MariaDB pursues parallel query execution: initial results, no release date

MariaDB is developing parallel query execution across multiple CPU cores, a capability the engine has never had: since inception, each query runs on a single thread. Oleg Smirnov from the Optimizer team explained this in the first episode of Inner Join, a new interview series published by the MariaDB Foundation. For administrators running MariaDB in production, the news matters most for analytical workloads: queries performing aggregations on large tables.

Work is focused on InnoDB. A table is divided into chunks starting from the root page of the clustered index B-tree: each entry in the root page becomes an initial chunk, processed by a worker or further subdivided if too large relative to available workers (typically 8 or 16, depending on core count). Each worker performs joins and aggregation on its chunk in a temporary table; partial results are then merged.

The practical finding from testing: a simple table scan gains nothing from parallelism, because sequential memory reads are already optimal and splitting introduces random access. Joins, filters and particularly aggregations are the beneficiaries.

Initial measurements

Two tests were cited, both on laptops rather than production servers. First: an aggregation query on a 100-million-row table, running in 20 seconds serially, halved with 2 workers, with diminishing gains at 4, 8 and 16 workers. Second is the TPC-H benchmark: some queries run 5 to 7 times faster (Q3 and Q6 most promising), others like Q7 and Q8 improve little, whilst Q1 and Q2 don't appear because their form isn't yet supported.

No dates, stability first

The team makes no commitment to a release date: first comes a phase of queries "guaranteed" correct in parallel, then QA, only then potential release. Smirnov's stated goal is having a stable initial phase by 2027.

Worth noting: MariaDB is late to this. PostgreSQL introduced Parallel Query in version 9.6 in 2016, a decade earlier. MySQL has InnoDB parallelism from 8.0.14 onwards, but it's limited to count(*) queries and check table, not general execution.

Those wishing to test the feature when an initial public version arrives can follow developments on the MariaDB Foundation's YouTube channel, where the full interview is published.