Supporting MERGE#

The Trino engine provides APIs to support row-level SQL MERGE. To implement MERGE, a connector must provide the following:

  • An implementation of ConnectorMergeSink, which is typically layered on top of a ConnectorPageSink.

  • Methods in ConnectorMetadata to get a “rowId” column handle, get the row change paradigm, and to start and complete the MERGE operation.

The Trino engine machinery used to implement SQL MERGE is also used to support SQL DELETE and UPDATE. This means that all a connector needs to do is implement support for SQL MERGE, and the connector gets all the Data Modification Language (DML) operations.

Standard SQL MERGE#

Different query engines support varying definitions of SQL MERGE. Trino supports the strict SQL specification ISO/IEC 9075, published in 2016. As a simple example, given tables target_table and source_table defined as:

CREATE TABLE accounts (
    customer VARCHAR,
    purchases DECIMAL,
    address VARCHAR);
INSERT INTO accounts (customer, purchases, address) VALUES ...;
CREATE TABLE monthly_accounts_update (
    customer VARCHAR,
    purchases DECIMAL,
    address VARCHAR);
INSERT INTO monthly_accounts_update (customer, purchases, address) VALUES ...;

Here is a possible MERGE operation, from monthly_accounts_update to accounts:

MERGE INTO accounts t USING monthly_accounts_update s
    ON (t.customer = s.customer)
    WHEN MATCHED AND s.address = 'Berkeley' THEN
    WHEN MATCHED AND s.customer = 'Joe Shmoe' THEN
        UPDATE SET purchases = purchases + 100.0
            SET purchases = s.purchases + t.purchases, address = s.address
        INSERT (customer, purchases, address)
            VALUES (s.customer, s.purchases, s.address);

SQL MERGE tries to match each WHEN clause in source order. When a match is found, the corresponding DELETE, INSERT or UPDATE is executed and subsequent WHEN clauses are ignored.

SQL MERGE supports two operations on the target table and source when a row from the source table or query matches a row in the target table:

  • UPDATE, in which the columns in the target row are updated.

  • DELETE, in which the target row is deleted.

In the NOT MATCHED case, SQL MERGE supports only INSERT operations. The values inserted are arbitrary but usually come from the unmatched row of the source table or query.


Different connectors have different ways of representing row updates, imposed by the underlying storage systems. The Trino engine classifies these different paradigms as elements of the RowChangeParadigm enumeration, returned by enumeration, returned by method ConnectorMetadata.getRowChangeParadigm(...).

The RowChangeParadigm enumeration values are:

  • CHANGE_ONLY_UPDATED_COLUMNS, intended for connectors that can update individual columns of rows identified by a rowId. The corresponding merge processor class is ChangeOnlyUpdatedColumnsMergeProcessor.

  • DELETE_ROW_AND_INSERT_ROW, intended for connectors that represent a row change as a row deletion paired with a row insertion. The corresponding merge processor class is DeleteAndInsertMergeProcessor.

Overview of MERGE processing#

A MERGE statement is processed by creating a RIGHT JOIN between the target table and the source, on the MERGE criteria. The source may be a table or an arbitrary query. For each row in the source table or query, MERGE produces a ROW object containing:

  • the data column values from the UPDATE or INSERT cases. For the DELETE cases, only the partition columns, which determine partitioning and bucketing, are non-null.

  • a boolean column containing true for source rows that matched some target row, and false otherwise.

  • an integer that identifies whether the merge case operation is UPDATE, DELETE or INSERT, or a source row for which no case matched. If a source row doesn’t match any merge case, all data column values except those that determine distribution are null, and the operation number is -1.

A SearchedCaseExpression is constructed from RIGHT JOIN result to represent the WHEN clauses of the MERGE. In the example preceding the MERGE is executed as if the SearchedCaseExpression were written as:

   WHEN present AND s.address = 'Berkeley' THEN
       -- Null values for delete; present=true; operation DELETE=2, case_number=0
       row(null, null, null, true, 2, 0)
   WHEN present AND s.customer = 'Joe Shmoe' THEN
       -- Update column values; present=true; operation UPDATE=3, case_number=1
       row(t.customer, t.purchases + 100.0, t.address, true, 3, 1)
   WHEN present THEN
       -- Update column values; present=true; operation UPDATE=3, case_number=2
       row(t.customer, s.purchases + t.purchases, s.address, true, 3, 2)
   WHEN (present IS NULL) THEN
       -- Insert column values; present=false; operation INSERT=1, case_number=3
       row(s.customer, s.purchases, s.address, false, 1, 3)
       -- Null values for no case matched; present=false; operation=-1,
       --     case_number=-1
       row(null, null, null, false, -1, -1)
 FROM (SELECT *, true AS present FROM target_table) t
   RIGHT JOIN source_table s ON s.customer = t.customer;

The Trino engine executes the RIGHT JOIN and CASE expression, and ensures that no target table row matches more than one source expression row, and ultimately creates a sequence of pages to be routed to the node that runs the ConnectorMergeSink.storeMergedRows(...) method.

