PostgreSQL includes an implementation of persistent on-disk hash indexes, which are fully crash recoverable. Any data type can be indexed by a hash index, including data types that do not have a well-defined linear ordering. Hash indexes store only the hash value of the data being indexed, thus there are no restrictions on the size of the data column being indexed.

Hash indexes support only single-column indexes and do not allow uniqueness checking.

Hash indexes support only the = operator, so WHERE clauses that specify range operations will not be able to take advantage of hash indexes.

Each hash index tuple stores just the 4-byte hash value, not the actual column value. As a result, hash indexes may be much smaller than B-trees when indexing longer data items such as UUIDs, URLs, etc. The absence of the column value also makes all hash index scans lossy. Hash indexes may take part in bitmap index scans and backward scans.

Hash indexes are best optimized for SELECT and UPDATE-heavy workloads that use equality scans on larger tables. In a B-tree index, searches must descend through the tree until the leaf page is found. In tables with millions of rows, this descent can increase access time to data. The equivalent of a leaf page in a hash index is referred to as a bucket page. In contrast, a hash index allows accessing the bucket pages directly, thereby potentially reducing index access time in larger tables. This reduction in "logical I/O" becomes even more pronounced on indexes/data larger than shared_buffers/RAM.

Hash indexes have been designed to cope with uneven distributions of hash values. Direct access to the bucket pages works well if the hash values are evenly distributed. When inserts mean that the bucket page becomes full, additional overflow pages are chained to that specific bucket page, locally expanding the storage for index tuples that match that hash value. When scanning a hash bucket during queries, we need to scan through all of the overflow pages. Thus an unbalanced hash index might actually be worse than a B-tree in terms of number of block accesses required, for some data.

As a result of the overflow cases, we can say that hash indexes are most suitable for unique, nearly unique data or data with a low number of rows per hash bucket. One possible way to avoid problems is to exclude highly non-unique values from the index using a partial index condition, but this may not be suitable in many cases.

Like B-Trees, hash indexes perform simple index tuple deletion. This is a deferred maintenance operation that deletes index tuples that are known to be safe to delete (those whose item identifier's LP_DEAD bit is already set). If an insert finds no space is available on a page we try to avoid creating a new overflow page by attempting to remove dead index tuples. Removal cannot occur if the page is pinned at that time. Deletion of dead index pointers also occurs during VACUUM.

If it can, VACUUM will also try to squeeze the index tuples onto as few overflow pages as possible, minimizing the overflow chain. If an overflow page becomes empty, overflow pages can be recycled for reuse in other buckets, though we never return them to the operating system. There is currently no provision to shrink a hash index, other than by rebuilding it with REINDEX. There is no provision for reducing the number of buckets, either.

Hash indexes may expand the number of bucket pages as the number of rows indexed grows. The hash key-to-bucket-number mapping is chosen so that the index can be incrementally expanded. When a new bucket is to be added to the index, exactly one existing bucket will need to be "split", with some of its tuples being transferred to the new bucket according to the updated key-to-bucket-number mapping.

The expansion occurs in the foreground, which could increase execution time for user inserts. Thus, hash indexes may not be suitable for tables with rapidly increasing number of rows.

There are four kinds of pages in a hash index: the meta page (page zero), which contains statically allocated control information; primary bucket pages; overflow pages; and bitmap pages, which keep track of overflow pages that have been freed and are available for re-use. For addressing purposes, bitmap pages are regarded as a subset of the overflow pages.

Both scanning the index and inserting tuples require locating the bucket where a given tuple ought to be located. To do this, we need the bucket count, highmask, and lowmask from the metapage; however, it's undesirable for performance reasons to have to have to lock and pin the metapage for every such operation. Instead, we retain a cached copy of the metapage in each backend's relcache entry. This will produce the correct bucket mapping as long as the target bucket hasn't been split since the last cache refresh.

Primary bucket pages and overflow pages are allocated independently since any given index might need more or fewer overflow pages relative to its number of buckets. The hash code uses an interesting set of addressing rules to support a variable number of overflow pages while not having to move primary bucket pages around after they are created.

Each row in the table indexed is represented by a single index tuple in the hash index. Hash index tuples are stored in bucket pages, and if they exist, overflow pages. We speed up searches by keeping the index entries in any one index page sorted by hash code, thus allowing binary search to be used within an index page. Note however that there is *no* assumption about the relative ordering of hash codes across different index pages of a bucket.

The bucket splitting algorithms to expand the hash index are too complex to be worthy of mention here, though are described in more detail in src/backend/access/hash/README. The split algorithm is crash safe and can be restarted if not completed successfully.