[ptrdom]

Motivation

apache/pekko-persistence-r2dbc#534 added the mariadb dialect to the persistence plugin. R2DBC projections can't use it yet:

• R2dbcOffsetStore.fromConfig has no offset store for Dialect.MariaDB.
• The MySQL offset store upserts with the VALUES (...) AS excluded row alias, which MariaDB doesn't support ([1064] ... near 'AS excluded ON DUPLICATE KEY UPDATE ...').

Draft until pekko-persistence-r2dbc is released. This builds against 2.0.0-M1+120-5a212bab-SNAPSHOT, which includes apache/pekko-persistence-r2dbc#534 (and the r2dbc-mariadb 1.3.2 bump in apache/pekko-persistence-r2dbc#538). It is stacked on #668, which needs the same snapshot handling (the close-calls-exceeding executor change) and adds Dialect.H2 to the same matches. Until #668 is merged, the diff here includes its commits; the MariaDB change is the last commit.

Modification

• Dialect: R2dbcOffsetStore.fromConfig creates a MariaDBR2dbcOffsetStore for Dialect.MariaDB. It extends MySQLR2dbcOffsetStore and only overrides the offset and management upserts. These use ON DUPLICATE KEY UPDATE col = VALUES(col), the same approach as MariaDBSnapshotDao in core: support MariaDB pekko-persistence-r2dbc#534. Everything else is the MySQL store:
  • the NOW(6) consumed timestamp,
  • binding each batch row before add(),
  • the CONCAT(...) NOT IN (...) delete.
• Timestamps: the MariaDB driver doesn't support java.time.Instant. Nothing in the projection handles that: ConnectionFactoryProvider in the plugin wraps the MariaDB connection pool so that Instant is bound and read as UTC LocalDateTime, with a UTC session. The projection gets its connections from that provider, so the offset store and the statements of an R2dbcSession are converted the same way.
• Schema: MariaDB uses the MySQL DDL (create_tables_mysql.sql), like the persistence plugin.
• Dependencies: org.mariadb:r2dbc-mariadb:1.3.2 (Apache 2.0), the version pekko-persistence-r2dbc is built against, test only in r2dbc-int-test. The r2dbc module doesn't reference driver classes.
• CI: a new "Build and Test Integration for R2DBC with MariaDB" job (Scala 2.13 and Scala 3 next). It uses docker-files/docker-compose-mariadb.yml (MariaDB 12.3.3, same as the plugin), with the server in America/New_York so that a non-UTC session would show up as wrong timestamps.
• Docs: a MariaDB entry in the schema list and a MariaDB section in r2dbc.md that links to the plugin's MariaDB docs. reference.conf lists mariadb as a dialect.
• Tests:
  • TestConfig gets a mariadb case.
  • R2dbcProjectionSpec's test repository gets a VALUES() upsert for MariaDB.
  • R2dbcTimestampOffsetProjectionSpec's spy offset store gets a MariaDB case.
  • New MariaDBR2dbcOffsetStoreSpec.

Result

R2DBC projections can store offsets in MariaDB 10.11 or later with the mariadb dialect.

Tests

• New MariaDBR2dbcOffsetStoreSpec (MariaDB 12.3.3 in America/New_York, Windows 11, JDK 25): 3 passed.
  • Dialect selection and the upsert SQL. This test makes no database calls, so it runs in every dialect's job.
  • Offset and management state updates.
  • Timestamp offsets at two instants whose UTC local times are ambiguous (2026-11-01T01:30Z) or don't exist (2027-03-14T02:30Z) in the server zone. The test checks the stored instant with UNIX_TIMESTAMP, which doesn't depend on the session time zone, and then reads the offset back.
    • With a New York session, MariaDB stores 2026-11-01 01:30 as 06:30Z but still reads it back as 01:30, so only UNIX_TIMESTAMP catches it. It rejects 2027-03-14 02:30 (ERROR 1292). Both checked with the mariadb CLI against the test image.
    • Not run through the plugin with a non-UTC session: the plugin rejects a non-UTC timezone option, and sessionVariables=time_zone=SYSTEM in the URL left the session in UTC.
  • The other two tests are canceled with other dialects.
  • With the MySQL upserts for MariaDB, 2 of 3 fail: the SQL check, and the update test with R2dbcBadGrammarException ... near 'AS excluded ...'.
• sbt r2dbc-int-test/test against PostgreSQL: 137 succeeded, 0 failed, 2 canceled.
• sbt -Dpekko.persistence.r2dbc.dialect=h2 -Dpekko.projection.r2dbc.dialect=h2 r2dbc-int-test/test: 137 succeeded, 0 failed, 2 canceled.
• sbt -Dpekko.persistence.r2dbc.dialect=mariadb -Dpekko.projection.r2dbc.dialect=mariadb r2dbc-int-test/test (MariaDB 12.3.3), 3 runs: 137 to 138 of 139 succeeded.
  • The failures were EventSourcedPubSubSpec, EventSourcedEndToEndSpec and once a R2dbcTimestampOffsetProjectionSpec test.
  • All of them pass when run on their own. The log shows Too many events stored with the same timestamp ..., buffer size [10], from app timestamps on the Windows clock.
  • With MySQL 9.5.0, EventSourcedPubSubSpec fails the same way locally, on this branch and on r2dbc: support H2 database (port akka-projection #914, #938) #668. CI passes it on r2dbc: support H2 database (port akka-projection #914, #938) #668.
• sbt -Dpekko.persistence.r2dbc.dialect=mysql -Dpekko.projection.r2dbc.dialect=mysql r2dbc-int-test/test (MySQL 9.5.0): 136 succeeded, 1 failed (EventSourcedPubSubSpec, see above), 2 canceled.
• ++3.3.8! and ++3.9.0! r2dbc/compile r2dbc-int-test/Test/compile: pass.
• sbt "checkCodeStyle; mimaReportBinaryIssues", headerCheckAll, Test/headerCheckAll, docs/paradox: pass. sbt headerCreateAll added the ASF headers.
• scalafmt (native, diff against r2dbc: support H2 database (port akka-projection #914, #938) #668's branch plus the new files): formatted. git diff --check: clean.
• Yugabyte not run locally; relying on CI.

Notes

• Copied code: none from outside this repo. The VALUES() upserts follow MariaDBSnapshotDao and docker-compose-mariadb.yml follows the plugin's file from core: support MariaDB pekko-persistence-r2dbc#534 (same ASF project).
• No check that the MariaDB driver and the mariadb dialect are used together, unlike the plugin's MariaDBConnections.requireMatchingDialect, which is private[r2dbc] in the plugin's package. The projection dialect defaults to the plugin's, and the plugin's DAOs already check the pairing. A projection set to mysql on the MariaDB driver fails on the first offset save with the syntax error above.
• The snapshot bump from +113 to +120 also brings in core: snapshots as starting points for eventsBySlices and eventsByPersistenceId pekko-persistence-r2dbc#515 (snapshots as starting points for eventsBySlices and eventsByPersistenceId). The PostgreSQL and H2 suites pass with it.
• After the pekko-persistence-r2dbc release: bump currentVersion to the release, rebase onto main once r2dbc: support H2 database (port akka-projection #914, #938) #668 is merged, re-run CI and mark ready.

References

Refs apache/pekko-persistence-r2dbc#534, Refs #668