When a transaction reads a row, the isolation level of the transaction determines if a read lock is acquired. Once a row is read locked, no other transaction can obtain a write lock on it. Acquiring a read lock ensures that a different transaction does not modify or delete a row while it is being read. Any number of transactions can acquire read locks on any row at the same time, so read locks are sometimes referred to as shared locks, or non-exclusive locks.
Read locks can be held for different durations. At isolation levels 2 and 3, any read locks acquired by a transaction are held until the transaction completes through a COMMIT or a ROLLBACK. These read locks are called long-term read locks.
For transactions executing at isolation level 1, the database server acquires a short-term read lock on the row upon which a cursor is positioned. As the application scrolls through the cursor, the short-term read lock on the previously-positioned row is released, and a new short-term read lock is acquired on the subsequent row. This technique is called cursor stability. Because the application holds a read lock on the current row, another transaction cannot make changes to the row until the application moves off the row. Note that more than one lock can be acquired if the cursor is over a query involving multiple tables. Short-term read locks are acquired only when the position within a cursor must be maintained across requests (ordinarily, these would be FETCH statements issued by the application). For example, short-term read locks are not acquired when processing a SELECT COUNT(*) query since a cursor opened over this statement will never be positioned on a particular base table row. In this case, the database server only needs to guarantee read committed semantics, that is, that the rows processed by the statement have been committed by other transactions.
Transactions executing at isolation level 0 (read uncommitted) do not acquire long-term or short-term read locks, and do not conflict with other transactions (except for exclusive schema locks). However, isolation level 0 transactions may process uncommitted changes made by other concurrent transactions. You can avoid processing uncommitted changes by using snapshot isolation. See Snapshot isolation.
Discuss this page in DocCommentXchange.
|
Copyright © 2010, iAnywhere Solutions, Inc. - SQL Anywhere 12.0.0 |