Help Center/ Distributed Database Middleware/ Best Practices/ Transaction Models
Updated on 2023-01-13 GMT+08:00
View PDF
Share

Transaction Models

Tables in each DDM instance are usually sharded, so data in the tables may be distributed across different database shards in multiple RDS instances. In DDM, one transaction to add, delete, update, or query a logical table is possibly executed on different database shards in multiple RDS instances, so a series of operations on data tables in one shard of an RDS instance can be deemed as a local transaction. Therefore, one DDM transaction is actually a distributed transaction that consists of local transactions on multiple RDS instances. These local transactions either all succeed, or all fail.

Implementation of Distributed Transactions in DDM

DDM executes 2PC distributed transactions based on the MySQL XA protocol, and these transactions can guarantee strong write consistency. For more information about the consistency, see the MySQL official documentation.

In a distributed system, each participant knows whether their operation succeeds or fails, but cannot know the status of the operation on other nodes. If a transaction encompasses multiple nodes, to guarantee atomicity and consistency of a transaction, a coordinator is introduced to control operation results of all participants so that all of the transaction's changes either take effect or do not. The DDM node acts as a coordinator in the distributed transaction, and RDS instances are participants.

The two-phase commit protocol breaks a database commit into two phases:

  1. Prepare phase: Participants notify the coordinator of the operation result. If the participants give a response that they are prepared, they must reserve the required resources before the coordinator makes a decision.
  2. Commit phase: After receiving a notification from the participants, the coordinator sends a notification to the participants again and determines whether the participant needs to commit or roll back the operation based on their response.

For example:

Four people A, B, C, and D want to have a dinner party, so they need to determine the time. Now assume that A is the coordinator and B, C, and D are participants.

Prepare phase:

  1. A sends a text message to B, C, and D, asking them whether they are available on Tuesday noon.
  2. D replies yes.
  3. B replies yes.
  4. C does not reply for a long time. The time is not determined, so A, B, C, and D cannot continue their dinner party.
  5. C replies yes or no.

Commit phase:

  1. Coordinator A sends the collected results (Tuesday dinner party) to B, C, and D. What the result is and when does A feed it back to the other three depends on the time when C replies.
  2. B received.
  3. C received.
  4. D received.
  5. If any of them does not receive the result, A continues to send it until all of them receive the result.

2-Phase Commit

When DDM processes a transaction, applications run BEGIN/COMMIT as they do on a common transaction, without worrying about whether there are distributed transactions at the underlying layer. DDM automatically processes distributed transactions using the two-phase commit (2PC) protocol. If a transaction involves only one data shard, DDM will process it using the one-phase commit protocol. The one-phase commit protocol will not be described in detail here.

Figure 1 Flowchart for executing a distributed transaction

XID is the global unique serial number of each distributed transaction. It consists of the transaction serial number, node ID, and timestamp. The XID and the corresponding shard name together constitute the XA_ID that initiates a distributed transaction on a physical database.

Suggestions for Using Transactions

  • DDM supports distributed transactions. If all SQL statements in a transaction use the same sharding key, DDM can process the transaction as a single-shard transaction to achieve optimal performance.
  • DDM uses the 2PC protocol to process distributed transactions. RDS deadlock detection does not take effect for distributed transactions across RDS instances. If a lock wait times out, there may be a cross-shard deadlock. In this case, check your business model.