This is an automated email from the ASF dual-hosted git repository.
dzamo pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/drill-site.git
The following commit(s) were added to refs/heads/master by this push:
new 322040cf3 Document new implicit casting.
322040cf3 is described below
commit 322040cf3ba8d793bfc0320d742d61487581109e
Author: James Turton <[email protected]>
AuthorDate: Mon Mar 6 16:21:00 2023 +0200
Document new implicit casting.
---
.../data-types/010-supported-data-types.md | 49 ++++++++++++----------
1 file changed, 28 insertions(+), 21 deletions(-)
diff --git a/_docs/en/sql-reference/data-types/010-supported-data-types.md
b/_docs/en/sql-reference/data-types/010-supported-data-types.md
index a12671fd1..2eb11f971 100644
--- a/_docs/en/sql-reference/data-types/010-supported-data-types.md
+++ b/_docs/en/sql-reference/data-types/010-supported-data-types.md
@@ -124,27 +124,34 @@ In some cases, Drill converts schema-less data to
correctly-typed data implicitl
## Implicit Casting Precedence of Data Types
-The following list includes data types Drill uses in descending order of
precedence. Casting precedence shown in the following table applies to the
implicit casting that Drill performs. For example, Drill might implicitly cast
data when a query includes a function or filter on mismatched data types:
-
- SELECT myBigInt FROM mytable WHERE myBigInt = 2.5;
-
-As shown in the table, Drill can cast a NULL value, which has the lowest
precedence, to any other type; you can cast a SMALLINT (not supported in this
release) value to INT. Drill might deviate from these precedence rules for
performance reasons. Under certain circumstances, such as queries involving
SUBSTR and CONCAT functions, Drill reverses the order of precedence and allows
a cast to VARCHAR from a type of higher precedence than VARCHAR, such as
BIGINT. The INTERVALDAY and INTERVALYE [...]
-
-### Casting Precedence
-
-| Precedence | Data Type | Precedence | Data Type |
-|------------|------------------------|------------|---------------|
-| 1 | INTERVALYEAR (highest) | 12 | UINT2 |
-| 2 | INTERVALDAY | 13 | SMALLINT[^5] |
-| 3 | TIMESTAMP | 14 | UINT1 |
-| 4 | DATE | 15 | VAR16CHAR |
-| 5 | TIME | 16 | FIXED16CHAR |
-| 6 | DOUBLE | 17 | VARCHAR |
-| 7 | DECIMAL | 18 | CHAR |
-| 8 | UINT8 | 19 | VARBINARY |
-| 9 | BIGINT | 20 | FIXEDBINARY |
-| 10 | UINT4 | 21 | NULL (lowest) |
-| 11 | INT | | |
+Drill's implicit casting logic was overhauled in version 1.21, changing
internally from a linear ranking of the SQL data types by "precedence" to a
directed graph over the SQL data types which encodes the cost[^1] of casting
between them in a transitive way. Without detailing the cost of every cast, the
following principles give a good idea of what implicit cast Drill will attempt
to insert, if any.
+
+[^1]: Casting costs in Drill are set so as to achieve a target set of query
behaviours, as measured by the test suite. They are not the computational cost
nor are do they derive from any similar single principle.
+
+1. The cost of casting from type A to type B is the path of least cost over
the data type DAG.
+2. The NULL type is castable to any type with VARCHAR being the cheapest.
+3. It is cheaper to cast a NULL argument to any non-NULL type than to cast a
non-NULL to another non-NULL.
+4. One expensive cast is generally cheaper than multiple cheap casts.
+5. Where type widenings naturally occur, e.g. INT4 → INT8, these casts are
cheap.
+
+Here are three examples of implcit casting at work.
+
+```
+select sqrt('2');
+EXPR$0 1.4142135623730951
+
+1 row selected (0.24 seconds)
+
+select current_date > '1970-01-01';
+EXPR$0 true
+
+1 row selected (0.262 seconds)
+
+select 'False' = false;
+EXPR$0 true
+
+1 row selected (0.133 seconds)
+```
## Explicit Casting
