ערוך

שתף באמצעות


Peer-to-Peer - Transactional Replication

Applies to: SQL Server

Peer-to-peer replication provides a scale-out and high-availability solution by maintaining copies of data across multiple server instances, also referred to as nodes. Built on the foundation of transactional replication, peer-to-peer replication propagates transactionally consistent changes in near real-time. This enables applications that require scale-out of read operations to distribute the reads from clients across multiple nodes. Because data is maintained across the nodes in near real-time, peer-to-peer replication provides data redundancy, which increases the availability of data.

Consider a web application. This can benefit from peer-to-peer replication in the following ways:

  • Catalog queries and other reads are spread across multiple nodes. This enables performance to remain consistent as reads increase.

  • If one of the nodes in the system fails, an application layer can redirect the writes for that node to another node. This maintains availability.

  • If a node requires maintenance or the whole system requires an upgrade, each node can be taken offline and added back to the system without affecting the availability of the application.

Although peer-to-peer replication enables scaling out of read operations, write performance for the topology is like that for a single node. This is because ultimately all inserts, updates, and deletes are propagated to all nodes. Replication recognizes when a change has been applied to a given node and prevents changes from cycling through the nodes more than one time. We strongly recommend that write operations for each row be performed at only one node, for the following reasons:

  • If a row is modified at more than one node, it can cause a conflict or even a lost update when the row is propagated to other nodes.

  • There is always some latency involved when changes are replicated. For applications that require the latest change to be seen immediately, dynamically load balancing the application across multiple nodes can be problematic.

Peer-to-peer replication includes the option to enable conflict detection across a peer-to-peer topology. This option helps prevent the issues that are caused from undetected conflicts, including inconsistent application behavior and lost updates. By enabling this option, by default a conflicting change is treated as a critical error that causes the failure of the Distribution Agent. In the event of a conflict, the topology remains in an inconsistent state until the conflict is resolved manually and the data is made consistent across the topology. For more information, see Conflict Detection in Peer-to-Peer Replication.

Note

To avoid potential data inconsistency, make sure that you avoid conflicts in a peer-to-peer topology, even with conflict detection enabled. To ensure that write operations for a particular row are performed at only one node, applications that access and change data must partition insert, update, and delete operations. This partitioning ensures that modifications to a given row originating at one node are synchronized with all other nodes in the topology before the row is modified by a different node. If an application requires sophisticated conflict detection and resolution capabilities, use merge replication. For more information, see Merge Replication and Detect and Resolve Merge Replication Conflicts.

Peer-to-Peer Topologies

The following scenarios illustrate typical uses for peer-to-peer replication.

Topology That Has Two Participating Databases

Peer-to-peer replication, two nodes

Both of the preceding illustrations show two participating databases, with user traffic directed to the databases through an application server. This configuration can be used for a number of applications, from Web sites to workgroup applications, and provides the following benefits:

  • Improved read performance, because reads are spread out over two servers.

  • Higher availability if maintenance is required or of failure at one node.

In both illustrations, read activity is load-balanced between the participating databases, but updates are handled differently:

  • On the left, updates are partitioned between the two servers. If the database contained a product catalog, you could, for example, have a custom application direct updates to node A for product names that start with A through M, and direct updates to node B for product names that start with N through Z. Updates are then replicated to the other node.

  • On the right, all updates are directed to node B. From there, updates are replicated to node A. If B is offline (for example, for maintenance), the application server can direct all activity to A. When B is back online, updates can flow to it, and the application server can move all updates back to B or keep directing them to A.

Peer-to-peer replication can support either approach, but the central update example on the right is also often used with standard transactional replication.

Topologies That Have Three or More Participating Databases

Peer-to-peer replication to dispersed locations

The preceding illustration shows three participating databases that provide data for a worldwide software support organization, with offices in Los Angeles, London, and Taipei. The support engineers at each office take customer calls and enter and update information about each customer call. The time zones for the three offices are eight hours apart, so there's no overlap in the workday. As the Taipei office closes, the London office is opening for the day. If a call is still in progress as one office is closing, the call is transferred to a representative at the next office to open.

Each location has a database and an application server, which are used by the support engineers as they enter and update information about customer calls. The topology is partitioned by time. Therefore, updates occur only at the node that is currently open for business, and then the updates flow to the other participating databases. This topology provides the following benefits:

  • Independence without isolation: Each office can insert, update, or delete data independently but can also share the data because it's replicated to all other participating databases.

  • Higher availability in case of failure or to allow maintenance at one or more of the participating databases.

    Peer-to-peer replication, three and four nodes

    The preceding illustration shows the addition of a node to the three-node topology. A node could be added in this scenario for the following reasons:

  • Because another office is opened.

  • To provide higher availability to support maintenance or increase fault tolerance if a disk failure, or other major failure occurs.

