4 66.2.1. Out-of-Line, On-Disk TOAST Storage
5 66.2.2. Out-of-Line, In-Memory TOAST Storage
7 This section provides an overview of TOAST (The Oversized-Attribute
10 PostgreSQL uses a fixed page size (commonly 8 kB), and does not allow
11 tuples to span multiple pages. Therefore, it is not possible to store
12 very large field values directly. To overcome this limitation, large
13 field values are compressed and/or broken up into multiple physical
14 rows. This happens transparently to the user, with only small impact on
15 most of the backend code. The technique is affectionately known as
16 TOAST (or “the best thing since sliced bread”). The TOAST
17 infrastructure is also used to improve handling of large data values
20 Only certain data types support TOAST — there is no need to impose the
21 overhead on data types that cannot produce large field values. To
22 support TOAST, a data type must have a variable-length (varlena)
23 representation, in which, ordinarily, the first four-byte word of any
24 stored value contains the total length of the value in bytes (including
25 itself). TOAST does not constrain the rest of the data type's
26 representation. The special representations collectively called TOASTed
27 values work by modifying or reinterpreting this initial length word.
28 Therefore, the C-level functions supporting a TOAST-able data type must
29 be careful about how they handle potentially TOASTed input values: an
30 input might not actually consist of a four-byte length word and
31 contents until after it's been detoasted. (This is normally done by
32 invoking PG_DETOAST_DATUM before doing anything with an input value,
33 but in some cases more efficient approaches are possible. See
34 Section 36.13.1 for more detail.)
36 TOAST usurps two bits of the varlena length word (the high-order bits
37 on big-endian machines, the low-order bits on little-endian machines),
38 thereby limiting the logical size of any value of a TOAST-able data
39 type to 1 GB (2^30 - 1 bytes). When both bits are zero, the value is an
40 ordinary un-TOASTed value of the data type, and the remaining bits of
41 the length word give the total datum size (including length word) in
42 bytes. When the highest-order or lowest-order bit is set, the value has
43 only a single-byte header instead of the normal four-byte header, and
44 the remaining bits of that byte give the total datum size (including
45 length byte) in bytes. This alternative supports space-efficient
46 storage of values shorter than 127 bytes, while still allowing the data
47 type to grow to 1 GB at need. Values with single-byte headers aren't
48 aligned on any particular boundary, whereas values with four-byte
49 headers are aligned on at least a four-byte boundary; this omission of
50 alignment padding provides additional space savings that is significant
51 compared to short values. As a special case, if the remaining bits of a
52 single-byte header are all zero (which would be impossible for a
53 self-inclusive length), the value is a pointer to out-of-line data,
54 with several possible alternatives as described below. The type and
55 size of such a TOAST pointer are determined by a code stored in the
56 second byte of the datum. Lastly, when the highest-order or
57 lowest-order bit is clear but the adjacent bit is set, the content of
58 the datum has been compressed and must be decompressed before use. In
59 this case the remaining bits of the four-byte length word give the
60 total size of the compressed datum, not the original data. Note that
61 compression is also possible for out-of-line data but the varlena
62 header does not tell whether it has occurred — the content of the TOAST
63 pointer tells that, instead.
65 The compression technique used for either in-line or out-of-line
66 compressed data can be selected for each column by setting the
67 COMPRESSION column option in CREATE TABLE or ALTER TABLE. The default
68 for columns with no explicit setting is to consult the
69 default_toast_compression parameter at the time data is inserted.
71 As mentioned, there are multiple types of TOAST pointer datums. The
72 oldest and most common type is a pointer to out-of-line data stored in
73 a TOAST table that is separate from, but associated with, the table
74 containing the TOAST pointer datum itself. These on-disk pointer datums
75 are created by the TOAST management code (in
76 access/common/toast_internals.c) when a tuple to be stored on disk is
77 too large to be stored as-is. Further details appear in Section 66.2.1.
78 Alternatively, a TOAST pointer datum can contain a pointer to
79 out-of-line data that appears elsewhere in memory. Such datums are
80 necessarily short-lived, and will never appear on-disk, but they are
81 very useful for avoiding copying and redundant processing of large data
82 values. Further details appear in Section 66.2.2.
84 66.2.1. Out-of-Line, On-Disk TOAST Storage #
86 If any of the columns of a table are TOAST-able, the table will have an
87 associated TOAST table, whose OID is stored in the table's
88 pg_class.reltoastrelid entry. On-disk TOASTed values are kept in the
89 TOAST table, as described in more detail below.
91 Out-of-line values are divided (after compression if used) into chunks
92 of at most TOAST_MAX_CHUNK_SIZE bytes (by default this value is chosen
93 so that four chunk rows will fit on a page, making it about 2000
94 bytes). Each chunk is stored as a separate row in the TOAST table
95 belonging to the owning table. Every TOAST table has the columns
96 chunk_id (an OID identifying the particular TOASTed value), chunk_seq
97 (a sequence number for the chunk within its value), and chunk_data (the
98 actual data of the chunk). A unique index on chunk_id and chunk_seq
99 provides fast retrieval of the values. A pointer datum representing an
100 out-of-line on-disk TOASTed value therefore needs to store the OID of
101 the TOAST table in which to look and the OID of the specific value (its
102 chunk_id). For convenience, pointer datums also store the logical datum
103 size (original uncompressed data length), physical stored size
104 (different if compression was applied), and the compression method
105 used, if any. Allowing for the varlena header bytes, the total size of
106 an on-disk TOAST pointer datum is therefore 18 bytes regardless of the
107 actual size of the represented value.
