Delen via


Master-Subordinate Transactional Incremental Replication

Retired Content

This content is outdated and is no longer being maintained. It is provided as a courtesy for individuals who are still using these technologies. This page may contain URLs that were valid when originally published, but now link to sites or pages that no longer exist. Please see the patterns & practices guidance for the most current information.

Ff648233.DesTransactionalReplication(en-us,PandP.10).png

Version 1.0.0

GotDotNet community for collaboration on this pattern

Complete List of patterns & practices

Context

You are about to design the handling of transmissions on a replication link. Your requirements are:

The replication set that you are sending to the target consists only of changes made to the source data, so you are designing an Incremental Replication.

All needed data must be available to the target applications at any point in time; this includes related data such as reference data in other tables.

You have decided to achieve this by replicating the changes on a transactional level. Thus, you must retrieve transactional information from the source by either by accessing the transaction log of the database system or by preparing the source database as described in the Capture Transactional Detailspattern . Since the granularity of a transmission is a transaction, both source and target have to be database management systems (DBMSs), which we will describe in relational DBMS (RDBMS) terms.

Note: This pattern uses concepts, terms, and definitions that are introduced in the Data Replication architectural pattern. It uses the services described in the Capture Transaction Details pattern as well.

Problem

How do you design a replication link to support the transmission of transactional changes, and to replay them on the target while meeting all integrity requirements?

Forces

Any of the following compelling forces justify using the solution described in this pattern:

Availability of consistent data on a complex target schema. When applications read data from the target, all related data (for example, referenced data in other tables) must be available and consistently up-to-date. This requires that each transaction that updates the source be transmitted to the target as one transaction.

The following enabling forces facilitate the adoption of the solution, and their absence may hinder such a move:

Small volume of changes compared to the volume of the replication set. The size of the changes being moved from the source to the target on each transmission is significantly smaller than the size of the replication set. Thus, the changes will be transmitted faster than a new snapshot of the replication set.

Similarity in the replication sets. The schemas of both source and target must be nearly identical, although you can tolerate the following types of differences:

Different data types for corresponding columns at source and target.

One column of the source table can be split into different columns in the target table.

Several columns of the source table can be combined into a single column of the target table.

Solution

The solution is to acquire the information about committed transactions from the source and to replay the transactions in the correct sequence when they are written to the target.

Note: This pattern uses the terms transactions' and operations' with the following meanings:

A transaction is a collection of SQL commands that form a unit of work. Depending on the RDBMS, a transaction is started explicitly by a command like Begin Transaction, or implicitly by the first SQL command outside of a transaction. The transaction is ended either explicitly by a commit or a rollback, or implicitly at the end of every SQL command in autocommit mode.

An operation is the change (INSERT, UPDATE, or DELETE) of an individual row within a transaction.

Figure 1 shows the replication building block for this type of replication.

Ff648233.Des_TransactionalReplication_Fig01(en-us,PandP.10).gif

Figure 1: Replication building block for Master-Subordinate Transactional Incremental Replication

Depending on the features of the RDBMS and on the requirements of the replication link, the transactional information can be acquired either from the logging system of the source or from additional schema objects on the source. In both cases, the effects of the transaction are acquired as a collection of operations on the rows being updated by the transaction. The transmission to the target results in the corresponding rows being updated by transactions of the same size and in the correct sequence (according to their completion time at the source).

The detailed description of this pattern is separated into:

Background

Prerequisites for this pattern

Recorded transactions executed on the target

Background

Before starting to describe the design of a transactional replication link, here are two considerations of more general nature:

Transaction order for replay

Handling triggers, if the Write service uses standard SQL

Transaction Order for Replay

When replaying the transactions on the target, you do not have to mirror the database connection environment on the target to achieve data integrity. As concurrent transactions are isolated from each other, you can execute the transactions on the target sequentially instead. Thus, you can use a single database connection to replay all transactions of the current transmission. However, you must execute the individual transactions in the correct sequence, which is given by the time of their completion on the source.

The following example shows why the transactions must be ordered by their completion time, not their start time.

Two tables are written by two concurrent transactions. Figure 2 shows how both transactions are running in parallel.

Ff648233.Des_TransactionalReplication_Fig02(en-us,PandP.10).gif

Figure 2: Two concurrent transactions

Transaction 1 starts first and updates tableA. Then Transaction 2 starts, updates tableB, and completes with a commit. Finally, Transaction 1 also updates tableB. After both transactions, colB of tableB has a value of 20.

