InnoDB performs a bulk load instead of
inserting one index record at a time when creating or rebuilding
indexes. This method of index creation is also known as a sorted
index build. Sorted index builds are not supported for spatial
There are three phases to an index build. In the first phase, the clustered index is scanned, and index entries are generated and added to the sort buffer. When the sort buffer becomes full, entries are sorted and written out to a temporary intermediate file. This process is also known as a “run”. In the second phase, with one or more runs written to the temporary intermediate file, a merge sort is performed on all entries in the file. In the third and final phase, the sorted entries are inserted into the B-tree.
Prior to the introduction of sorted index builds, index entries were inserted into the B-tree one record at a time using insert APIs. This method involved opening a B-tree cursor to find the insert position and then inserting entries into a B-tree page using an optimistic insert. If an insert failed due to a page being full, a pessimistic insert would be performed, which involves opening a B-tree cursor and splitting and merging B-tree nodes as necessary to find space for the entry. The drawbacks of this “top-down” method of building an index are the cost of searching for an insert position and the constant splitting and merging of B-tree nodes.
Sorted index builds use a “bottom-up” approach to building an index. With this approach, a reference to the right-most leaf page is held at all levels of the B-tree. The right-most leaf page at the necessary B-tree depth is allocated and entries are inserted according to their sorted order. Once a leaf page is full, a node pointer is appended to the parent page and a sibling leaf page is allocated for the next insert. This process continues until all entries are inserted, which may result in inserts up to the root level. When a sibling page is allocated, the reference to the previously pinned leaf page is released, and the newly allocated leaf page becomes the right-most leaf page and new default insert location.
To set aside space for future index growth, you can use the
option to reserve a percentage of B-tree page space. For example,
innodb_fill_factor to 80
reserves 20 percent of the space in B-tree pages during a sorted
index build. This setting applies to both B-tree leaf and non-leaf
pages. It does not apply to external pages used for
BLOB entries. The amount of space
that is reserved may not be exactly as configured, as the
innodb_fill_factor value is
interpreted as a hint rather than a hard limit.
Sorted index builds are supported for fulltext indexes. Previously, SQL was used to insert entries into a fulltext index.
For compressed tables, the previous index creation method appended entries to both compressed and uncompressed pages. When the modification log (representing free space on the compressed page) became full, the compressed page would be recompressed. If compression failed due to a lack of space, the page would be split. With sorted index builds, entries are only appended to uncompressed pages. When an uncompressed page becomes full, it is compressed. Adaptive padding is used to ensure that compression succeeds in most cases, but if compression fails, the page is split and compression is attempted again. This process continues until compression is successful. For more information about compression of B-Tree pages, see Section 18.104.22.168, “How Compression Works for InnoDB Tables”.
Redo logging is disabled during a sorted index build. Instead, there is a checkpoint to ensure that the index build can withstand an unexpected exit or failure. The checkpoint forces a write of all dirty pages to disk. During a sorted index build, the page cleaner thread is signaled periodically to flush dirty pages to ensure that the checkpoint operation can be processed quickly. Normally, the page cleaner thread flushes dirty pages when the number of clean pages falls below a set threshold. For sorted index builds, dirty pages are flushed promptly to reduce checkpoint overhead and to parallelize I/O and CPU activity.
Sorted index builds may result in optimizer statistics that differ from those generated by the previous method of index creation. The difference in statistics, which is not expected to affect workload performance, is due to the different algorithm used to populate the index.