109 The TOAST management code is triggered only when a row value to be
110 stored in a table is wider than TOAST_TUPLE_THRESHOLD bytes (normally 2
111 kB). The TOAST code will compress and/or move field values out-of-line
112 until the row value is shorter than TOAST_TUPLE_TARGET bytes (also
113 normally 2 kB, adjustable) or no more gains can be had. During an
114 UPDATE operation, values of unchanged fields are normally preserved
115 as-is; so an UPDATE of a row with out-of-line values incurs no TOAST
116 costs if none of the out-of-line values change.
118 The TOAST management code recognizes four different strategies for
119 storing TOAST-able columns on disk:
120 * PLAIN prevents either compression or out-of-line storage. This is
121 the only possible strategy for columns of non-TOAST-able data
123 * EXTENDED allows both compression and out-of-line storage. This is
124 the default for most TOAST-able data types. Compression will be
125 attempted first, then out-of-line storage if the row is still too
127 * EXTERNAL allows out-of-line storage but not compression. Use of
128 EXTERNAL will make substring operations on wide text and bytea
129 columns faster (at the penalty of increased storage space) because
130 these operations are optimized to fetch only the required parts of
131 the out-of-line value when it is not compressed.
132 * MAIN allows compression but not out-of-line storage. (Actually,
133 out-of-line storage will still be performed for such columns, but
134 only as a last resort when there is no other way to make the row
135 small enough to fit on a page.)
137 Each TOAST-able data type specifies a default strategy for columns of
138 that data type, but the strategy for a given table column can be
139 altered with ALTER TABLE ... SET STORAGE.
141 TOAST_TUPLE_TARGET can be adjusted for each table using ALTER TABLE ...
142 SET (toast_tuple_target = N)
144 This scheme has a number of advantages compared to a more
145 straightforward approach such as allowing row values to span pages.
146 Assuming that queries are usually qualified by comparisons against
147 relatively small key values, most of the work of the executor will be
148 done using the main row entry. The big values of TOASTed attributes
149 will only be pulled out (if selected at all) at the time the result set
150 is sent to the client. Thus, the main table is much smaller and more of
151 its rows fit in the shared buffer cache than would be the case without
152 any out-of-line storage. Sort sets shrink also, and sorts will more
153 often be done entirely in memory. A little test showed that a table
154 containing typical HTML pages and their URLs was stored in about half
155 of the raw data size including the TOAST table, and that the main table
156 contained only about 10% of the entire data (the URLs and some small
157 HTML pages). There was no run time difference compared to an un-TOASTed
158 comparison table, in which all the HTML pages were cut down to 7 kB to
161 66.2.2. Out-of-Line, In-Memory TOAST Storage #
163 TOAST pointers can point to data that is not on disk, but is elsewhere
164 in the memory of the current server process. Such pointers obviously
165 cannot be long-lived, but they are nonetheless useful. There are
166 currently two sub-cases: pointers to indirect data and pointers to
169 Indirect TOAST pointers simply point at a non-indirect varlena value
170 stored somewhere in memory. This case was originally created merely as
171 a proof of concept, but it is currently used during logical decoding to
172 avoid possibly having to create physical tuples exceeding 1 GB (as
173 pulling all out-of-line field values into the tuple might do). The case
174 is of limited use since the creator of the pointer datum is entirely
175 responsible that the referenced data survives for as long as the
176 pointer could exist, and there is no infrastructure to help with this.
178 Expanded TOAST pointers are useful for complex data types whose on-disk
179 representation is not especially suited for computational purposes. As
180 an example, the standard varlena representation of a PostgreSQL array
181 includes dimensionality information, a nulls bitmap if there are any
182 null elements, then the values of all the elements in order. When the
183 element type itself is variable-length, the only way to find the N'th
184 element is to scan through all the preceding elements. This
185 representation is appropriate for on-disk storage because of its
186 compactness, but for computations with the array it's much nicer to
187 have an “expanded” or “deconstructed” representation in which all the
188 element starting locations have been identified. The TOAST pointer
189 mechanism supports this need by allowing a pass-by-reference Datum to
190 point to either a standard varlena value (the on-disk representation)
191 or a TOAST pointer that points to an expanded representation somewhere
192 in memory. The details of this expanded representation are up to the
193 data type, though it must have a standard header and meet the other API
194 requirements given in src/include/utils/expandeddatum.h. C-level
195 functions working with the data type can choose to handle either
196 representation. Functions that do not know about the expanded
197 representation, but simply apply PG_DETOAST_DATUM to their inputs, will
198 automatically receive the traditional varlena representation; so
199 support for an expanded representation can be introduced incrementally,
200 one function at a time.
202 TOAST pointers to expanded values are further broken down into
203 read-write and read-only pointers. The pointed-to representation is the
204 same either way, but a function that receives a read-write pointer is
205 allowed to modify the referenced value in-place, whereas one that
206 receives a read-only pointer must not; it must first create a copy if
207 it wants to make a modified version of the value. This distinction and
208 some associated conventions make it possible to avoid unnecessary
209 copying of expanded values during query execution.
211 For all types of in-memory TOAST pointer, the TOAST management code
212 ensures that no such pointer datum can accidentally get stored on disk.
213 In-memory TOAST pointers are automatically expanded to normal in-line
214 varlena values before storage — and then possibly converted to on-disk
215 TOAST pointers, if the containing tuple would otherwise be too big.