# Supporting DELETE and UPDATE#

The Trino engine provides APIs to support row-level SQL DELETE and UPDATE. To implement DELETE or UPDATE, a connector must layer an UpdatablePageSource on top of the connector’s usual ConnectorPageSource, and define ConnectorMetadata methods to get a “rowId” column handle; to start the operation; and to finish the operation.

## DELETE and UPDATE Data Flow#

DELETE and UPDATE have a similar flow:

• For each split, the connector will create an UpdatablePageSource instance, layered over the connector’s ConnectorPageSource, to read pages on behalf of the Trino engine, and to write deletions and/or updates to the underlying data store.

• The connector’s UpdatablePageSource.getNextPage() implementation fetches the next page from the underlying ConnectorPageSource, optionally reformats the page, and returns it to the Trino engine.

• The Trino engine performs filtering and projection on the page read, producing a page of filtered, projected results.

• The Trino engine passes that filtered, projected page of results to the connector’s UpdatablePageSource deleteRows() or updateRows() method. Those methods persist the deletions or updates in the underlying data store.

• When all the pages for a specific split have been processed, the Trino engine calls UpdatablePageSource.finish(), which returns a Collection<Slice> of “fragments” representing connector-specific information about the rows processed by the calls to deleteRows or updateRows.

• When all pages for all splits have been processed, the Trino engine calls ConnectorMetadata.finishDelete() or finishUpdate, passing a collection containing all the “fragments” from all the splits. The connector does what is required to finalize the operation, for example, committing the transaction.

## The rowId Column Abstraction#

The Trino engine and connectors use a “rowId” column handle abstraction to agree on the identities of rows to be updated or deleted. The rowId column handle is opaque to the Trino engine. Depending on the connector, the rowId column handle abstraction could represent several physical columns. For the JDBC connector, the rowId column handle points might be the primary key for the table. For deletion in Hive ACID tables, the rowId consists of the three ACID columns that uniquely identify rows.

### The rowId Column for DELETE#

The Trino engine identifies the rows to be deleted using a connector-specific rowId column handle, returned by the connector’s ConnectorMetadata.getDeleteRowIdColumnHandle() method, whose full signature is:

ColumnHandle getDeleteRowIdColumnHandle(
ConnectorSession session,
ConnectorTableHandle tableHandle)


### The rowId Column for UPDATE#

The Trino engine identifies rows to be updated using a connector-specific rowId column handle, returned by the connector’s ConnectorMetadata.getUpdateRowIdColumnHandle() method. In addition to the columns that identify the row, for UPDATE the rowId column will contain any columns that the connector requires in order to perform the UPDATE operation. In Hive ACID, for example, the rowId column contains the values of all columns not updated by the UPDATE operation, since Hive ACID implements UPDATE as a DELETE paired with an INSERT.

## UpdatablePageSource API#

As mentioned above, to support DELETE or UPDATE, the connector must define a subclass of UpdatablePageSource, layered over the connector’s usual ConnectorPageSource. The interesting methods are:

• Page getNextPage(). When the Trino engine calls getNextPage(), the UpdatablePageSource calls its underlying ConnectorPageSource.getNextPage() method to get a page. Some connectors will reformat the page before returning it to the Trino engine.

• void deleteRows(Block rowIds). The Trino engine calls the deleteRows() method of the same UpdatablePageSource instance that supplied the original page, passing a block of “rowIds”, created by the Trino engine based on the column handle returned by ConnectorMetadata.getDeleteRowIdColumnHandle()

• void updateRows(Page page, List<Integer> columnValueAndRowIdChannels). The Trino engine calls the updateRows() method of the same UpdatablePageSource instance that supplied the original page, passing a page of projected columns, one for each updated column and the last one for the rowId column. The order of projected columns is defined by the Trino engine, and that order is reflected in the columnValueAndRowIdChannels argument. The job of updateRows() is to:

• Extract the updated column blocks and the rowId block from the projected page.

• Assemble them in whatever order is required by the connector for storage.

• Store the update result in the underlying file store.

In the case of Hive ACID, updateRows() stores a file of records that delete the previous contents of the updated rows, and a separate file that inserts completely new rows containing the updated and non-updated column values.

• CompletableFuture<Collection<Slice>> finish(). The Trino engine calls finish() when all the pages of a split have been processed. The connector returns a future containing a collection of Slice, representing connector-specific information about the rows processed. Usually this will include the row count, and might include information like the files or partitions created or changed.

