MariaDB config generator for WordPress

Store writes, orders, stock and scheduled jobs need transactional safety. HPOS remains an application setting.

As checked on 15 October 2026: 12.3.3 is the latest LTS patch; 13.0.2 is the latest stable rolling release. Stable does not imply long-term maintenance. Test upgrades and plugin compatibility first.

A web server also needs room for PHP workers, Redis, the OS and backup jobs. Dedicated database guidance cannot be applied unchanged to that host.

The browser cannot detect server resources. Enter the host limit, rather than the RAM of this computer. GiB means 1024 MiB.

Replace the estimate with measured peak resident memory. Dedicated database hosts still need a reserve for monitoring and backups.

This is a planning assumption, not a measured OS footprint. Exhausting RAM or swapping can stall checkout even when the database cache is large.

More connections do not create CPU capacity. A thread pool changes scheduling and should be load-tested.

Device labels do not reveal sustainable IOPS or latency. Measure the actual volume under load before increasing I/O budgets.

One active PHP request often uses one connection, but plugins and persistent connections can change that. Include every site sharing this server.

Avoid scheduling large imports, analytics and backups at the same checkout peak. This allowance is not a concurrency limit for WordPress jobs.

A query can allocate more than one temporary table or join buffer. The model is not a maximum-memory calculator.

Keep database credentials in the application or a protected client file. No password is needed by this generator.

A prefix change here does not rename tables. Multisite and plugin tables may use additional names.

This is the main cache to size. On a shared WordPress host leave measured space for PHP, Redis and connections. A larger pool helps only when useful pages would otherwise be read from storage.

Current target releases otherwise reserve a much larger virtual ceiling. Setting a deliberate ceiling makes future manual resizing easier to budget; it does not automatically grow the pool.

Supported by both selected releases on Linux. It does not automatically grow back, and a smaller pool can increase disk reads. Do not use it to justify overcommitting RAM.

Choose a floor below the initial pool and no lower than a practical minimum for the site. Watch buffer-pool misses after shrink events.

This stores page identifiers, not a backup of the database. It reduces cache warm-up work after a clean restart.

WordPress application tables should normally be InnoDB. Aria still has internal uses, so account for its cache alongside InnoDB rather than assigning all RAM to one pool.

Check table engines first. A large MyISAM cache wastes memory when application data is InnoDB; this does not convert existing tables.

Only increase after observing Innodb_log_waits or large transactions. This buffer is RAM; redo file capacity is disk space.

The lower of the two limits applies. Several tables can exist per query and many queries can run together; this is not a global memory cap. Some queries use disk tables regardless.

Large global sort buffers multiply under load and rarely fix a missing index. Review slow-query plans before raising this.

An unindexed join is usually a query or index problem. Raising every session’s join allocation can amplify memory pressure.

This is shared by all sites. Keep administration capacity and watch Max_used_connections, connection errors and CPU; a huge limit can turn overload into an out-of-memory event.

Very short values can interrupt imports or jobs that pause between queries. Persistent connections and plugins may need a different policy.

Increase deliberately for large imports or plugin payloads. It is not reserved in full for every idle connection.

A pool can help many concurrent short requests on CPU-constrained hosts. It can also queue work; test checkout latency and long-running jobs before adopting it.

Tune from queueing, CPU use and tail latency. Connection capacity is still controlled by max_connections.

Watch Threads_created relative to Connections before increasing. This setting is omitted when using the thread pool.

Many sites and plugin tables can need more handles. Check Opened_tables, cache misses, and the systemd file-descriptor limit rather than forcing a huge open_files_limit.

Base this on table count and metadata churn. Large multisite installations need a separate observed workload estimate.

Measure redo generation, checkpoint age and recovery time. This is not MySQL innodb_redo_log_capacity, and the redo file is not an independent backup. Never delete redo files to apply this setting.

This generator keeps full transaction flushing enabled. Storage must honour flush requests; backups and payment reconciliation are still needed.

Do not disable crash protection for an unmeasured speed gain. Filesystem and power-loss behavior still matter.

It is not a disk speed switch or hard total-I/O limit. Shared cloud disks and backups may have less sustained capacity than their headline rate.

Keep foreground checkout I/O responsive. The sizing button offers conservative examples, not hardware measurements.

Rotational disks may benefit from coalesced writes. On SSD, extra neighbouring-page writes usually add work.

Both targets default to unbuffered data I/O. Keep platform behavior unless storage testing supports a change; the deprecated innodb_flush_method is omitted.

A binlog alone is not a backup. Preserve the existing server’s format unless you have a tested migration plan. InnoDB binlogs are opt-in on 12.3+ and change tool compatibility.

1 is a standalone example. Coordinate IDs before joining a replication topology; this generator configures neither replicas nor Galera.

14 days is an example, not a guarantee. MariaDB keeps binary logs for at most 99 days (8,553,600 seconds), so keep older recovery points as full backups. Size disk storage, archive logs off-host and verify the backup chain; expired logs cannot be recovered from this setting.

For the new InnoDB binlog this option is unnecessary and omitted; commit durability is governed by innodb_flush_log_at_trx_commit instead.

Never expose the database directly to the public internet. Private binding still requires firewall rules and least-privilege accounts.

Only RFC1918 IPv4 addresses are accepted here. A private bind does not authenticate clients or configure a firewall.

A different port is not an access control. Match DB_HOST and firewall rules. This control is omitted in socket-only mode.

Hostname-based grants can stop matching. Existing localhost accounts are a separate case; review SHOW GRANTS and test the actual application connection.

Modern MariaDB has automatic TLS, but PHP/mysqlnd trust behavior differs from MariaDB Connector/C. Use a verified CA and matching server name for remote WordPress connections. This browser only takes file paths.

Local TCP without TLS will also be rejected. Confirm WordPress/PHP and maintenance tools before enforcing this on an existing service.

Normal WordPress and WooCommerce requests do not need this. Some bulk-import tools do; enable deliberately and restrict client behavior too.

This does not grant FILE privileges, affect backups, or restrict LOCAL INFILE. Create the directory with correct ownership before starting the service.

unicode_ci is a conservative interoperability baseline. UCA 14 is available on both targets; use only after testing imports, comparisons and replication. This does not convert existing schemas or tables, and WordPress can choose its own collation.

Invalidations and contention can outweigh benefits for changing store data. Redis object caching is a separate application cache; it does not replace InnoDB’s page cache.

SQL text may contain personal data or tokens. Restrict access, set a retention policy, and measure overhead. General query logging remains disabled.

Temporarily lower during a controlled investigation if needed. Logging every query on a busy store can be expensive and exposes more sensitive data.

Details help distinguish scans, sorting and engine waits. The generator uses file logging; it does not enable general statement logging.

Instrumentation has overhead. Capture representative slow queries, lock waits and peak sessions before changing indexes or adding CPU. Changing this requires a restart.

Aggregates help prioritize actual workload costs. Access to observability tables still needs appropriate permissions; SQL text and metadata remain sensitive.