*** DRAFT ***
A "value" is a single number, string, BLOB or NULL. Sometimes the qualified name "scalar value" is used to emphasize that only a single quantity is involved.
A "row value" is an ordered list of two or more scalar values. In other words, a "row value" is a vector or tuple.
The "size" of a row value is the number of scalar values the row value contains. The size of a row value is always at least 2. A row value with a single column is just a scalar value. A row value with no columns is a syntax error.
SQLite allows row values to be expressed in two ways:
SQLite can use row values in two contexts:
The syntax for row values and the circumstances in which row values can be used are illustrated in examples below.
Two row values are compared by looking at the constituent scalar values from left to right. A NULL means of "unknown". The overall result of comparison is NULL if it is possible to make the result either true or false by substituting alternative values in place of the constituent NULLs. The following query demonstrates some row value comparisons:
SELECT (1,2,3) = (1,2,3), -- 1 (1,2,3) = (1,NULL,3), -- NULL (1,2,3) = (1,NULL,4), -- 0 (1,2,3) < (2,3,4), -- 1 (1,2,3) < (1,2,4), -- 1 (1,2,3) < (1,3,NULL), -- 1 (1,2,3) < (1,2,NULL), -- NULL (1,3,5) < (1,2,NULL), -- 0 (1,2,NULL) IS (1,2,NULL); -- 1
The result of "(1,2,3)=(1,NULL,3)" is NULL because the result might be true if we replaced NULL→2 or false if we replaced NULL→9. The result of "(1,2,3)=(1,NULL,4)" is not NULL because there is no substitutions of the constituent NULL that will make the expression true, since 3 will never equal 4 in the third column.
Any of the row values in the previous example could be replace by a subquery that returns three columns and the same answer would result. For example:
CREATE TABLE t1(a,b,c); INSERT INTO t1(a,b,c) VALUES(1,2,3); SELECT (1,2,3)=(SELECT * FROM t1); -- 1
For a row-value IN operator, the left-hand side (hereafter "LHS") can be either a parenthesized list of values or a subquery with multiple columns. But the right-hand side (hereafter "RHS") must be a subquery expression.
CREATE TABLE t2(x,y,z); INSERT INTO t2(x,y,z) VALUES(1,2,3),(2,3,4),(1,NULL,5); SELECT (1,2,3) IN (SELECT * FROM t2), -- 1 (7,8,9) IN (SELECT * FROM t2), -- 0 (1,3,5) IN (SELECT * FROM t2); -- NULL
Row values can also be used in the SET clause of an UPDATE statement. The LHS must be a list of column names. The RHS can be any row value. For example:
UPDATE tab3
SET (a,b,c) = (SELECT x,y,z
FROM tab4
WHERE tab4.w=tab3.d)
WHERE tab3.e BETWEEN 55 AND 66;
Suppose an application wants to display a list of contacts in alphabetical order by lastname, firstname, in a scrolling window that can only show 7 contacts at a time. Initialize the scrolling window to the first 7 entries is easy:
SELECT * FROM contacts ORDER BY lastname, firstname LIMIT 7;
When the user scrolls down, the application needs to find the second set of 7 entries. One way to do this is to use the OFFSET clause:
SELECT * FROM contacts ORDER BY lastname, firstname LIMIT 7 OFFSET 7;
OFFSET gives the correct answer. However, OFFSET requires time proportional to the offset value. What really happens with "LIMIT x OFFSET y" is that SQLite computes the query as "LIMIT x+y" and discards the first y values without returning them to the application. So as the window scrolls down toward the bottom of a long list, and the y value becomes larger and larger, successive offset computations take more and more time.
A more efficient approach is to remember the last entry currently displayed and then use a row value comparison in the WHERE clause:
SELECT * FROM contacts WHERE (lastname,firstname) > (?1,?2) ORDER BY lastname, firstname LIMIT 7;
If the lastname and firstname on the bottom row of the previous screen are bound to ?1 and ?2, then the query above computes the next 7 rows. And, assuming there is an appropriate index, it does so very efficiently — much more efficiently than OFFSET.
