Transaction visibility, errors and retry boundaries

Baseline: v4.2.5. Transaction mode, actual isolation, MySQL-compatible variables and freshness are different dimensions. The string returned by @@transaction_isolation does not prove the execution engine uses that named strategy.

Timestamps and states

Value/state

Purpose

Diagnostic boundary

SnapshotTS

Read version boundary

RC statement refresh differs from an SI transaction snapshot

PrepareTS / CommitTS

TN preparation/commit ordering

Neither UUID transaction ID nor WAL LSN

Logtail watermark

CN synchronized readable progress

Compare with required snapshot and last commit

Committing / unknown

Commit in progress or unresolved

A network error does not establish rollback

BEGIN → workspace / reads / writes / locks → commit request
                                            ├→ committed → release
                                            ├→ aborted → rollback/release
                                            └→ unknown → resolve; avoid blind replay

Pessimistic mode defaults to RC and optimistic mode to SI; confirm runtime configuration. RC needs a suitable snapshot per statement. Internal statement retry is not permission to retry an entire business transaction without idempotency. TN timestamp/MVCC validation decides optimistic conflicts; this is not pessimistic locking at commit.

Error scope

Case

Default principle

Application action

Ordinary constraint/syntax statement error

With mo_rollback_txn_on_error=0, roll back the statement; transaction may remain open

Explicitly continue, commit or roll back

Deadlock, lock timeout/binding changes and selected connection/state errors

Static error set ends the transaction

Start a new transaction after checking error and idempotency

mo_rollback_txn_on_error=1

Real statement errors roll back the whole transaction

Subsequent work is not the original transaction

Ambiguous commit response

Outcome may not yet be known

Check a business idempotency key/state instead of replaying on network error alone

Classify by error number/SQLSTATE and retain transaction ID, rather than message substrings. Internal/background sessions have their own variable and lifecycle handling.

Locks and node changes

CN LockService manages local lock tables and proxies remote ones. The TN allocator versions bindings; stale bindings can abort a commit. CN loss, unreachable holders and TN reconstruction require reliability handling. Reconstructible caches do not make active transactions stateless.

Deadlock/lock waits differ from WAL waits. Retain lock-table/service IDs, holders/waiters, binding versions and transaction timing, then correlate lock-service logs.

Cross-CN visibility

Source commit acknowledgement, destination Logtail application and destination snapshot selection are separate events. Same-session, another session on the same CN and a session on another CN are different tests. Reconnection requires subscription recovery; a healthy port does not establish catch-up.

See transaction usage, isolation and Logtail.

Source and regression entry points