Chapter 5: Types of Cursors, Shapes, OUT and OUT UNION, and FETCH

Chapter 5: Types of Cursors, Shapes, OUT and OUT UNION, and FETCH

In the previous chapters we have used cursor variables without fully discussing them. Most of the uses are fairly self-evident but a more exhaustive discussion is also useful.

First there are three types of cursors, as we will see below.

Statement Cursors

A statement cursor is based on a SQL SELECT statement. A full example might look like this:

-- elsewhere
create table xy_table(x integer, y integer);

declare C cursor for select x, y from xy_table;

When compiled, this will result in creating a SQLite statement object (type sqlite_stmt *) and storing it in a variable called C_stmt. This statement can then be used later in various ways.

Here’s perhaps the simplest way to use the cursor above:

declare x, y integer;
fetch C into x, y;

This will have the effect of reading one row from the results of the query into the local variables x and y.

These variables might then be used to create some output such as:

/* note use of double quotes so that \n is legal */
call printf("x:%d y:%d\n", ifnull(x, 0), ifnull(y,0));

More generally, there the cursor may or may not be holding fetched values. The cursor variable C can be used by itself as a boolean indicating the presence of a row. So a more complete example might be

if C then
  call printf("x:%d y:%d\n", ifnull(x, 0), ifnull(y,0));
else
  call printf("nada\n");
end if

And even more generally

loop fetch C into x, y
begin
  call printf("x:%d y:%d\n", ifnull(x, 0), ifnull(y,0));
end;

The last example above reads all the rows and prints them.

Now if the table xy_table had instead had dozens of columns, those declarations would be very verbose and error prone, and frankly annoying, especially if the table definition was changing over time.

To make this a little easier, there are so-called ‘automatic’ cursors. These happen implicitly and include all the necessary storage to exactly match the rows in their statement. Using the automatic syntax for the above looks like so:

declare C cursor for select * from xy_table;
fetch C;
if C then
  call printf("x:%d y:%d\n", ifnull(C.x, 0), ifnull(C.y,0));
end if;

or the equivalent loop form:

declare C cursor for select * from xy_table;
loop fetch C
begin
  call printf("x:%d y:%d\n", ifnull(C.x, 0), ifnull(C.y,0));
end;

All the necessary local state is automatically created, hence “automatic” cursor. This pattern is generally preferred, but the loose variables pattern is in some sense more general.

In all the cases if the number or type of variables do not match the select statement, semantic errors are produced.

Value Cursors

The purpose of value cursors is to make it possible for a stored procedure to work with structures as a unit rather than only field by field. SQL doesn’t have the notion of structure types, but structures actually appear pretty directly in many places. Generally we call these things “Shapes” and there are a variety of source for shapes including:

  • the columns of a table
  • the projection of a SELECT statement
  • the columns of a cursor
  • the result type of a procedure that returns a select
  • the arguments of a procedure
  • other things derived from the above

Let’s first start with how you declare a value cursor. It is providing one of the shape sources above.

So:

declare C cursor like xy_table;
declare C cursor like select 1 a, 'x' b;
declare C cursor like (a integer not null, b text not null);
declare C cursor like my_view;
declare C cursor like my_other_cursor;
declare C cursor like my_previously_declared_stored_proc;
declare C cursor like my_previously_declared_stored_proc arguments;

Any of those forms define a valid set of columns – a shape. Note that the select example in no way causes the query provided to run. Instead, the select statement is analyzed and the column names and types are computed. The cursor gets the same field names and types. Nothing happens at run time.

The last two examples assume that there is a stored procedure defined somewhere earlier in the same translation unit and that the procedure returns a result set or has arguments, respectively.

In all cases the cursor declaration makes a cursor that could hold the indicated result. That result can then be loaded with FETCH or emitted with OUT or OUT UNION which will be discussed below.

