_*).
• references/examples.md : end-to-end working examples (async, semi-sync, parallel-applier, multi-source, TLS, monitoring).
• references/anti-patterns.md : 10+ field-tested anti-patterns with code, why-it-fails, and the correct alternative.

Sources

• MariaDB KB : semi-synchronous replication. https://mariadb.com/kb/en/semisynchronous-replication/
• MariaDB KB : parallel replication. https://mariadb.com/kb/en/parallel-replication/
• MariaDB KB : Global Transaction ID (GTID). https://mariadb.com/kb/en/gtid/
• MariaDB KB : Binary log formats. https://mariadb.com/kb/en/binary-log-formats/
• MariaDB KB : Replication and binary log system variables. https://mariadb.com/kb/en/replication-and-binary-log-system-variables/
• MariaDB KB : Multi-source replication. https://mariadb.com/kb/en/multi-source-replication/
• MariaDB KB : SHOW SLAVE STATUS. https://mariadb.com/kb/en/show-slave-status/