Like DELETE and UPDATE, MERGE target table rows are identified by a connector-specific rowId column handle. For MERGE, the rowId handle is returned by ConnectorMetadata.getMergeRowIdColumnHandle(...).

MERGE redistribution#

The Trino MERGE implementation allows UPDATE to change the values of columns that determine partitioning and/or bucketing, and so it must “redistribute” rows from the MERGE operation to the worker nodes responsible for writing rows with the merged partitioning and/or bucketing columns.

Since the MERGE process in general requires redistribution of merged rows among Trino nodes, the order of rows in pages to be stored are indeterminate. Connectors like Hive that depend on an ascending rowId order for deleted rows must sort the deleted rows before storing them.

To ensure that all inserted rows for a given partition end up on a single node, the redistribution hash on the partition key/bucket columns is applied to the page partition keys. As a result of the hash, all rows for a specific partition/bucket hash together, whether they were MATCHED rows or NOT MATCHED rows.

For connectors whose RowChangeParadigm is DELETE_ROW_AND_INSERT_ROW, inserted rows are distributed using the layout supplied by ConnectorMetadata.getInsertLayout(). For some connectors, the same layout is used for updated rows. Other connectors require a special layout for updated rows, supplied by ConnectorMetadata.getUpdateLayout().

Connector support for MERGE#

To start MERGE processing, the Trino engine calls:

  • ConnectorMetadata.getMergeRowIdColumnHandle(...) to get the rowId column handle.

  • ConnectorMetadata.getRowChangeParadigm(...) to get the paradigm supported by the connector for changing existing table rows.

  • ConnectorMetadata.beginMerge(...) to get the a ConnectorMergeTableHandle for the merge operation. That ConnectorMergeTableHandle object contains whatever information the connector needs to specify the MERGE operation.

  • ConnectorMetadata.getInsertLayout(...), from which it extracts the the list of partition or table columns that impact write redistribution.

  • ConnectorMetadata.getUpdateLayout(...). If that layout is non-empty, it is used to distribute updated rows resulting from the MERGE operation.

On nodes that are targets of the hash, the Trino engine calls ConnectorPageSinkProvider.createMergeSink(...) to create a ConnectorMergeSink.

To write out each page of merged rows, the Trino engine calls ConnectorMergeSink.storeMergedRows(Page). The storeMergedRows(Page) method iterates over the rows in the page, performing updates and deletes in the MATCHED cases, and inserts in the NOT MATCHED cases.

When using RowChangeParadigm.DELETE_ROW_AND_INSERT_ROW, the engine translates UPDATE operations into a pair of DELETE and INSERT operations before storeMergedRows(Page) is called.

To complete the MERGE operation, the Trino engine calls ConnectorMetadata.finishMerge(...), passing the table handle and a collection of JSON objects encoded as Slice instances. These objects contain connector-specific information specifying what was changed by the MERGE operation. Typically this JSON object contains the files written and table and partition statistics generated by the MERGE operation. The connector takes appropriate actions, if any.

RowChangeProcessor implementation for MERGE#

In the MERGE implementation, each RowChangeParadigm corresponds to an internal Trino engine class that implements interface RowChangeProcessor. RowChangeProcessor has one interesting method: Page transformPage(Page). The format of the output page depends on the RowChangeParadigm.

The connector has no access to the RowChangeProcessor instance – it is used inside the Trino engine to transform the merge page rows into rows to be stored, based on the connector’s choice of RowChangeParadigm.

The page supplied to transformPage() consists of:

  • The write redistribution columns if any

  • For partitioned or bucketed tables, a long hash value column.

  • The rowId column for the row from the target table if matched, or null if not matched

  • The merge case RowBlock

  • The integer case number block

  • The byte is_distinct block, with value 0 if not distinct.

The merge case RowBlock has the following layout:

  • Blocks for each column in the table, including partition columns, in table column order.

  • A block containing the boolean “present” value which is true if the source row matched a target row, and false otherwise.

  • A block containing the MERGE case operation number, encoded as INSERT = 1, DELETE = 2, UPDATE = 3 and if no MERGE case matched, -1.

  • A block containing the number, starting with 0, for the WHEN clause that matched for the row, or -1 if no clause matched.

The page returned from transformPage consists of:

  • All table columns, in table column order.

  • The merge case operation block.

  • The rowId block.

  • A byte block containing 1 if the row is an insert derived from an update operation, and 0 otherwise. This block is used to correctly calculate the count of rows changed for connectors that represent updates and deletes plus inserts.

transformPage must ensure that there are no rows whose operation number is -1 in the page it returns.

Detecting duplicate matching target rows#

The SQL MERGE specification requires that in each MERGE case, a single target table row must match at most one source row, after applying the MERGE case condition expression. The first step toward finding these error is done by labeling each row in the target table with a unique id, using an AssignUniqueId node above the target table scan. The projected results from the RIGHT JOIN have these unique ids for matched target table rows as well as the WHEN clause number. A MarkDistinct node adds an is_distinct column which is true if no other row has the same unique id and WHEN clause number, and false otherwise. If any row has is_distinct equal to false, a MERGE_TARGET_ROW_MULTIPLE_MATCHES exception is raised and the MERGE operation fails.