Once we have declared a value cursor we can load it with values using FETCH in its value form. Here are some examples:

Fetch from compatible values:

fetch C from values(1,2);

Fetch from a call to a procedure that returns a single row:

fetch C from call my_previously_declared_stored_proc();

Fetch from another cursor:

fetch C from D;

In this last case if D is a statement cursor it must also be “automatic” (i.e. it has the storage). This form lets you copy a row and save it for later. For instance, in a loop you could copy the current max-value row into a value cursor and use it after the loop, like so:

declare C cursor for select * from somewhere;
declare D cursor like C;

loop fetch C
begin
  if (not D or D.something < C.something) then
    fetch D from C;
  end if;
end;

After the loop, D either empty because there were no rows (thus if D would fail) or else it has the row with the maximum value of something, whatever that is.

Value cursors are always have their own storage, so you could say all value cursors are “automatic”.

And as we saw above, value cursors may or may not be holding a row.

declare C cursor like xy_table;
if not C then
  call printf("this will always print because C starts empty\n");
end if;

When you call a procedure you may or may not get a row as we’ll see below.

The third type of cursor is a “result set” cursor but that won’t make any sense until we’ve discussed result sets a little which requires OUT and/or OUT UNION and so we’ll go on to those statements next. As it happens, we are recapitulating the history of cursor features in the CQL language by exploring the system in this way.

Benefits of using named typed to declare a cursor

This form allows any kind of declaration, for instance:

declare C cursor like ( id integer not null, val real, flag boolean );

This wouldn’t really give us much more than the other forms, however typed name lists can include LIKE in them again, as part of the list. Which means you can do this kind of thing:

declare C cursor like (like D, extra1 real, extra2 bool)

You could then load that cursor like so:

fetch C from values (from D, 2.5, false);

and now you have D plus 2 more fields which maybe you want to output.

Importantly this way of doing it means that C always includes D, even if D changes over time. As long as the extra1 and extra2 fields don’t conflict names it will always work.

OUT Statement

Value cursors were initially designed to create a convenient way for a procedure to return a single row from a complex query without having a crazy number of OUT parameters. It’s easiest to illustrate this with an example.

Suppose you want to return several variables, the “classic” way to do so would be a procedure like this:

create proc get_a_row(
  id_ integer not null,
  out got_row bool not null,
  out w integer not null,
  out x integer,
  out y text not null,
  out z real)
begin
  declare C cursor for
    select w, x, y, z from somewhere where id = id_;
  fetch C into w, x, y, z;
  set got_row := C;
end;

This is already verbose, but you can imagine the situation gets very annoying if get_a_row has to produce a couple dozen column values. And of course you have to get the types exactly right. And they might evolve over time. Joy.

On the receiving side you get to do something just as annoying:

declare w integer not null
declare x integer;
declare y text;
declare z real;
declare got_row bool not null;
call get_a_row(id, got_row, w, x, y, z);

Using the out statement we get the equivalent functionality with a much simplified pattern. It looks like this:

create proc get_a_row(id_ integer not null)
begin
  declare C cursor for
    select w, x, y, z from somewhere where id = id_;
  fetch C;
  out C;
end;

To use the new procedure you simply do this:

declare C cursor like get_a_row;
fetch C from call get_a_row(id);

In fact, originally you did the two steps above in one statement and that was the only way to load a value cursor. Later, the calculus was generalized. The original form still works:

declare C cursor fetch from call get_a_row(id);

The OUT statement lets you return a single row economically and lets you then test if there actually was a row and if so, read the columns. It infers all the various column names and types so it is resilient to schema change and generally a lot less error prone than having a large number of out arguments to your procedure.

Once you have the result in a value cursor you can do the usual cursor operations to move it around or otherwise work with it.

The use of the LIKE keyword to refer to groups of columns spread to other places in CQL as a very useful construct, but it began here with the need to describe a cursor shape economically, by reference.

OUT UNION Statement

The semantics of the OUT statement are that it always produces one row of output (a procedure can produce no row if an out never actually rans but the procedure does use OUT).

If an OUT statement runs more than once, the most recent row becomes the result. So the OUT statement really does mirror having one out variable for each output column. This was its intent and procedures that return at most, or exactly, one row are very common so it works well enough.

However, in general, one row results do not suffice; you might want to produce a result set from various sources, possibly with some computation as part of the row creation process. To make general results, you need to be able to emit multiple rows from a computed source. This is exactly what OUT UNION provides.

