UniqueID Generator: A Pattern for Primary Key Generation

Motivation
Enterprise JavaBeans is a server-side component model that targets the specific business domain of online transaction processing (OLTP) applications. OLTP applications generally have the need to store information persistently. The data records or objects for each transaction require unique identifiers to allow them to be stored and retrieved accurately. Thus there's a need to generate unique identifiers for the data involved in an EJB transaction processing system.

For the purposes of this pattern, I'm assuming a relational database is used as the data store rather than an object database, which would provide its own ID generator. A relational database table typically has a primary key column that's indexed to prevent duplicate IDs, which would lead to data integrity problems. Thus each row stored in the table is uniquely identifiable. For instance, in EJB the return value from the call to myHome.findByPrimaryKey is a single entity bean, not multiple entities!

There are numerous ways to generate unique IDs in a Java application. Let's review some approaches before discussing the UniqueID generator pattern presented in this article.

System.currentTimeMillis()
An easy way to generate unique IDs is to utilize the System.currentTimeMillis() method to get the current time in milliseconds and use it as your ID. Although it's an easy way to start creating applications, this approach has implications across high-volume applications. A high-volume OLTP application may perform several calls to System.currentTimeMillis() at the same time, resulting in the generation of duplicate IDs. Thus the generator must perform some sort of synchronization on ID requests. Typically, this is done with a wrapper object that uses the synchronized modifier to queue threads accessing it.

Next, you ask, "What about synchronization across multiple JVMs?" Certainly a clustered-server, multi-JVM architecture will be the norm for an enterprise application, but a clustered application poses two problems for this approach toward generating unique IDs:

1. The system time on each machine may be different. Thus, in a multi-JVM architecture, calls to System.currentTimeMillis() can lead to duplicate values.
2. The Java synchronize operator doesn't work across multiple JVMs, so even if the system times were equal, simultaneous calls to different machines could still result in duplicate IDs.

EntityBean Key Generator
This approach uses an entity bean to select the next value from a relational database table, which holds the latest value for unique ID generation. An entity bean can encapsulate the SQL, which it executes in a generic fashion by loading in a bean environment property containing a SQL string for each type of database. For instance, in Oracle the dual table should automatically return a sequence value. The SQL statement might look something like this:

"select mysequence.nextval from dual"

Oracle sequences are a proprietary feature and increment the value automatically for the caller by a specified increment count. Other databases that don't have this feature must increment a next ID field and select it in one call. Stored procedures are a common cure for the need to execute multiple database operations within a single request. To get the same ID generation capabilities from Microsoft SQL Server, a stored procedure could be called that both returns a value and increments it for the next caller. Since the stored procedure performs its work inside a transaction in one database request, it eliminates a "window" of opportunity when simultaneous calls for another ID get the same row and attempt to update the value.

Thus our entity bean key generator executes SQL, which it retrieved generically from its bean environment property and executed on the database. However, there are some drawbacks to this approach:

• Each call to get another ID is a remote method invocation, which can create unwanted chatter depending on where the entity bean is deployed in your system.
• Some might argue that entity beans are for business entities rather than utilitarian functionality like generating IDs for business entities.
• Few application servers provide synchronized caching across a cluster. Thus the ability to cache a set of IDs, improving ID generation by preventing database operations, is negated when the application is clustered.

A more scalable approach to ID generation that provides both local caching and guaranteed unique IDs is a singleton object that hands out IDs from its local cache (see Figure 1). Each time the cache of the singleton object is depleted, it gets a set of IDs representing the next available IDs for the application. The singleton fetches IDs from a stateless session bean that accesses the database in a portable manner, allowing it to interact with any database.

Applicability
Use UniqueID Generator: (Source code)

1. To return a unique identifier for object and/or database row identity
2. To cache sets of IDs across multiple JVMs for scalable ID generation
3. To abstract database implementations from the core ID generation classes

Structure
Participants

• Singleton: Acts as the wrapper of the UniqueIDEJB and caches sets of unique IDs to prevent chatty remote invocations and database operations
• UniqueIDEJB: Component made up of UniqueID (remote interface) and UniqueIDHome (home interface); performs generic SQL to fetch a set of IDs from the database and return them to the singleton client

  Collaborations

• Singleton: Calls UniqueIDEJB component to fetch a set of IDs; set is cached locally and a new set is fetched when the maximum is surpassed

Consequences
The UniqueID generator has several key benefits and some limitations:

1. Limited remote method invocations and database operations: The UniqueID generator's singleton object caches a set of IDs that are depleted before another request is made to the UniqueIDEJB. This limits the number of remote method invocations on the EJB and the number of hits the database must handle to generate new IDs.
2. Guaranteed UniqueIDs: The UniqueID generator pattern guarantees unique identifiers for your objects/data across an n-tiered solution. If a set of IDs is requested from multiple JVMs at the same time, the transaction isolation of the UniqueIDEJB and/or database layer will automatically queue multiple requests to ensure that only one singleton's request for an ID set is processed at a time.
3. Portable across EJB servers: The pattern represents a component that's portable across EJB servers. It's a session bean, and session beans have been mandatory in the specification since its inception, whereas any EJB server still not up to the 1.1 version of the specification wouldn't be able to utilize an entity bean-based pattern.

