Merge (SQL)


A relational [database management system] uses SQL statements to new records or or existing records depending on whether condition matches. It was officially introduced in the SQL:2003 standard, and expanded in the SQL:2008 standard.

Usage


MERGE INTO tablename USING table_reference ON
WHEN MATCHED THEN
UPDATE SET column1 = value1
WHEN NOT MATCHED THEN
INSERT VALUES ;

A right join is employed over the Target and the Source --where Target is the left table and Source is the right one. The four possible combinations yield these rules:
  • If the field in the Source matches the field in the Target, then
  • If the field in the Source does not match the field in the Target, then
  • If the field does not exist in the Source but does exist in the Target, then no action is performed.
  • If the field does not exist in either the Source or Target, then no action is performed.
If multiple Source rows match a given Target row, an error is mandated by SQL:2003 standards. You cannot update a Target row multiple times with a statement

Implementations

Database management systems PostgreSQL, Oracle Database, IBM Db2, Teradata, EXASOL, Firebird, CUBRID, H2, HSQLDB, MS SQL, MonetDB, Vectorwise and Apache Derby support the standard syntax. Some also add non-standard SQL extensions.

Synonymous

Some database implementations adopted the term upsert to a database statement, or combination of statements, that inserts a record to a table in a database if the record does not exist or, if the record already exists, updates the existing record. This synonym is used in PostgreSQL and SQLite. It is also used to abbreviate the "MERGE" equivalent pseudo-code.
It is used in Microsoft Azure SQL Database.

Other non-standard implementations

Some other database management systems support this, or very similar behavior, through their own, non-standard SQL extensions.
MySQL, for example, supports the use of syntax which can be used to achieve a similar effect with the limitation that the join between target and source has to be made only on PRIMARY KEY or UNIQUE constraints, which is not required in the ANSI/ISO standard. It also supports syntax, which first attempts an insert, and if that fails, deletes the row, if exists, and then inserts the new one. There is also an clause for the statement, which tells the server to ignore "duplicate key" errors and go on.
SQLite's works similarly. It also supports as an alias for compatibility with MySQL.
Firebird supports though fails to throw an error when there are multiple Source data rows. Additionally there is a single-row version,, but the latter does not give you the option to take different actions on insert versus update
IBM Db2 extends the syntax with multiple and clauses, distinguishing them with guards.
Microsoft SQL Server extends with supporting guards and also with supporting Left Join via clauses.
PostgreSQL supports merge since version 15 but previously supported merging via.
CUBRID supports statement. And supports the use of syntax. It also supports for compatibility with MySQL.
Apache Phoenix supports and UPSERT SELECT syntax.
Spark SQL supports and clauses in actions.
Apache Impala supports.

Usage in NoSQL

A similar concept is applied in some NoSQL databases.
In MongoDB the fields in a value associated with a key can be updated with an operation. The raises an error if the key is not found. In the operation it is possible to set the flag: in this case a new value is stored associated to the given key if it does not exist, otherwise the whole value is replaced.
In Redis the operations sets the value associated with a given key. Redis does not know any detail of the internal structure of the value, so an update would have no meaning. So the operation has always a set or replace semantics.