Here’s a (somewhat contrived) example of the kind of thing you can do with this form:

create proc foo(n integer not null)
begin
  declare C cursor like select 1 value;
  let i := 0;
  while i < n
  begin
     -- emit one row for every integer
     fetch C from values(i);
     out union C;
     set i := i + 1;
  end;
end;

In foo above, we make an entire result set out of thin air. It isn’t a very interesting result, but of course any computation would have been possible.

This pattern is very flexible as we see below in bar where we merge two different data streams.

create table t1(id integer, stuff text, [other things too]);
create table t2(id integer, stuff text, [other things too]);

create proc bar()
begin
  declare C cursor for select * from t1 order by id;
  declare D cursor for select * from t2 order by id;

  fetch C;
  fetch D;

  -- we're going to merge these two queries
  while C or D
  begin
    -- if both have a row pick the smaller id
    if C and D then
       if C.id < D.id then
         out union C;
         fetch C;
       else
         out union D;
         fetch D;
       end if;
    else if C then
      -- only C has a row, emit that
      out union C;
      fetch C;
    else
      -- only D has a row, emit that
      out union D;
      fetch D;
    end if;
  end;
end;

Just like foo, in bar, each time OUT UNION runs a new row is accumulated.

Now, if you build a procedure that ends with a SELECT statement CQL automatically creates a fetcher function that does something like an OUT UNION loop – it loops over the SQLite statement for the SELECT and fetches each row, materializing a result.

With OUT UNION you take manual control of this process, allowing you to build arbitrary result sets. Note that either of C or D above could have been modified, replaced, skipped, normalized, etc. with any kind of computation. Even entirely synthetic rows can be computed and inserted into the output as we saw in foo.

Result Set Cursors

Now that we have OUT UNION it makes sense to talk about the final type of cursor.

OUT UNION makes it possible to create arbitrary result sets using a mix of sources and filtering. Unfortunately this result type is not a simple row, nor is it a SQLite statement. This meant that neither of the existing types of cursors could hold the result of a procedure that used OUT UNION. – CQL could not itself consume its own results.

To address this hole, we need an additional cursor type. The syntax is exactly the same as the statement cursor cases described above but, instead of holding a SQLite statement, the cursor holds a result set pointer and the current and maximum row numbers. Stepping through the cursor simply increments the row number and fetches the next row out of the rowset instead of from SQLite.

Example:

-- reading the above
create proc reader()
begin
  declare C cursor for call bar();
  loop fetch C
  begin
    call printf("%d %s\n", C.id, C.stuff);  -- or whatever fields you need
  end;
end;

If bar had been created with a SELECT, UNION ALL, and ORDER BY to merge the results, the above would have worked with C being a standard statement cursor, iterating over the union. Since foo produces a result set, CQL transparently produces a suitable cursor implementation behind the scenes, but otherwise the usage is the same.

Note this is a lousy way to simply iterate over rows; you have to materialize the entire result set so that you can just step over it. Re-consuming like this is not recommended at all for production code, but it is ideal for testing result sets that were made with OUT UNION which otherwise would require C/C++ to test. Testing CQL with CQL is generally a lot easier.

Reshaping Data, Cursor LIKE forms

There are lots of cases where you have big rows with many columns, and there are various manipulations you might need to do.

What follows is a set of useful syntactic sugar constructs that simplify handling complex rows. The idea is that pretty much anywhere you can specify a list of columns you can instead use the LIKE x construct to get the columns as they appear in the shape x – which is usually a table or a cursor.

It’s a lot easier to illustrate with examples, even though these are, again, a bit contrived.

