9 PostgreSQL includes an implementation of persistent on-disk hash
10 indexes, which are fully crash recoverable. Any data type can be
11 indexed by a hash index, including data types that do not have a
12 well-defined linear ordering. Hash indexes store only the hash value of
13 the data being indexed, thus there are no restrictions on the size of
14 the data column being indexed.
16 Hash indexes support only single-column indexes and do not allow
19 Hash indexes support only the = operator, so WHERE clauses that specify
20 range operations will not be able to take advantage of hash indexes.
22 Each hash index tuple stores just the 4-byte hash value, not the actual
23 column value. As a result, hash indexes may be much smaller than
24 B-trees when indexing longer data items such as UUIDs, URLs, etc. The
25 absence of the column value also makes all hash index scans lossy. Hash
26 indexes may take part in bitmap index scans and backward scans.
28 Hash indexes are best optimized for SELECT and UPDATE-heavy workloads
29 that use equality scans on larger tables. In a B-tree index, searches
30 must descend through the tree until the leaf page is found. In tables
31 with millions of rows, this descent can increase access time to data.
32 The equivalent of a leaf page in a hash index is referred to as a
33 bucket page. In contrast, a hash index allows accessing the bucket
34 pages directly, thereby potentially reducing index access time in
35 larger tables. This reduction in "logical I/O" becomes even more
36 pronounced on indexes/data larger than shared_buffers/RAM.
38 Hash indexes have been designed to cope with uneven distributions of
39 hash values. Direct access to the bucket pages works well if the hash
40 values are evenly distributed. When inserts mean that the bucket page
41 becomes full, additional overflow pages are chained to that specific
42 bucket page, locally expanding the storage for index tuples that match
43 that hash value. When scanning a hash bucket during queries, we need to
44 scan through all of the overflow pages. Thus an unbalanced hash index
45 might actually be worse than a B-tree in terms of number of block
46 accesses required, for some data.
48 As a result of the overflow cases, we can say that hash indexes are
49 most suitable for unique, nearly unique data or data with a low number
50 of rows per hash bucket. One possible way to avoid problems is to
51 exclude highly non-unique values from the index using a partial index
52 condition, but this may not be suitable in many cases.
54 Like B-Trees, hash indexes perform simple index tuple deletion. This is
55 a deferred maintenance operation that deletes index tuples that are
56 known to be safe to delete (those whose item identifier's LP_DEAD bit
57 is already set). If an insert finds no space is available on a page we
58 try to avoid creating a new overflow page by attempting to remove dead
59 index tuples. Removal cannot occur if the page is pinned at that time.
60 Deletion of dead index pointers also occurs during VACUUM.
62 If it can, VACUUM will also try to squeeze the index tuples onto as few
63 overflow pages as possible, minimizing the overflow chain. If an
64 overflow page becomes empty, overflow pages can be recycled for reuse
65 in other buckets, though we never return them to the operating system.
66 There is currently no provision to shrink a hash index, other than by
67 rebuilding it with REINDEX. There is no provision for reducing the
68 number of buckets, either.
70 Hash indexes may expand the number of bucket pages as the number of
71 rows indexed grows. The hash key-to-bucket-number mapping is chosen so
72 that the index can be incrementally expanded. When a new bucket is to
73 be added to the index, exactly one existing bucket will need to be
74 "split", with some of its tuples being transferred to the new bucket
75 according to the updated key-to-bucket-number mapping.
77 The expansion occurs in the foreground, which could increase execution
78 time for user inserts. Thus, hash indexes may not be suitable for
79 tables with rapidly increasing number of rows.
81 65.6.2. Implementation #
83 There are four kinds of pages in a hash index: the meta page (page
84 zero), which contains statically allocated control information; primary
85 bucket pages; overflow pages; and bitmap pages, which keep track of
86 overflow pages that have been freed and are available for re-use. For
87 addressing purposes, bitmap pages are regarded as a subset of the
90 Both scanning the index and inserting tuples require locating the
91 bucket where a given tuple ought to be located. To do this, we need the
92 bucket count, highmask, and lowmask from the metapage; however, it's
93 undesirable for performance reasons to have to have to lock and pin the
94 metapage for every such operation. Instead, we retain a cached copy of
95 the metapage in each backend's relcache entry. This will produce the
96 correct bucket mapping as long as the target bucket hasn't been split
97 since the last cache refresh.
99 Primary bucket pages and overflow pages are allocated independently
100 since any given index might need more or fewer overflow pages relative
101 to its number of buckets. The hash code uses an interesting set of
102 addressing rules to support a variable number of overflow pages while
103 not having to move primary bucket pages around after they are created.
105 Each row in the table indexed is represented by a single index tuple in
106 the hash index. Hash index tuples are stored in bucket pages, and if
107 they exist, overflow pages. We speed up searches by keeping the index
108 entries in any one index page sorted by hash code, thus allowing binary
109 search to be used within an index page. Note however that there is *no*
110 assumption about the relative ordering of hash codes across different
111 index pages of a bucket.
113 The bucket splitting algorithms to expand the hash index are too
114 complex to be worthy of mention here, though are described in more
115 detail in src/backend/access/hash/README. The split algorithm is crash
116 safe and can be restarted if not completed successfully.