Distributed transaction and at mode principle of Seata

1. Distributed transactions

In a distributed system architecture , Complex businesses require cross database operations , To ensure the global transaction consistency, we need to solve the distributed transaction problem .

1.1 XA standard

XA Specification , Three roles are defined ：AP Applications 、RM Explorer ( database )、TM Transaction manager .XA The specification defines RM-TM Communication interface .

1.2 2PC Two-phase commit

2PC and XA Norms are not necessarily linked , however 2PC The pattern is right XA An abstract implementation and optimization of the specification .2PC Altogether 2 Stages ： Pre submission and commit.

     1. TM Inform each RM Prepare to commit their local transactions .
     1. TM Follow the steps 1 each RM The result of the feedback will be submitted in advance to decide whether to release commit The order of .

2PC There are several serious problems ：

     -  The blocking delay is large .
     - TM A single point of failure .
     -  Global data inconsistency , In stage two ,TM It's down commit operation , But only a part RM received commit, Some did not receive , There is a problem of inconsistency .

1.3 3PC Three stage commit

3PC yes 2PC Improved version , But there is no essential difference , and 2PC The difference is ：

     -  The timeout mechanism is introduced , stay TM and RM There is a timeout mechanism .
     -  stay 2PC Withholding and payment of commit A preparatory stage is interspersed in the middle , therefore 3PC Yes 3 Stages ：CanCommit、PreCommit、Commit.

1.4 TCC

Try-Confirm-Cancel, In essence, it is also a two-stage submission agreement for business service ,TCC Altogether 3 A step , The first stage requires the application to reserve resources , When all participants try The interfaces all worked , Then the transaction manager will commit the transaction confirm, Otherwise to cancel Rollback resource release .TCC Of try In the stage, it is generally necessary to take a pre snapshot of resources to record the changes of transaction data , Subsequent consistency alignment operations are performed based on this record when committing or rolling back .

1.5 Flexible business

Flexible transactions refer to services that are basically available 、 Soft state / Intermediate state 、 Final agreement ; The core idea is that strong consistency cannot be guaranteed in a distributed environment , Instead, make some compromises based on your own business scenarios . Common flexible transaction solutions are ： Best effort notification 、 compensate 、 Final consistency / Asynchronously ensures 、 Two stage submission .

2.Seata

Seata It is an Alibaba open source distributed transaction resolution framework . It supports AT、TCC、Saga、XA Multiple modes .Seata There are three roles in our organizational structure ：

• TC（Transaction Coordinator） Transaction coordinator ： Maintain the information of global transactions and branch transactions , Control the commit or rollback of global transactions ,TC It is a service that needs to be deployed independently , and TM and RM It's a client The client depends on the project .
• TM（Transaction Manager） Transaction manager ： The opening of global transactions and the definition of boundaries .
• RM（Resource Manager） Explorer ： Management of local branch transactions , Responsible for and TC Interact and report transaction status , Perform specific commit or rollback .

3.Seata AT Pattern ( Main stream )

Seata Support multiple distributed transaction modes , The official main push is AT Pattern ,AT Patterns are non intrusive 、 Similar to automation implicit TCC Pattern .AT The mode is divided into 2 Stages ：

1. Business data and undoLog Record in the same transaction and commit , Then release the local lock and connect resources .
2. If you need to submit , Asynchronous commit . If you need to roll back , Local execution undoLog Reverse compensation .

3.1 AT The advantages of patterns

  -  Automated agents DataSource, Automation realizes the compensation logic , Reduce business code intrusion .
  - TC Introduction of transaction coordinator , Simplified some transaction control logic .
  -  The isolation of distributed transactions ensures the implementation of write isolation and read isolation through global locks and local locks .

One stage local transaction Commit You need to get it before Global transaction lock ;
If the global lock is not obtained, the local transaction cannot be committed , There is a time limit for trying to obtain a global lock , If the threshold is exceeded, the rollback of the local transaction is executed .

3.2 AT The inadequacy of the model

     -  performance , Many times are involved in the transaction RPC Remote calls and various SQL Parsing and snapshot dropping .
     -  For the isolation of hot data, there are performance problems .
     -  Because the lock mechanism is introduced , May cause deadlock , Although there is a timeout , But it still has an impact on performance .