22.1.7.3 Limits Relating to Transaction Handling in NDB Cluster

A number of limitations exist in NDB Cluster with regard to the handling of transactions. These include the following:

Transaction isolation level. The NDBCLUSTER storage engine supports only the READ COMMITTED transaction isolation level. (InnoDB, for example, supports READ COMMITTED, READ UNCOMMITTED, REPEATABLE READ, and SERIALIZABLE.) You should keep in mind that NDB implements READ COMMITTED on a per-row basis; when a read request arrives at the data node storing the row, what is returned is the last committed version of the row at that time.
Uncommitted data is never returned, but when a transaction modifying a number of rows commits concurrently with a transaction reading the same rows, the transaction performing the read can observe “before” values, “after” values, or both, for different rows among these, due to the fact that a given row read request can be processed either before or after the commit of the other transaction.
To ensure that a given transaction reads only before or after values, you can impose row locks using SELECT ... LOCK IN SHARE MODE. In such cases, the lock is held until the owning transaction is committed. Using row locks can also cause the following issues:
- Increased frequency of lock wait timeout errors, and reduced concurrency
- Increased transaction processing overhead due to reads requiring a commit phase
- Possibility of exhausting the available number of concurrent locks, which is limited by MaxNoOfConcurrentOperations
NDB uses READ COMMITTED for all reads unless a modifier such as LOCK IN SHARE MODE or FOR UPDATE is used. LOCK IN SHARE MODE causes shared row locks to be used; FOR UPDATE causes exclusive row locks to be used. Unique key reads have their locks upgraded automatically by NDB to ensure a self-consistent read; BLOB reads also employ extra locking for consistency.
See Section 22.5.3.4, “NDB Cluster Backup Troubleshooting”, for information on how NDB Cluster's implementation of transaction isolation level can affect backup and restoration of NDB databases.
Transactions and BLOB or TEXT columns. NDBCLUSTER stores only part of a column value that uses any of MySQL's BLOB or TEXT data types in the table visible to MySQL; the remainder of the BLOB or TEXT is stored in a separate internal table that is not accessible to MySQL. This gives rise to two related issues of which you should be aware whenever executing SELECT statements on tables that contain columns of these types:
1. For any SELECT from an NDB Cluster table: If the SELECT includes a BLOB or TEXT column, the READ COMMITTED transaction isolation level is converted to a read with read lock. This is done to guarantee consistency.
2. For any SELECT which uses a unique key lookup to retrieve any columns that use any of the BLOB or TEXT data types and that is executed within a transaction, a shared read lock is held on the table for the duration of the transaction—that is, until the transaction is either committed or aborted.
  This issue does not occur for queries that use index or table scans, even against NDB tables having BLOB or TEXT columns.
  For example, consider the table t defined by the following CREATE TABLE statement:
  1. CREATE TABLE t (
  2. a INT NOT NULL AUTO_INCREMENT PRIMARY KEY,
  3. b INT NOT NULL,
  4. c INT NOT NULL,
  5. d TEXT,
  6. INDEX i(b),
  7. UNIQUE KEY u(c)
  8. ) ENGINE = NDB,
  Either of the following queries on t causes a shared read lock, because the first query uses a primary key lookup and the second uses a unique key lookup:
  1. SELECT * FROM t WHERE a = 1;
  3. SELECT * FROM t WHERE c = 1;
  However, none of the four queries shown here causes a shared read lock:
  1. SELECT * FROM t WHERE b = 1;
  3. SELECT * FROM t WHERE d = '1';
  5. SELECT * FROM t;
  7. SELECT b,c WHERE a = 1;
  This is because, of these four queries, the first uses an index scan, the second and third use table scans, and the fourth, while using a primary key lookup, does not retrieve the value of any BLOB or TEXT columns.
  You can help minimize issues with shared read locks by avoiding queries that use unique key lookups that retrieve BLOB or TEXT columns, or, in cases where such queries are not avoidable, by committing transactions as soon as possible afterward.
Rollbacks. There are no partial transactions, and no partial rollbacks of transactions. A duplicate key or similar error causes the entire transaction to be rolled back.
This behavior differs from that of other transactional storage engines such as InnoDB that may roll back individual statements.
Transactions and memory usage. As noted elsewhere in this chapter, NDB Cluster does not handle large transactions well; it is better to perform a number of small transactions with a few operations each than to attempt a single large transaction containing a great many operations. Among other considerations, large transactions require very large amounts of memory. Because of this, the transactional behavior of a number of MySQL statements is affected as described in the following list:
- TRUNCATE TABLE is not transactional when used on NDB tables. If a TRUNCATE TABLE fails to empty the table, then it must be re-run until it is successful.
- DELETE FROM (even with no WHERE clause) is transactional. For tables containing a great many rows, you may find that performance is improved by using several DELETE FROM ... LIMIT ... statements to “chunk” the delete operation. If your objective is to empty the table, then you may wish to use TRUNCATE TABLE instead.
- LOAD DATA statements. LOAD DATA is not transactional when used on NDB tables.
  
  Important
  
  When executing a LOAD DATA statement, the NDB engine performs commits at irregular intervals that enable better utilization of the communication network. It is not possible to know ahead of time when such commits take place.
- ALTER TABLE and transactions. When copying an NDB table as part of an ALTER TABLE, the creation of the copy is nontransactional. (In any case, this operation is rolled back when the copy is deleted.)
Transactions and the COUNT() function. When using NDB Cluster Replication, it is not possible to guarantee the transactional consistency of the COUNT() function on the slave. In other words, when performing on the master a series of statements (INSERT, DELETE, or both) that changes the number of rows in a table within a single transaction, executing SELECT COUNT(*) FROM table queries on the slave may yield intermediate results. This is due to the fact that SELECT COUNT(...) may perform dirty reads, and is not a bug in the NDB storage engine. (See Bug #31321 for more information.)

Rechercher dans le manuel MySQL

Traduction non disponible

Le manuel MySQL n'est pas encore traduit en français sur l'infobrol. Seule la version anglaise est disponible pour l'instant.

Document créé le 26/06/2006, dernière modification le 26/10/2018
Source du document imprimé : https://www.gaudry.be/mysql-rf-mysql-cluster-limitations-transactions.html

L'infobrol est un site personnel dont le contenu n'engage que moi. Le texte est mis à disposition sous licence CreativeCommons(BY-NC-SA). Plus d'info sur les conditions d'utilisation et sur l'auteur.

Références

Manuel MySQL : https://dev.mysql.com/

Ces références et liens indiquent des documents consultés lors de la rédaction de cette page, ou qui peuvent apporter un complément d'information, mais les auteurs de ces sources ne peuvent être tenus responsables du contenu de cette page.
L'auteur de ce site est seul responsable de la manière dont sont présentés ici les différents concepts, et des libertés qui sont prises avec les ouvrages de référence. N'oubliez pas que vous devez croiser les informations de sources multiples afin de diminuer les risques d'erreurs.

Table des matières Haut