The usual way of storing a date in a database table is as a single field, as either a unix timestamp, a julian day number, or an ISO-8601 dates string. But some application store dates as three separate fields for the year, month, and day.
CREATE TABLE info( year INT, -- 4 digit year month INT, -- 1 through 12 day INT, -- 1 through 31 other_stuff BLOB -- blah blah blah );
When dates are stored this way, row value comparisons provide a convenient way to compare dates:
SELECT * FROM info WHERE (year,month,day) BETWEEN (2015,9,12) AND (2016,9,12);
Suppose we want to know the order number, product number, and quantity for any item in which the product number and quantity match the product number and quantity of any item in order number 365:
SELECT ordid, prodid, qty
FROM item
WHERE (prodid, qty) IN (SELECT prodid, qty
FROM item
WHERE ordid = 365);
The query above could be rewritten as a join and without the use of row values:
SELECT t1.ordid, t1.prodid, t1.qty FROM item AS t1, item AS t2 WHERE t1.prodid=t2.prodid AND t1.qty=t2.qty AND t2.ordid=365;
Because the same query could be written without the use of row values, row values do not provide new capabilities. However, many developers say that the row value format is easier to read, write, and debug.
Even in the JOIN form, the query can be made clearer through the use of row values:
SELECT t1.ordid, t1.prodid, t1.qty FROM item AS t1, item AS t2 WHERE (t1.prodid,t1.qty) = (t2.prodid,t2.qty) AND t2.ordid=365;
This later query generates exactly the same bytecode as the previous scalar formulation, but using syntax that it cleaner and easier to read.
The row-value notation is useful for updating two or more columns of a table from the result of a single query. An example of this is in the full-text search feature of the Fossil version control system.
In the Fossil full-text search system, documents that participate in the full-text search (wiki pages, tickets, check-ins, documentation files, etc) are tracked by a table called "ftsdocs" (full text search documents). As new documents are added to the repository, they are not indexed right away. Indexing is deferred until there is a search request. The ftsdocs table contains an "idxed" field which is true if the document has been indexed and false if not.
When a search request occurs and pending documents are indexed for the first time, the ftsdocs table must be updated by setting the idxed column to true and also filling in several other columns with information pertinent to the search. That other information is obtained from a join. The query is this:
UPDATE ftsdocs SET
idxed=1,
name=NULL,
(label,url,mtime) =
(SELECT printf('Check-in [%%.16s] on %%s',blob.uuid,
datetime(event.mtime)),
printf('/timeline?y=ci&c=%%.20s',blob.uuid),
event.mtime
FROM event, blob
WHERE event.objid=ftsdocs.rid
AND blob.rid=ftsdocs.rid)
WHERE ftsdocs.type='c' AND NOT ftsdocs.idxed
(See the source code for further detail. Other examples here and here.)
Five out of nine columns in the ftsdocs table are updated. Two of the modified columns, "idxed" and "name", can be updated independently of the query. But the three columns "label", "url", and "mtime" all require a join query against the "event" and "blob" tables. Without row values, the equivalent UPDATE would require that the join be repeated three times, once for each column to be updated.
Sometimes the use of row values just makes the SQL easier to read and write. Consider the following two UPDATE statements:
UPDATE tab1 SET (a,b)=(b,a); UPDATE tab1 SET a=b, b=a;
Both UPDATE statements do exactly the same thing. (They generate identical bytecode.) But the first form, the row value form, seems to make it clearer that the intent of the statement is to swap the values in columns A and B.
Or consider these identical queries:
SELECT * FROM tab1 WHERE a=?1 AND b=?2; SELECT * FROM tab1 WHERE (a,b)=(?1,?2);
Once again, the SQL statements generate identical bytecode and thus do exactly the same job in exactly the same way. But the second form is made easier for humans to read by grouping the query parameters together into a single row value rather than scattering them across the WHERE clause.
Row values were added to SQLite version 3.15.0 (2016-10-14). Attempts to use row values in prior versions of SQLite will generate syntax errors.