Notice that in both the three- and four-node topologies, all databases publish and subscribe to all other databases. This provides maximum availability if there are maintenance needs or failures of one or more nodes. As nodes are added, you must balance availability and scalability needs against performance and the complexity of deployment and administration.

Configuring Peer-to-Peer Replication

Configuring a peer-to-peer replication topology is similar to configuring a series of standard transactional publications and subscriptions. The steps described in the following articles show the configuration of a three-node system, similar to the configuration shown on the left in the previous illustration that shows peer-to-peer topology.

Considerations for Using Peer-to-Peer Replication

This section provides information and guidelines to consider when you use peer-to-peer replication.

General Considerations

  • Peer-to-peer replication is available only in enterprise edition of SQL Server.

  • All databases that participate in peer-to-peer replication should contain identical schema and data:

    • Object names, object schema, and publication names should be identical.

    • Publications must allow schema changes to be replicated. (This is a setting of 1 for the publication property replicate_ddl, which is the default setting.) For more information, see Make Schema Changes on Publication Databases.

    • Row and column filtering aren't supported.

  • Each node should use its own distribution database. This eliminates the potential of having a single point of failure.

  • Tables and other objects can't be included in multiple peer-to-peer publications in a single publication database.

  • A publication must be enabled for peer-to-peer replication before any subscriptions are created.

  • Subscriptions must be initialized by using a backup or with the replication support only option. For more information, see Initialize a Transactional Subscription Without a Snapshot.

  • Don't use identity columns. When using identities, you must manually manage the ranges assigned to the tables at each participating database. For more information, see Assigning ranges for manual identity range management.

Feature Restrictions

Peer-to-peer replication supports the core features of transactional replication, but doesn't support the following options:

  • Initialization and reinitialization with a snapshot.

  • Row and column filters.

  • Timestamp columns.

  • Non-SQL Server Publishers and Subscribers.

  • Immediate updating and queued updating subscriptions.

  • Anonymous subscriptions.

  • Partial subscriptions.

  • Attachable subscriptions and transformable subscriptions. (Both of these options were deprecated in SQL Server 2005 (9.x).)

  • Shared Distribution Agents.

  • The Distribution Agent parameter -SubscriptionStreams and the Log Reader Agent parameter -MaxCmdsInTran.

  • The article properties @destination_owner and @destination_table.

  • Peer-to-Peer transactional replication doesn't support creating a one-way transactional subscription to a Peer-to-Peer publication

  • Peer-to-Peer transactional replication doesn't support publishing tables with computed columns as part of their primary key.

The following properties have special considerations:

  • The publication property @allow_initialize_from_backup requires a value of true.

  • The article property @replicate_ddl requires a value of true; @identityrangemanagementoption requires a value of manual; and @status requires that option 24 is set.

  • The value for article properties @ins_cmd, @del_cmd, and @upd_cmd can't be set to SQL.

  • The subscription property @sync_type requires a value of none or automatic.

  • SQL Server 2019 (15.x) CU 13 introduces the publication property, @p2p_confictdetection_policy. The default parameter value is originatorid and it resolves the conflict on basis of originator ID. The lastwriter parameter value resolves the conflict on basis of last writer.

Maintenance Considerations

Some actions require the system to be quiescent. This means stopping activity on published tables at all nodes and making sure that each node has received all changes from all other nodes.

Action SQL Server 2005 peers only or mix of SQL Server 2005 peers with SQL Server 2008 peers and higher SQL Server 2005 peers only or mix of SQL Server 2005 peers with SQL Server 2008 peers and higher SQL2008 peers and higher SQL2008 peers and higher
Adding a node to the topology 2 nodes in complete topology: No quiescing required. Use sync_type = 'initialize with backup'. More than 2 nodes: Quiescing required. sync_type = 'replication support only': Quiescing required. sync_type = 'initialize with backup' and 'initialize from lsn': No quiescing required.

Topology schema changes (adding or dropping an article) requires quiescing. For more information, see Administer a Peer-to-Peer Topology (Replication Transact-SQL Programming).

Removing a node from the topology never requires quiescing.

Changing the article properties by using sp_changearticle never requires quiescing. Allowable changes (for P2P) are the description, ins_cmd, upd_cmd, and del_cmd properties.

Article Schema changes (adding/dropping column) never requires quiescing.

  • Adding article: For adding an article to an existing configuration- we need to quiesce the system, execute CREATE TABLE statement & load initial data at each node in the topology and add the new article at each node in the topology.

  • Dropping article: If we want a consistent state on all nodes, we need to quiesce the topology

For more information, see Quiesce a Replication Topology (Replication Transact-SQL Programming) and Administer a Peer-to-Peer Topology (Replication Transact-SQL Programming).

  • If you add a new node to a peer-to-peer topology, you should restore only from backups that were created after the new node was added.

  • You can't reinitialize subscriptions in a peer-to-peer topology. If you have to ensure that a node has a new copy of the data, restore a backup at the node.