Distributed Transaction (distributed_transaction)

Supports normal transactions and force_commit. In normal mode, the initiator prepares participators, asks the coordinator to record the global decision, then notifies participators to execute and acknowledge local work. force_commit executes during prepare without creating a coordinator record.

Location: src/component/distributed_transaction/ (see the README in that directory).

Quick Start

1. Enable dtcoordsvr and prepare Redis; configure transaction timeout, recovery, and cleanup in its chart.
2. Add distributed-transaction-sdk to USE_COMPONENTS; use sdk/transaction_client_handle.h, transaction_participator_handle.h, and rpc/transaction/transaction_api.h.
3. Implement participant callbacks using the component README's contract, prepare participants, resource keys, and transaction data, then submit through the client handle. Persist business data consistently with participant snapshots.
4. Verify a normal commit and rejection after a prepare failure. Local actions/recovery callbacks must be idempotent. Start with normal transaction mode and follow the decision/recovery rules below.

Customization and Design

Composition

Part	Location	Description
dtcoordsvr	dtcoordsvr/	Coordinator service: transaction_manager + a set of task actions
SDK	sdk/	transaction_client_handle (initiator), transaction_participator_handle (participator), transaction_api
Protocol	protocol/	distributed_transaction.proto, dtcoordsvr_config.proto

Coordinator RPC (task action)

create / commit / reject / query / remove / commit_participator / reject_participator.

Flow

When the global terminal state is unconfirmed, the client performs bounded queries. Partial replica responses from a failed call cannot decide notification direction. Later local or delivery failures cannot reverse a confirmed decision.

See Distributed Transaction Design for conflicts, replication, recovery, callback states, and runtime contracts. Also follow the component README for callbacks and parameters.