fix: Abort running R2DBC statement when transaction coroutine is cancelled

[Matthew de Detrich (mdedetrich)]

Description

Summary of the change: Cancelling a coroutine running an R2DBC transaction now aborts the in-flight statement on the database (PostgreSQL), instead of waiting for it to run to completion.

Detailed description:

• Why: R2DBC drivers drain a cancelled subscription rather than aborting the query. Exposed's rollback and connection release then queue behind the still-running statement, so a cancelled coroutine only completes once the query finishes (e.g. a pg_sleep(30) cancelled after 2s took ~28s to join). Exposed holds the executing connection internally but never used it to cancel.

  At work, I am doing batch jobs in long running transactions on Postgres and currently there isn't a way to cancel the underlying transaction if the batches get cancelled. It would be ideal to model this using coroutine cancellation so that if an Exposed transaction in a coroutine gets cancelled it also cancels the database transaction.

• What:

  • Each R2dbcConnectionImpl tracks whether a statement may still be executing.
  • On CancellationException in a transaction block, a driver-specific cancel request is sent on the transaction's own connection before rollback.
  • New R2dbcDatabaseConfig.cancelRunningStatementOnCancellation (default true) to opt out.

• How:

  • New internal InFlightStatementTracker wraps statement result publishers; a statement counts as in flight from subscription until a terminal signal (which never arrives after a downstream cancel).
  • Connections are unwrapped via io.r2dbc.spi.Wrapped (so pooled connections work) down to PostgresqlConnection, and cancelRequest() is called. The Postgres driver is compileOnly, so this is guarded by a classpath check; other drivers are a no-op.
  • The cancel is sent only when a statement is tracked as in flight and a connection has already been acquired (it never acquires one). It is awaited before rollback/release, so it cannot hit a later statement or the next pool borrower; a cancel that arrives after the statement finished lands on an idle backend and is ignored.

---

Type of Change

Please mark the relevant options with an "X":

• Bug fix
• New feature
• Documentation update

Updates/remove existing public API methods:

• Is breaking change

Affected databases:

• MariaDB
• Mysql5
• Mysql8
• Oracle
• Postgres
• Redshift
• SqlServer
• H2
• SQLite

Checklist

• Unit tests are in place
• The build is green (including the Detekt check)
• All public methods affected by my PR has up to date API docs
• Documentation for my change is up to date