## ConnectorMetadataDELETE API#

A connector implementing DELETE must specify three ConnectorMetadata methods.

• getDeleteRowIdColumnHandle():

ColumnHandle getDeleteRowIdColumnHandle(
ConnectorSession session,
ConnectorTableHandle tableHandle)


The ColumnHandle returned by this method provides the “rowId” used by the connector to identify rows to be deleted, as well as any other fields of the row that the connector will need to complete the DELETE operation. For a JDBC connector, that rowId is usually the primary key for the table and no other fields are required. For other connectors, the information needed to identify a row usually consists of multiple physical columns. With the Hive connector and a Hive ACID table, for example, the DELETE rowId consists of the three ORC ACID columns that identify the row.

• beginDelete():

ConnectorTableHandle beginDelete(
ConnectorSession session,
ConnectorTableHandle tableHandle)


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

beginDelete() performs any orchestration needed in the connector to start processing the DELETE. This orchestration varies from connector to connector. In the Hive ACID connector, for example, beginDelete() checks that the table is transactional and starts a Hive Metastore transaction.

beginDelete() returns a ConnectorTableHandle with any added information the connector needs when the handle is passed back to finishDelete() and the split generation machinery. For most connectors, the returned table handle contains a flag identifying the table handle as a table handle for a DELETE operation.

• finishDelete():

void finishDelete(
ConnectorSession session,
ConnectorTableHandle tableHandle,
Collection<Slice> fragments)


During DELETE processing, the Trino engine accumulates the Slice collections returned by UpdatablePageSource.finish(). After all splits have been processed, the engine calls finishDelete(), passing the table handle and that collection of Slice fragments. In response, the connector takes appropriate actions to complete the Delete operation. Those actions might include committing the transaction, assuming the connector supports a transaction paradigm.

## ConnectorMetadataUPDATE API#

A connector implementing UPDATE must specify three ConnectorMetadata methods.

• getUpdateRowIdColumnHandle:

ColumnHandle getUpdateRowIdColumnHandle(
ConnectorSession session,
ConnectorTableHandle tableHandle,
List<ColumnHandle> updatedColumns)


The updatedColumns list contains column handles for all columns updated by the UPDATE operation in table column order.

The ColumnHandle returned by this method provides the “rowId” used by the connector to identify rows to be updated, as well as any other fields of the row that the connector will need to complete the UPDATE operation. For a JDBC connector, that rowId is usually the primary key for the table and no other fields are required. For other connectors, the information needed to identify a row usually consists of multiple physical columns. Moreover, some connectors may need the values of columns that are not updated to complete the UPDATE operation. With the Hive connector and a Hive ACID table, for example, the UPDATE rowId consists of the three ORC ACID columns that identify the row, plus the values of all the data columns not updated.

• beginUpdate:

ConnectorTableHandle beginUpdate(
ConnectorSession session,
ConnectorTableHandle tableHandle,
List<ColumnHandle> updatedColumns)


As the last step in creating the UPDATE execution plan, the connector’s beginUpdate() method is called, passing arguments that define the UPDATE to the connector. In addition to the session and tableHandle, the arguments includes the list of the updated columns handles, in table column order.

beginUpdate() performs any orchestration needed in the connector to start processing the UPDATE. This orchestration varies from connector to connector. In the Hive ACID connector, for example, beginUpdate() starts the Hive Metastore transaction; checks that the updated table is transactional and that neither parition columns nor bucket columns are updated.

beginUpdate returns a ConnectorTableHandle with any added information the connector needs when the handle is passed back to finishUpdate() and the split generation machinery. For most connectors, the returned table handle contains a flag identifying the table handle as a table handle for a UPDATE operation. For some connectors that support partitioning, the table handle will reflect that partitioning.

• finishUpdate:

void finishUpdate(
ConnectorSession session,
ConnectorTableHandle tableHandle,
Collection<Slice> fragments)


During UPDATE processing, the Trino engine accumulates the Slice collections returned by UpdatablePageSource.finish(). After all splits have been processed, the engine calls finishUpdate(), passing the table handle and that collection of Slice fragments. In response, the connector takes appropriate actions to complete the UPDATE operation. Those actions might include committing the transaction, assuming the connector supports a transaction paradigm.