Using transactions

flowchart LR
    B["BEGIN\nbuffer statements"] --> W["writes\nWAL Data + memory buffer"]
    W --> C["COMMIT\none storage txn: batch + fsync + publish"]
    W --> R["ROLLBACK\nAbort + discard"]
    C --> V["visible per\nMVCC snapshot rules"]

flowchart LR
    B["BEGIN\nbuffer statements"] --> W["writes\nWAL Data + memory buffer"]
    W --> C["COMMIT\none storage txn: batch + fsync + publish"]
    W --> R["ROLLBACK\nAbort + discard"]
    C --> V["visible per\nMVCC snapshot rules"]

See Run transactions for the task guide. This page is the precise reference.

BEGIN [TRANSACTION|WORK] [ISOLATION LEVEL …|READ ONLY|…]
START TRANSACTION […] | SET [SESSION|LOCAL] TRANSACTION … | SET … CHARACTERISTICS …
COMMIT [opts] | ROLLBACK [TO [SAVEPOINT] name] [opts]

• execute_all splits batches depth-aware on ;; BEGIN buffers into pending; COMMIT|ROLLBACK runs execute_transaction_group as one storage txn.
• In-txn writes allowed: INSERT/UPDATE/DELETE + CREATE TABLE|VIEW. COPY and UPDATE…FROM in-txn → Unsupported.
• SET search_path in-batch affects later statements.
• preview_dml_count / validate_dml_strict cover staged DML; maintenance accounting runs only after committed writes.

Isolation options parse but the engine is single-writer with MVCC snapshots (see MVCC); there is no multi-writer serializable scheduler to tune in v0.1.0.