Chapter 1. Introduction

Table of Contents

Overview
Replication Group Participants
Replicated Environments
Selecting a Master
Replication Streams
Managing Data Guarantees
Durability
Managing Data Consistency
Replication Group Life Cycle
Terminology
Node States
New Replication Group Startup
Subsequent Startups
Replica Startup
Master Failover
Two Node Groups

This book provides a thorough introduction to replication as used with Berkeley DB, Java Edition (JE). It begins by offering a general overview to replication and the benefits it provides. It also describes the APIs that you use to implement replication, and it describes architecturally the things that you need to do to your application code in order to use the replication APIs.

You should understand the concepts from the Berkeley DB, Java Edition Getting Started with Transaction Processing guide before reading this book.

Overview

Welcome to the JE High Availability (HA) product. JE HA is a replicated, single-master, embedded database engine based on Berkeley DB, Java Edition. JE HA offers important improvements in application availability, as well as offering improved read scalability and performance. JE HA does this by extending the data guarantees offered by a traditional transactional system to processes running on multiple physical hosts.

The JE replication APIs allow you to distribute your database contents (performed on a read-write Master) to one or more read-only Replicas. For this reason, JE's replication implementation is said to be a single master, multiple replica replication strategy.

Replication offers your application a number of benefits that can be a tremendous help. Primarily, replication's benefits revolve around performance, but there is also a benefit in terms of data durability guarantees.

Briefly, some of the reasons why you might choose to implement replication in your JE application are:

  • Improved application availability.

    By spreading your data across multiple machines, you can ensure that your application's data continues to be available even in the event of a hardware failure on any given machine in the replication group.

  • Improve read performance.

    By using replication you can spread data reads across multiple machines on your network. Doing so allows you to vastly improve your application's read performance. This strategy might be particularly interesting for applications that have readers on remote network nodes; you can push your data to the network's edges thereby improving application data read responsiveness.

  • Improve transactional commit performance

    In order to commit a transaction and achieve a transactional durability guarantee, the commit must be made durable. That is, the commit must be written to disk (usually, but not always, synchronously) before the application's thread of control can continue operations.

    Replication allows you to batch disk I/O so that it is performed as efficiently as possible while still maintaining a degree of durability by committing to the network. In other words, you relax your transactional durability guarantees on the machine where you perform the database write, but by virtue of replicating the data across the network you gain some additional durability guarantees beyond what is provided locally.

  • Improve data durability guarantee.

    In a traditional transactional application, you commit your transactions such that data modifications are saved to disk. Beyond this, the durability of your data is dependent upon the backup strategy that you choose to implement for your site.

    Replication allows you to increase this durability guarantee by ensuring that data modifications are written to multiple machines. This means that multiple disks, disk controllers, power supplies, and CPUs are used to ensure that your data modification makes it to stable storage. In other words, replication allows you to minimize the problem of a single point of failure by using more hardware to guarantee your data writes.

    If you are using replication for this reason, then you probably will want to configure your application such that it waits to hear about a successful commit from one or more replicas before continuing with the next operation. This will obviously impact your application's write performance to some degree — with the performance penalty being largely dependent upon the speed and stability of the network connecting your replication group.

Replication Group Participants

Processes that participate in a JE HA application are generically called nodes. Most nodes participate as a read-only Replica. One node in the HA application can perform database writes. This is the Master node.

The sum totality of all the nodes participating in the replicated application is called the replication group. While it is only a logical entity (there is no object that you instantiate and destroy which represents the replication group), the replication group is the first-order element of management for a replicated HA application. It is very important to remember that the replication group is persistent in that it exists regardless of whether its member nodes are currently running. In fact, all nodes that have been added to a replication group will remain in the group until they are manually removed from the group by you or your application's administrator.

Replication groups consist of electable nodes and, optionally, Monitor nodes. Electable nodes are replication group participants that can become the group's Master node through a replication election. If an electable node is not a Master, then it participates in the replication group as a read-only Replica. A key defining characteristic of an electable node is that it has access to a JE environment.

Note

Beyond Master and Replica, a node can also be in several other states. See Replication Group Life Cycle for more information.