ConnectorMergeTableHandle API#

Interface ConnectorMergeTableHandle defines one method, getTableHandle() to retrieve the ConnectorTableHandle originally passed to ConnectorMetadata.beginMerge().

ConnectorPageSinkProvider API#

To support SQL MERGE, ConnectorPageSinkProvider must implement the method that creates the ConnectorMergeSink:

  • createMergeSink:

    ConnectorMergeSink createMergeSink(
        ConnectorTransactionHandle transactionHandle,
        ConnectorSession session,
        ConnectorMergeTableHandle mergeHandle)

ConnectorMergeSink API#

To support MERGE, the connector must define an implementation of ConnectorMergeSink, usually layered over the connector’s ConnectorPageSink.

The ConnectorMergeSink is created by a call to ConnectorPageSinkProvider.createMergeSink().

The only interesting methods are:

  • storeMergedRows:

    void storeMergedRows(Page page)

    The Trino engine calls the storeMergedRows(Page) method of the ConnectorMergeSink instance returned by ConnectorPageSinkProvider.createMergeSink(), passing the page generated by the RowChangeProcessor.transformPage() method. That page consists of all table columns, in table column order, followed by the TINYINT operation column, followed by the rowId column.

    The job of storeMergedRows() is iterate over the rows in the page, and process them based on the value of the operation column, INSERT, DELETE, UPDATE, or ignore the row. By choosing appropriate paradigm, the connector can request that the UPDATE operation be transformed into DELETE and INSERT operations.

  • finish:

    CompletableFuture<Collection<Slice>> finish()

    The Trino engine calls finish() when all the data has been processed by a specific ConnectorMergeSink instance. The connector returns a future containing a collection of Slice, representing connector-specific information about the rows processed. Usually this includes the row count, and might include information like the files or partitions created or changed.

ConnectorMetadata MERGE API#

A connector implementing MERGE must implement these ConnectorMetadata methods.

  • getRowChangeParadigm():

    RowChangeParadigm getRowChangeParadigm(
        ConnectorSession session,
        ConnectorTableHandle tableHandle)

    This method is called as the engine starts processing a MERGE statement. The connector must return a RowChangeParadigm enumeration instance. If the connector doesn’t support MERGE, then it should throw a NOT_SUPPORTED exception to indicate that SQL MERGE isn’t supported by the connector. Note that the default implementation already throws this exception when the method isn’t implemented.

  • getMergeRowIdColumnHandle():

    ColumnHandle getMergeRowIdColumnHandle(
        ConnectorSession session,
        ConnectorTableHandle tableHandle)

    This method is called in the early stages of query planning for MERGE statements. The ColumnHandle returned provides the rowId used by the connector to identify rows to be merged, as well as any other fields of the row that the connector needs to complete the MERGE operation.

  • getInsertLayout():

    Optional<ConnectorTableLayout> getInsertLayout(
        ConnectorSession session,
        ConnectorTableHandle tableHandle)

    This method is called during query planning to get the table layout to be used for rows inserted by the MERGE operation. For some connectors, this layout is used for rows deleted as well.

  • getUpdateLayout():

    Optional<ConnectorTableLayout> getUpdateLayout(
        ConnectorSession session,
        ConnectorTableHandle tableHandle)

    This method is called during query planning to get the table layout to be used for rows deleted by the MERGE operation. If the optional return value is present, the Trino engine uses the layout for updated rows. Otherwise, it uses the result of ConnectorMetadata.getInsertLayout to distribute updated rows.

  • beginMerge():

    ConnectorMergeTableHandle beginMerge(
         ConnectorSession session,
         ConnectorTableHandle tableHandle)

    As the last step in creating the MERGE execution plan, the connector’s beginMerge() method is called, passing the session, and the tableHandle.

    beginMerge() performs any orchestration needed in the connector to start processing the MERGE. This orchestration varies from connector to connector. In the case of Hive connector operating on transactional tables, for example, beginMerge() checks that the table is transactional and starts a Hive Metastore transaction.

    beginMerge() returns a ConnectorMergeTableHandle with any added information the connector needs when the handle is passed back to finishMerge() and the split generation machinery. For most connectors, the returned table handle contains at least a flag identifying the table handle as a table handle for a MERGE operation.

  • finishMerge():

    void finishMerge(
        ConnectorSession session,
        ConnectorMergeTableHandle tableHandle,
        Collection<Slice> fragments)

    During MERGE processing, the Trino engine accumulates the Slice collections returned by ConnectorMergeSink.finish(). The engine calls finishMerge(), passing the table handle and that collection of Slice fragments. In response, the connector takes appropriate actions to complete the MERGE operation. Those actions might include committing an underlying transaction, if any, or freeing any other resources.