67.3. B-Tree Support Functions
As shown in Table 38.9, btree defines one required and four optional support functions. The five user-defined methods are:
order
For each combination of data types that a btree operator family provides comparison operators for, it must provide a comparison support function, registered in
pg_amproc
with support function number 1 andamproclefttype
/amprocrighttype
equal to the left and right data types for the comparison (i.e., the same data types that the matching operators are registered with inpg_amop
). The comparison function must take two non-null valuesA
andB
and return anint32
value that is<
0
,0
, or>
0
whenA
<
B
,A
=
B
, orA
>
B
, respectively. A null result is disallowed: all values of the data type must be comparable. Seesrc/backend/access/nbtree/nbtcompare.c
for examples.If the compared values are of a collatable data type, the appropriate collation OID will be passed to the comparison support function, using the standard
PG_GET_COLLATION()
mechanism.sortsupport
Optionally, a btree operator family may provide sort support function(s), registered under support function number 2. These functions allow implementing comparisons for sorting purposes in a more efficient way than naively calling the comparison support function. The APIs involved in this are defined in
src/include/utils/sortsupport.h
.in_range
Optionally, a btree operator family may provide in_range support function(s), registered under support function number 3. These are not used during btree index operations; rather, they extend the semantics of the operator family so that it can support window clauses containing the
RANGE
offset
PRECEDING
andRANGE
offset
FOLLOWING
frame bound types (see Section 4.2.8). Fundamentally, the extra information provided is how to add or subtract anoffset
value in a way that is compatible with the family's data ordering.An
in_range
function must have the signaturein_range(
val
type1,base
type1,offset
type2,sub
bool,less
bool) returns boolval
andbase
must be of the same type, which is one of the types supported by the operator family (i.e., a type for which it provides an ordering). However,offset
could be of a different type, which might be one otherwise unsupported by the family. An example is that the built-intime_ops
family provides anin_range
function that hasoffset
of typeinterval
. A family can providein_range
functions for any of its supported types and one or moreoffset
types. Eachin_range
function should be entered inpg_amproc
withamproclefttype
equal totype1
andamprocrighttype
equal totype2
.The essential semantics of an
in_range
function depend on the two Boolean flag parameters. It should add or subtractbase
andoffset
, then compareval
to the result, as follows:if
!
sub
and!
less
, returnval
>=
(base
+
offset
)if
!
sub
andless
, returnval
<=
(base
+
offset
)if
sub
and!
less
, returnval
>=
(base
-
offset
)if
sub
andless
, returnval
<=
(base
-
offset
)
Before doing so, the function should check the sign of
offset
: if it is less than zero, raise errorERRCODE_INVALID_PRECEDING_OR_FOLLOWING_SIZE
(22013) with error text like “invalid preceding or following size in window function”. (This is required by the SQL standard, although nonstandard operator families might perhaps choose to ignore this restriction, since there seems to be little semantic necessity for it.) This requirement is delegated to thein_range
function so that the core code needn't understand what “less than zero” means for a particular data type.An additional expectation is that
in_range
functions should, if practical, avoid throwing an error ifbase
+
offset
orbase
-
offset
would overflow. The correct comparison result can be determined even if that value would be out of the data type's range. Note that if the data type includes concepts such as “infinity” or “NaN”, extra care may be needed to ensure thatin_range
's results agree with the normal sort order of the operator family.The results of the
in_range
function must be consistent with the sort ordering imposed by the operator family. To be precise, given any fixed values ofoffset
andsub
, then:If
in_range
withless
= true is true for someval1
andbase
, it must be true for everyval2
<=
val1
with the samebase
.If
in_range
withless
= true is false for someval1
andbase
, it must be false for everyval2
>=
val1
with the samebase
.If
in_range
withless
= true is true for someval
andbase1
, it must be true for everybase2
>=
base1
with the sameval
.If
in_range
withless
= true is false for someval
andbase1
, it must be false for everybase2
<=
base1
with the sameval
.
Analogous statements with inverted conditions hold when
less
= false.If the type being ordered (
type1
) is collatable, the appropriate collation OID will be passed to thein_range
function, using the standard PG_GET_COLLATION() mechanism.in_range
functions need not handle NULL inputs, and typically will be marked strict.equalimage
Optionally, a btree operator family may provide
equalimage
(“equality implies image equality”) support functions, registered under support function number 4. These functions allow the core code to determine when it is safe to apply the btree deduplication optimization. Currently,equalimage
functions are only called when building or rebuilding an index.An
equalimage
function must have the signatureequalimage(
opcintype
oid
) returns boolThe return value is static information about an operator class and collation. Returning
true
indicates that theorder
function for the operator class is guaranteed to only return0
(“arguments are equal”) when itsA
andB
arguments are also interchangeable without any loss of semantic information. Not registering anequalimage
function or returningfalse
indicates that this condition cannot be assumed to hold.The
opcintype
argument is the
of the data type that the operator class indexes. This is a convenience that allows reuse of the same underlyingpg_type
.oidequalimage
function across operator classes. Ifopcintype
is a collatable data type, the appropriate collation OID will be passed to theequalimage
function, using the standardPG_GET_COLLATION()
mechanism.As far as the operator class is concerned, returning
true
indicates that deduplication is safe (or safe for the collation whose OID was passed to itsequalimage
function). However, the core code will only deem deduplication safe for an index when every indexed column uses an operator class that registers anequalimage
function, and each function actually returnstrue
when called.Image equality is almost the same condition as simple bitwise equality. There is one subtle difference: When indexing a varlena data type, the on-disk representation of two image equal datums may not be bitwise equal due to inconsistent application of TOAST compression on input. Formally, when an operator class's
equalimage
function returnstrue
, it is safe to assume that thedatum_image_eq()
C function will always agree with the operator class'sorder
function (provided that the same collation OID is passed to both theequalimage
andorder
functions).The core code is fundamentally unable to deduce anything about the “equality implies image equality” status of an operator class within a multiple-data-type family based on details from other operator classes in the same family. Also, it is not sensible for an operator family to register a cross-type
equalimage
function, and attempting to do so will result in an error. This is because “equality implies image equality” status does not just depend on sorting/equality semantics, which are more or less defined at the operator family level. In general, the semantics that one particular data type implements must be considered separately.The convention followed by the operator classes included with the core PostgreSQL distribution is to register a stock, generic
equalimage
function. Most operator classes registerbtequalimage()
, which indicates that deduplication is safe unconditionally. Operator classes for collatable data types such astext
registerbtvarstrequalimage()
, which indicates that deduplication is safe with deterministic collations. Best practice for third-party extensions is to register their own custom function to retain control.options
Optionally, a B-tree operator family may provide
options
(“operator class specific options”) support functions, registered under support function number 5. These functions define a set of user-visible parameters that control operator class behavior.An
options
support function must have the signatureoptions(
relopts
local_relopts *
) returns voidThe function is passed a pointer to a
local_relopts
struct, which needs to be filled with a set of operator class specific options. The options can be accessed from other support functions using thePG_HAS_OPCLASS_OPTIONS()
andPG_GET_OPCLASS_OPTIONS()
macros.Currently, no B-Tree operator class has an
options
support function. B-tree doesn't allow flexible representation of keys like GiST, SP-GiST, GIN and BRIN do. So,options
probably doesn't have much application in the current B-tree index access method. Nevertheless, this support function was added to B-tree for uniformity, and will probably find uses during further evolution of B-tree in PostgreSQL.