If the ordering were done on the transaction start time, the update of Transaction 2 would be the last one. Thus, colB of tableB would have a value of 10 at the end, instead of 20. However, if the ordering is done on the end time of each transaction, colB of tableB has a value of 20 at the end, which is the correct value.

If Transaction 1 had written tableB before Transaction 2 (for example, instead of tableA), then Transaction 2 would have been blocked until the end of Transaction 1. Transaction 2 would have been executed after Transaction 1, and colB of tableB would have a value of 10. Here again the transactions have to be ordered on their completion, which produces the correct result.

A sufficiently precise timestamp is needed to distinguish the completion time of any two transactions and thereby correctly order the transactions.

Handling Triggers, If the Write Uses Standard SQL

Triggers are schema objects that perform additional operations on behalf of an initial operation. Triggered operations are also part of the initiating transaction and are logged in the same way as any other operation.

There is an issue around how the RDBMS behaves when logged changes get applied to the target database. If you can apply the log information to the target without ever triggering a secondary operation, then you are fine. But if a trigger could be fired by applying a log record, then you have to be concerned.

This applies to either triggers that match source ones, or totally different triggers at the target.

For example, a source table has a trigger that is fired on every UPDATE. This would INSERT the old state of the row into a history table. There is a similar trigger on the target table. An UPDATE of the row in the source log is recorded, and the triggered INSERT into the history table is also recorded as part of the transaction. When executing this transaction on the target, the UPDATE of the table is performed first. This fires the trigger on the target, which causes an entry into the history table. But the source also recorded an INSERT into the history table. When replaying the next operation, a second INSERT into the history table is performed. Thus, the history table would have two new entries instead of one.

Hence, if you don't take special precaution around the handling of triggers, you might perform more operations on the target than you performed on the source.

To solve the problem, you must eliminate the effect of the trigger on the target during transmissions. To achieve this, the replication link should connect to the target database with a dedicated user or a dedicated role that is only used for transmissions, but not by any other applications. Additionally, the trigger has to be defined in a manner (depending on the SQL dialect of the RDBMS) that does not perform any operations if the database connection uses this special user or role.

Prerequisites

This pattern depends on two features that the DBMS must provide and on a prerequisite for the data model:

A fine-grained clock. The order in which transactions are executed on the source must be the same as the order in which they are replayed on the target. Thus, the clock must provide a sufficiently fine resolution to preserve the order. A clock grain of a millisecond is generally sufficient; many systems provide even microseconds. A clock with a resolution of whole seconds only will definitely prevent the use of this pattern.

Transaction Identifiers. The RDBMS must provide a means to identify the operations that belonging to the same transaction. This is called a Transaction Identifier throughout the remaining discussion. It can be an opaque data type, and is generally provided to handle distributed transactions.

Unique key. All tables of the replication set must have either unique keys or another combination of columns that uniquely identifies every row. The unique identifier of every row is referred to as the Replication Key throughout this document. After the Manipulate process, the resulting Replication Key must also identify every row in the target uniquely.

Elements of the Replication Building Block

The following paragraphs describe the elements of the replication building block for this type of replication.

Source

The source contains the replication set, which is a log of all changes that you want to acquire.

Acquire

The steps of Acquire are:

1.Connect to the source.

2.Find the transaction pending for transmission to the target that has the oldest completion timestamp.

3.Find the first operation of this transaction.

4.Pass the Transaction Identifier, the table name, the type of operation (INSERT, UPDATE, or DELETE) and the names and values of the columns to Manipulate.

5.Continue with step 4 until all operations for the current transaction are read.

6.Delete the record of the transmitted transaction, unless it is needed for other replication links.

7.Continue with step 2 until all transactions are read.

Manipulate

This service performs the following steps:

1.Get the first row from Acquire.

2.If the table needs some manipulation, perform the appropriate action, such as converting data types, and combining or splitting fields of the row.

3.Pass the Transaction Identifier, the table name, the type of operation (INSERT, UPDATE, or DELETE) and the names and values of the columns to the Write service.

4.Get the next row from Acquire and continue with step 2 until all rows are processed.

Write

The Write service performs the following steps:

1.Get the first row from Manipulate.

2.If the Transaction Identifier differs from the Transaction Identifier of the previously handled row, COMMIT the transaction.

3.Depending on the remaining attributes, build a SQL statement that INSERTs, UPDATEs, or DELETEs a single row in the target.

4.Get the next row from the Manipulate service and continue with step 2 until all rows are processed.

Target

The target is the database where the transactions are replayed. You must ensure that no triggered operations are executed, as described above.

Example

The Implementing Master-Subordinate Transactional Incremental Replication Using SQL Server pattern presents an implementation of the design pattern by the means provided with Microsoft SQL Server.