Most of the nodes in a replication group are electable nodes. However, it is also possible to have Monitor nodes, which are nodes that do not have access to a JE environment. For this reason, they cannot serve as either a Master or a Replica. Instead, they merely monitor the composition of the replication group as changes are made by adding and removing electable nodes, and as elections are held to select a new Master. Monitor nodes are therefore used by applications external to the JE replicated application to route data requests to the various participants of the replication group.

Note that all nodes in a replication group have a unique group-wide name. Further, all replication groups are also assigned a unique name. This is necessary because it is possible for a single process to have access to multiple replication groups. Further, any given collection of hardware can be running multiple replication groups (a production and a test group, for example.) By uniquely identifying the replication group with a unique name, it is possible for JE HA to internally check that nodes have not been misconfigured and so make sure that messages are being routed to the correct location.

Replicated Environments

All electable nodes must have access to a database environment. Further, no electable node can share a database environment with another electable node.

More to the point, in order to create an electable node in a replication group, you use a specialized form of the environment handle: ReplicatedEnvironment.

There is no JE-specified limit to the number of environments which can participate in a replication group. The only limitation here is one of resources — network bandwidth, for example.

We discuss ReplicatedEnvironment handle usage in Using Replicated Environments. For an introduction to database environments, see the Getting Started with Berkeley DB, Java Edition guide.

Selecting a Master

Every replication group is allowed one and only one Master. Masters are selected by holding an election. All such elections are performed by the underlying Berkeley DB, Java Edition replication code.

When a node joins a replication group, it attempts to locate the Master. If it is the first node added to the replication group, then it automatically becomes the Master. If it is not the first node to startup in the replication group, and it cannot locate the Master, it calls for an election. Further, if at any time the Master becomes unavailable to the replication group, the individual replicas will call for an election.

When holding an election, replicas vote on who should be the Master. Among replicas participating in the election, the node with the most up-to-date set of logs will win the election. In order to win an election, a node must win a simple majority of the votes.

Usually JE requires a majority of electable nodes to be available to hold an election. If a simple majority is not available, then the replication group will no longer be able to accept write requests as there will be no Master.

Note that an electable node is part of the replication group even if it is currently not running or is otherwise unreachable by the rest of the replication group. Membership in the replication group is persistent; once a node joins the group it remains in the group regardless of its current state. The only way a node leaves a replication group is if you manually remove it from the group (see Adding and Removing Nodes from the Group for details). This is a very important point to remember when considering elections. An election cannot be held if the majority of electable nodes in the group are not running or are otherwise unreachable.

Note

There are two circumstances under which a majority of nodes need not be available in order to hold an election. The first is for the special circumstance of the two-node group. See Configuring Two-Node Groups for details.

The second circumstance is if you explicitly relax the requirement for a majority of nodes to be available in order to hold an election. This is a dangerous thing to do, and your replication group should rarely (if ever) be configured this way. See Managing a Failure of the Majority for more information.

Once a node has been elected Master, it remains in that role until the replication group has a reason to hold another election. Currently, the only reason why the group will try to elect a new Master is if the current Master becomes unavailable to the group. This can happen because you shutdown the current Master, the current Master crashes due to bugs in your application code, or a network outage causes the current Master to be unreachable by a majority of the nodes in your replication group.

In the event of a tie in the number of votes, JE's underlying implementation of the election code will pick the Master. Moreover, the election code will always make a consistent choice when settling a tie. That is, all things being even, the same node will always be picked to win a tied election.

Replication Streams

Write transactions can only be performed at the Master. The results of these transactions are replicated to Replicas using a logical replication stream.

Logical replication streams are performed over a TCP/IP connection. The stream contains a description of the logical changes (for example, insert, update or delete) operations that were performed on the database as a result of the transaction commit. Each such replicated change is assigned a group-wide unique identifier called a Virtual Log Sequence Number (VLSN). The VLSN can be used to locate the replicated change in the log files associated with any member of the group. Through the use of the VLSN, each operation described by the replication stream can be replayed at each Replica using an efficient internal replay mechanism.

A consequence of this logical replaying of a transaction is that physical characteristics of the log files contained at the Replicas can be different across the replication group. The data contents of the environments found across the replication group, however, should be identical.

Note that there is a process by which a non-replicated environment can be converted such that it has the log structure and metadata required for replication. See Converting Existing Environments for Replication for more information.