SQL
MariaDB diagnostics, autonomous.
dbexpertAI covers MariaDB with dedicated detection paths — InnoDB health, Galera cluster state, replication, thread-pool behavior, and optimizer regressions — delivering root cause and resolution steps every 90 seconds.
MariaDB is not just MySQL with a different name: Galera clustering, a different optimizer lineage, the thread-pool scheduler, and divergent replication features (multi-source, semi-sync behavior) all fail in MariaDB-specific ways. The path library treats it as its own engine, sharing InnoDB fundamentals with MySQL where they genuinely overlap.
Failure modes we detect
How MariaDB actually fails in production
Galera flow control stalls the whole cluster
In a Galera cluster, one slow node throttles everyone: when its receive queue passes the flow-control threshold, the cluster pauses writes cluster-wide. Applications see 'the database froze'; the cause is one node's disk, one DDL, or one hot backup on one box.
Certification conflicts under write concurrency
Optimistic certification means two nodes accepting conflicting writes roll one back at commit time with a deadlock error the application never expects. Hot rows written on multiple nodes produce rollback storms that look like random deadlocks.
State transfers (SST/IST) at the worst moment
A node rejoining after a blip can trigger a full SST — saturating the donor node and, with the wrong donor selection, degrading the cluster you were trying to heal. Whether IST was possible (and why it was not) is evidence detection paths capture.
InnoDB fundamentals, MariaDB flavor
Lock chains, purge lag, buffer-pool pressure, and redo sizing behave essentially as in MySQL — with MariaDB-specific variables and defaults. Paths shared with the MySQL library run with MariaDB-correct evidence sources.
Optimizer divergence from MySQL
MariaDB's optimizer took its own path (different join strategies, histogram support, split-materialization behavior). Query plans that behaved on MySQL can regress after a MariaDB migration or version bump — a diagnosis that requires knowing which optimizer you're actually on.
A real diagnosis, walked through
What a MariaDB detection path delivers
// symptom
Every application region reports intermittent multi-second write freezes; reads are unaffected.
// diagnostic evidence
- wsrep_flow_control_paused = 0.31 — the cluster spent 31% of the last interval paused
- node db-3: wsrep_local_recv_queue_avg far above threshold; other nodes near zero
- db-3 I/O utilization saturated by a concurrently running logical backup
- no lock waits, no long transactions — single-node cause confirmed
// root cause
Cluster-wide write freezes are Galera flow control triggered by db-3, whose nightly backup saturated its disk and let its receive queue grow past gcs.fc_limit. One node paused the whole cluster.
// resolution steps
- Move the backup to a dedicated (or desynced: wsrep_desync=ON) node so it cannot trigger flow control
- Raise gcs.fc_limit modestly to absorb transient queue growth on I/O-constrained nodes
- Add per-node receive-queue watch so the next slow node is named before it pauses the cluster
Composite walkthrough drawn from the MariaDB detection-path library — the shape of every diagnosis: symptom, evidence, root cause, resolution. Yours will name your tables.
Path library coverage
What the MariaDB detection paths watch
- Galera: flow control, certification conflicts, quorum/partition state, SST/IST behavior, donor selection
- InnoDB: lock waits, purge/history health, buffer pool, redo sizing (MariaDB-correct variables)
- Replication: multi-source lag, semi-sync stalls, GTID consistency
- Query layer: optimizer regressions, histogram staleness, slow-query outliers
- Thread pool: scheduler saturation and priority starvation
- Capacity and configuration drift across cluster nodes
Connect your MariaDB — free.
Read-only agent, on-prem, 90 seconds to your first diagnosis. Your first database is free for life — make it the MariaDB that hurts.
First database free. For life.
Runs on Windows, macOS and Ubuntu — inside your network, read-only. Enterprise fleet or want humans in the loop? Talk to us.