Resulting Context

The use of this pattern inherits the benefits and liabilities from Master-Subordinate Replication and has the following additional benefit and liability:

Benefit

Basis for other useful services. Other services might have a similar need to use transactional information. For example, spin off the transmission data to a historical store, such as a data warehouse.

Liability

Dependencies of schemas. This pattern depends greatly on the similarity between the source and the target schemas. If one of them changes, the other must change accordingly. (For minor changes, it could be sufficient to adapt Manipulate for the data conversion from the source schema to the target schema.)

Security Considerations

The database connection used to replay the transactions on the target must have sufficient access rights to INSERT, UPDATE, or DELETE in all tables belonging to the replication set. In addition, the Acquire database connection account needs to have SELECT or READ privileges on the source transaction objects.

It is recommended that you create a dedicated user in the target database with appropriate privileges and use this user for all transmissions. In general, this user will not be used for any other purposes.

Defining such a user allows you to tailor the privileges to the specific needs of Acquire or Write.

Operational Considerations

Before implementing Master-Subordinate Transactional Incremental Replication, the following are considerations to achieve smooth running operations:

Load on the source database server. When a large number of targets want to use the same source for their replication links, then the source can have operational difficulties in meeting all their demands. In this case, consider using the Master-Subordinate Cascading Replication pattern.

Load on the target database server. Transmissions replay transactions in the same way they were executed on the source. Although the transactions are executed sequentially on the target, the transmission consumes significant resources, such as CPU time and I/O activity, on the target. This impacts the response times of the applications during transmissions.

Space requirements for replaying the transactions on the target. Every transaction on the source is replayed as a single transaction on the target. Thus, you must provide sufficient space in the logging system to complete the largest transaction that might be executed on the source. Although the source must deal with concurrent transactions that the target does not need, a good starting point is to configure the logging system of the target in the same way that it is configured on the source.

Variants

After introducing the solution for transmitting the changes of a replication set from the source to the target, two variants show possible enhancements. The first variant shows a way to obtain a higher robustness of the replication. The second one sketches the implementation of a change history.

Higher Robustness

When writing the transactions to the target, you might consider converting INSERTs into UPDATEs or vice versa. Thus, if an INSERT raises a duplicate key error, you perform an UPDATE instead; and if the number of rows being hit by an UPDATE is zero, you perform an INSERT instead. This is sometimes referred to as an UPSERT.

Although such error handling is not necessary as long as the content of the target corresponds exactly to the state of the source before the transaction, it offers a higher degree of robustness. If the content of the target had been changed, for example, by an erroneous action of an operation or administrator, the normal execution of the transaction would fail; however, such an error handling will again align the contents of the source and the target.

Implementing a Change History

Instead of updating changed rows, they can be appended by adding a version number. Therefore, the previous version of the target row will be kept to retain a change history at the target. For example, this information can be used to trace a stock's performance record.

Additional Prerequisites

Some preparations have to be made for this variant to work.

The data schema must be extended by a version column.

The version number must be managed by Write, which increases the version number before it writes the new row by performing an INSERT.

Operational Considerations

Because rows will not be updated but inserted, the target database grows faster than the source database. Therefore, you must provide an appropriate amount of free disk space at the target site.

For more information, see the following related patterns:

Patterns That May Have Led You Here

Move Copy of Data. This is the root pattern of this cluster; it presents the overall architecture for maintain copies of data after they have been updated.

Data Replication. This pattern presents the architecture of a replication.

Master-Subordinate Replication. This pattern describes design considerations for transmitting data from the source to the target by overwriting potential changes in the target on a higher level.

Capture Transaction Details. This pattern describes the underlying change-capture service to provide transactional information from the source when a DBMS log is not available or is not to be used for this purpose.

Patterns That You Can Use Next

Implementing Master-Subordinate Transactional Incremental Replication Using SQL Server. This pattern shows how to implement Master-Subordinate Transactional Incremental Replication by using SQL Server.

Other Patterns of Interest

Master-Subordinate Snapshot Replication. This pattern presents a design for transmitting a complete replication set. This can be used to equalize both databases as a starting point before establishing an Incremental Replication.

Master-Subordinate Cascading Replication. This pattern shows replication topologies where Master-Subordinate Transactional Incremental Replication can be used to design the individual replication links of the topology.

patterns & practices Developer Center

Retired Content

This content is outdated and is no longer being maintained. It is provided as a courtesy for individuals who are still using these technologies. This page may contain URLs that were valid when originally published, but now link to sites or pages that no longer exist.