TwoPhaseCommit

Perform Two Phase Commits

This is an incomplete rewrite of perform-two-phase-commits using TableDB terms. Just give you an example of emulated transactions. All code is seudo.

Synopsis

This document provides a pattern for doing multi-document updates or "multi-document transactions" using a two-phase commit approach for writing data to multiple documents. Additionally, you can extend this process to provide a rollback-like <2-phase-commits-rollback> functionality.

Background

Operations on a single table document are always atomic with TableDB databases; however, operations that involve multiple documents, which are often referred to as "multi-document transactions", are not atomic. Since documents can be fairly complex and contain multiple "nested" documents, single-document atomicity provides the necessary support for many practical use cases.

Despite the power of single-document atomic operations, there are cases that require multi-document transactions. When executing a transaction composed of sequential operations, certain issues arise, such as:

• Atomicity: if one operation fails, the previous operation within the transaction must "rollback" to the previous state (i.e. the "nothing," in "all or nothing").

• Consistency: if a major failure (i.e. network, hardware) interrupts the transaction, the database must be able to recover a consistent state.

For situations that require multi-document transactions, you can implement two-phase commit in your application to provide support for these kinds of multi-document updates. Using two-phase commit ensures that data is consistent and, in case of an error, the state that preceded the transaction is recoverable <2-phase-commits-rollback>. During the procedure, however, documents can represent pending data and states.

Because only single-document operations are atomic with TableDB, two-phase commits can only offer transaction-like semantics. It is possible for applications to return intermediate data at intermediate points during the two-phase commit or rollback.

Pattern

Overview

Consider a scenario where you want to transfer funds from account A to account B. In a relational database system, you can subtract the funds from A and add the funds to B in a single multi-statement transaction. In TableDB, you can emulate a two-phase commit to achieve a comparable result.

The examples in this tutorial use the following two collections:

#. A collection named accounts to store account information.

#. A collection named transactions to store information on the fund transfer transactions.

Initialize Source and Destination Accounts

Insert into the accounts collection a document for account A and a document for account B.

db.accounts.insert({
        { _id= "A", balance: 1000, pendingTransactions= {} },
        { _id= "B", balance: 1000, pendingTransactions= {} } 
      })

Initialize Transfer Record

For each fund transfer to perform, insert into the transactions collection a document with the transfer information. The document contains the following fields:

• source and destination fields, which refer to the _id fields from the accounts collection,

• value field, which specifies the amount of transfer affecting the balance of the source and destination accounts,

• state field, which reflects the current state of the transfer. The state field can have the value of initial, pending, applied, done, canceling, and canceled.

• lastModified field, which reflects last modification date.

To initialize the transfer of 100 from account A to account B, insert into the transactions collection a document with the transfer information, the transaction state of "initial", and the lastModified field set to the current date:

   db.transactions.insert( 
       { _id= 1, source= "A", destination="B", value= 100, state= "initial", lastModified= Date() }
   )

Transfer Funds Between Accounts Using Two-Phase Commit

Recovering from Failure Scenarios

The most important part of the transaction procedure is not the prototypical example above, but rather the possibility for recovering from the various failure scenarios when transactions do not complete successfully. This section presents an overview of possible failures and provides steps to recover from these kinds of events.

Recovery Operations

The two-phase commit pattern allows applications running the sequence to resume the transaction and arrive at a consistent state. Run the recovery operations at application startup, and possibly at regular intervals, to catch any unfinished transactions.

The time required to reach a consistent state depends on how long the application needs to recover each transaction.

The following recovery procedures uses the lastModified date as an indicator of whether the pending transaction requires recovery; specifically, if the pending or applied transaction has not been updated in the last 30 minutes, the procedures determine that these transactions require recovery. You can use different conditions to make this determination.

Transactions in Pending State

To recover from failures that occur after step "Update transaction state to pending." but before "Update transaction state to applied." step, retrieve from the transactions collection a pending transaction for recovery:

   local dateThreshold = Date();
   dateThreshold.setMinutes(dateThreshold.getMinutes() - 30);
   local t = db.transactions.findOne( { state= "pending", lastModified= { _lt= dateThreshold } } );

And resume from step "Apply the transaction to both accounts._"

Transactions in Applied State

To recover from failures that occur after step "Update transaction state to applied." but before "Update transaction state to done." step, retrieve from the transactions collection an applied transaction for recovery:

   local dateThreshold = new Date();
   dateThreshold.setMinutes(dateThreshold.getMinutes() - 30);

   local t = db.transactions.findOne( { state= "applied", lastModified= { _lt= dateThreshold } } );

And resume from "Update both accounts' list of pending transactions._"

Rollback Operations

In some cases, you may need to "roll back" or undo a transaction; e.g., if the application needs to "cancel" the transaction or if one of the accounts does not exist or stops existing during the transaction.

Transactions in Applied State

After the "Update transaction state to applied._" step, you should not roll back the transaction. Instead, complete that transaction and :ref:create a new transaction <initialize-transfer-record> to reverse the transaction by switching the values in the source and the destination fields.

Transactions in Pending State

After the "Update transaction state to pending." step, but before the "Update transaction state to applied." step, you can rollback the transaction using the following procedure:

Multiple Applications

Transactions exist, in part, so that multiple applications can create and run operations concurrently without causing data inconsistency or conflicts. In our procedure, to update or retrieve the transaction document, the update conditions include a condition on the state field to prevent reapplication of the transaction by multiple applications.

For example, applications App1 and App2 both grab the same transaction, which is in the initial state. App1 applies the whole transaction before App2 starts. When App2 attempts to perform the "Update transaction state to pending._" step, the update condition, which includes the state: "initial" criterion.

When multiple applications are running, it is crucial that only one application can handle a given transaction at any point in time. As such, in addition including the expected state of the transaction in the update condition, you can also create a marker in the transaction document itself to identify the application that is handling the transaction. Use :method:~db.collection.findAndModify() method to modify the transaction and get it back in one step:

   t = db.transactions.findAndModify(
          {
            query= { state= "initial", application={ exists=false } },
            update=
              {
                _set= { state= "pending", application="App1" },
                currentDate= { lastModified=true } 
              },
            new= true
          }
       )

Amend the transaction operations to ensure that only applications that match the identifier in the application field apply the transaction.

If the application App1 fails during transaction execution, you can use the recovery procedures <2-phase-commits-recovery>, but applications should ensure that they "own" the transaction before applying the transaction. For example to find and resume the pending job, use a query that resembles the following:

   local dateThreshold = new Date();
   dateThreshold.setMinutes(dateThreshold.getMinutes() - 30);

   db.transactions.find( 
      {
        application= "App1",
        state= "pending",
        lastModified= { _lt= dateThreshold }
      }
   )

Using Two-Phase Commits in Production Applications

The example transaction above is intentionally simple. For example, it assumes that it is always possible to roll back operations to an account and that account balances can hold negative values.

Production implementations would likely be more complex. Typically, accounts need information about current balance, pending credits, and pending debits.