Version 3.2.1
Custom Extractors

Introduction to Custom Extractors

A WiredTiger table can have zero or more associated indices. An index uses a different key to locate records than the table, and usually only stores a short key for each record, with the (larger) value in the table.

WiredTiger tables must be created with column names in order to create an index. This is required so that index cursors can support projections, and because WiredTiger optimizes some cases of "simple" tables without column names.

When the full schema of your records can be described in WiredTiger's packing format, you can create an index by specifying which columns from the record should appear in the index. However, for more complex records, or to associate multiple index keys to each record, applications can instead specify a custom extractor by implementing the WT_EXTRACTOR interface.

The main method in the interface is WT_EXTRACTOR::extract. This is called by WiredTiger each time a record is updated in a table. The extract method should determine the index key(s) and call WT_CURSOR::set_key followed by WT_CURSOR::insert on the supplied result_cursor for each index key.

If any operation fails, WT_EXTRACTOR::extract must return the failure to WiredTiger, or the index could become out of sync with the table.

Note that the extract callback is called for all operations that update the table, not just inserts. The callback sets the key and uses the WT_CURSOR::insert method to return the index key(s). WiredTiger will perform the required operation to keep the index in sync with the table.

Applications must register their WT_EXTRACTOR implementations using WT_CONNECTION::add_extractor. This is often done by creating a WiredTiger extension. They are then configured by passing "extractor=..." to WT_SESSION::create when creating an index.

See ex_extractor.c for an example of how to implement custom extractors.

Implementation notes

A WiredTiger index is a row store where the key columns contain all of the secondary and primary key columns, but only the secondary key columns are visible to applications. The value is empty, and WiredTiger's on-disk format optimizes for this case (empty values take up no space on disk).

Custom extractors only need to calculate the public index key columns. The result_cursor will be configured with a key_format corresponding to what was supplied to WT_SESSION::create when the index was created. WiredTiger will append the (hidden) primary key when populating the index.

If column names are specified for an index with a custom extractor, it is not permitted to use any column names from the table key. Custom index keys can include columns from the table value, but the extracted value must be equal to the value from that column of the record or the results of using a projection cursor on the index will be undefined.

Custom Collators in raw mode

If a custom extractor needs to operate in raw mode on the result_cursor, it must take into account an implementation detail. To avoid rewriting the extracted key, WiredTiger appends a padding byte to the raw key using a 'x' format. See Format types for more information. If the callback operates in raw mode, it must also append this padding byte.