All managed entity state transitions are translated to associated database statements when the current Persistence Context gets flushed. Hibernate’s flush behavior is not always as obvious as one might think.
Hibernate tries to defer the Persistence Context flushing up until the last possible moment. This strategy has been traditionally known as transactional write-behind.
The write-behind is more related to Hibernate flushing rather than any logical or physical transaction. During a transaction, the flush may occur multiple times.
The flushed changes are visible only for the current database transaction. Until the current transaction is committed, no change is visible by other concurrent transactions.
The persistence context, also known as the first level cache, acts as a buffer between the current entity state transitions and the database.
In caching theory, the write-behind synchronization requires that all changes happen against the cache, whose responsibility is to eventually synchronize with the backing store.
Reducing lock contention
Every DML statement runs inside a database transaction. Based on the current database transaction isolation level, locks (shared or explicit) may be acquired for the current selected/modified table rows.
Reducing the lock holding holding time lowers the dead-lock probability, and according to the scalability theory, it increases throughput. Locks always introduce serial executions, and according to Amdahl’s law, the maximum speedup is inversely proportional with the serial part of the currently executing program.
Even in READ_COMMITTED isolation level, UPDATE and DELETE statements acquire locks. This behavior prevents other concurring transactions from reading uncommitted changes or modify the rows in question.
So, deferring locking statements (UPDATE/DELETE) may increase performance, but we must make sure that data consistency is not affected whatsoever.
Postponing the entity state transition synchronization has another major advantage. Since all changes are being flushed at once, Hibernate may benefit from the JDBC batching optimization.
Batching improves performance by grouping multiple DML statements into a single operation, therefore reducing database round-trips.
Since queries are always running against the database (unless second level query cache is being hit), we need to make sure that all pending changes are synchronized before the query starts running.
Therefore, both JPA and Hibernate define a flush-before-query synchronization strategy.
From JPA to Hibernate flushing strategies
|JPAFLUSHMODETYPE||HIBERNATEFLUSHMODE||HIBERNATE IMPLEMENTATION DETAILS|
|AUTO||AUTO||The Session is sometimes flushed before query execution.|
|COMMIT||COMMIT||The Session is only flushed prior to a transaction commit.|
|ALWAYS||The Session is always flushed before query execution.|
|MANUAL||The Session can only be manually flushed.|
|Deprecated. Use MANUAL instead. This was the original name given to manual flushing, but it was misleading users into thinking that the Session won’t ever be flushed.|
Current Flush scope
The Persistence Context defines a default flush mode, that can be overridden upon Hibernate Session creation. Queries can also take a flush strategy, therefore overruling the current Persistence Context flush mode.
In my next post, you’ll find out that Hibernate FlushMode.AUTO breaks data consistency for SQL queries and you’ll see how you can overcome this shortcoming.