1. Portable across database servers: The implementation of the pattern really determines its portability, but the pattern itself allows for database-specific SQL to be loaded generically from the EJB's environment so that no code changes are required to deploy it against disparate databases.
2. Singleton knows nothing about the "incrementBy" value: The singleton object knows nothing about how many IDs are returned in a set. This is controlled at the UniqueIDEJB level by an environment property. To change the incrementBy value, simply redeploy the EJB with a hot-deploy mechanism to avoid downtime and the increment will be changed to the new value without repercussions on the rest of the application.
3. ID gaps: Gaps in IDs will occur with this solution. For example, if a set of 50 IDs is retrieved and the server crashes midway, 25 IDs would be lost. My recommendation is to increment only by one until ready for production, then set the UniqueIDEJB's incrementBy environment property to be sufficient for your application's load. This prevents wasting IDs in development and test.
4. Support for other data types: The structure section's class diagram shows a long data type used for IDs, yet this could be any data type your application requires. However, the UniqueID generator pattern would have to be extended to account for different ID data types in a single application. For instance, if your application mixed doubles, strings, longs, and ints as keys, the singleton would have to be extended.

Implementation
Consider the following issues when implementing a generic, portable UniqueID generator pattern:

• Determining incrementBy for your application: The amount of IDs contained in each ID set (and subsequently cached in the JVM) should be determined based on the number of users attempting "insert" transactions per JVM. Thus, if your incrementBy is 50 on a single J2EE server, set it to approximately 25 if you cluster two servers. Also, set incrementBy to a low number, even 1, for development and testing. As with any tunable parameter, metrics gathered from stress testing your application should ultimately drive your settings.
• Ensure portability across EJB servers: When communicating from the UniqueIDEJB to the database, don't use proprietary logic to obtain a connection from a connection pool. Use data sources looked up through JNDI instead.
• Ensuring portability across databases: There are a few ways to ensure that your code to fetch a set of IDs is portable. One way is to store the SQL that will be executed in the EJB's environment. The SQL is set in the deployment descriptor, which allows it to be modified during deployment rather than having to modify the codebase. Another option is to use data access objects (DAOs) to contain the database-specific SQL code. Last, your UniqueIDEJB component couldn't take advantage of database-specific sequence generators. Your component's implementation should create the table at runtime if it doesn't exist already. This approach isn't recommended, but it's an alternative nonetheless.

• Generating different IDs for different tables (classes): To generate a different ID for different tables, each class in the pattern would have to change its interface to take a "sequence" parameter, which indicates which class you want to get the next ID for and is forwarded throughout the pattern to the JDBC call against the database.
• Generating IDs shouldn't rely on a business transaction outcome: The UniqueIDEJB should create its own new transaction when getting an ID set. Incrementing the database to the "next ID value" shouldn't rely on whether or not the business transaction currently executing in your application succeeds or fails.
• Business transactions rely on generated ID success: An error fetching an ID set from the UniqueIDEJB will cause a business transaction to fail. Don't mix different approaches to generating unique IDs when one fails. For instance, let's say your application uses the System.currentTimeMillis() approach to key generation if the first approach, calling UniqueIDEJB, fails. This provides an extra layer of fault tolerance to your application, but you could encounter duplicate IDs! If your database counter is near the number generated by the System.currentTimeMillis(), you'll have problems.

Sample Code
The code in Listing 1 shows how to implement the singleton to cache the ID set, fetch a new set, and handle failures when calling the UniqueIDEJB.

Summary
This month I provided an EJB pattern to a common problem in the OLTP world, generating unique IDs. While there are variations of the pattern, the UniqueID generator overcomes many scalability and flexibility issues where other patterns fall short, such as local ID caching, limiting remote method invocations, and portability. I hope this pattern is helpful for your EJB engagement. For other EJB patterns see www.theServerSide.com. And let me know if you're interested in JDJ featuring more patterns in the future.

Several patterns exist for generating primary keys for your EJB application. This month I'll provide a pattern for generating PKs that's scalable, generic, and portable.