First we need some table with lots of columns – usually the column names are much bigger which makes it all the more important to not have to type them over and over, but in the interest of some brevity, here is a big table:

create table big (
  id integer primary key,
  id2 integer unique,
  a integer,
  b integer,
  c integer,
  d integer,
  e integer,
  f integer
);

This example showcases several of the cursor and shape slicing features by emitting two related rows:

create proc foo(id_ integer not null)
begin
  -- this is the shape of the result we want -- it's some of the columns of "big"
  -- note this query doesn't run, we just use its shape to create a cursor
  -- with those columns.
  declare result cursor like select id, b, c, d from big;

  -- fetch the main row, specified by id_
  -- main row has all the fields, including id2
  declare main_row cursor for select * from big where id = id_;
  fetch main_row;

  -- now fetch the result columns out of the main row
  -- `like result` here means to use the names of the result cursor
  -- to index into the columns of the main_row cursor, and then
  -- and store them in `result`
  fetch result from cursor main_row(like result);

  -- this is our first result row
  out union result;

  -- now we want the related row, but we only need two columns
  -- from the related row, 'b' and 'c'
  declare alt_row cursor for select b, c from big where big.id2 = main_row.id2;
  fetch alt_row;

  -- update some of the fields in 'result' from the `alt_row`
  update cursor result(like alt_row) from cursor alt_row;

  -- and emit the modified result, so we've generated two rows
  out union result;
end;

Now let’s briefly discuss what is above. The two essential parts are:

fetch result from cursor main_row(like result);

and

update cursor result(like alt_row) from cursor alt_row;

In the first case what we’re saying is that we want to load the columns of result from main_row but we only want to take the columns that are actually present in result. So this is a narrowing of a wide row into a smaller row. In this case, the smaller row, result, is what we want to emit. We needed the other columns to compute alt_row.

The second case, what we’re saying is that we want to update result by replacing the columns found in alt_row with the values in alt_row. So in this case we’re writing a smaller cursor into part of a wider cursor. Note that we used the update cursor form here because it preserves all other columns. If we used fetch we would be rewriting the entire row contents, using NULL if necessary, and that is not desired here.

Here is the rewritten version of the above procedure; this is what ultimately gets compiled into C.

CREATE PROC foo (id_ INTEGER NOT NULL)
BEGIN
  DECLARE result CURSOR LIKE SELECT id, b, c, d FROM big;
  DECLARE main_row CURSOR FOR SELECT * FROM big WHERE id = id_;
  FETCH main_row;

  FETCH result(id, b, c, d)
    FROM VALUES(main_row.id, main_row.b, main_row.c, main_row.d);
  OUT UNION result;

  DECLARE alt_row CURSOR FOR SELECT b, c FROM big WHERE big.id2 = main_row.id2;
  FETCH alt_row;

  UPDATE CURSOR result(b, c) FROM VALUES(alt_row.b, alt_row.c);
  OUT UNION result;
END;

Of course you could have typed the above directly but if there are 50 odd columns it gets old fast and is very error prone. The sugar form is going to be 100% correct and will require much less typing and maintenance.

Finally, while I’ve shown both LIKE forms separately, they can also be used together. For instance:

  update cursor C(like X) from cursor D(like X);

The above would mean, “move the columns that are found in X from cursor D to cursor C”, presuming X has columns common to both.

Fetch Statement Specifics

Many of the examples used the FETCH statement in a sort of demonstrative way that is hopefully self-evident but the statement has many forms and so it’s worth going over them specifically. Below we’ll use the letters C and D for the names of cursors. Usually C;

Fetch with Statement or Result Set Cursors

A cursor declared in one of these forms:

  • declare C cursor for select * from foo;
  • declare C cursor for call foo(); (foo might end with a select or use out union)

is either a statement cursor or a result set cursor. In either case the cursor moves through the results. You load the next row with:

  • FETCH C, or
  • FETCH C into x, y, z;

In the first form C is said to be automatic in that it automatically declares the storage needed to hold all its columns. As mentioned above, automatic cursors have storage for their row.

Having done this fetch you can use C as a scalar variable to see if it holds a row, e.g.

declare C cursor for select * from foo limit 1;
fetch C;
if C then
  -- bingo we have a row
  call printf("%s\n", C.whatever);
end if

You can easily iterate, e.g.

declare C cursor for select * from foo;
loop fetch C
begin
  -- one time for every row
  call printf("%s\n", C.whatever);
end;

Automatic cursors are so much easier to use than explicit storage that explicit storage is rarely seen. Storing to out parameters is one case where explicit storage actually is the right choice, as the out parameters have to be declared anyway.

Fetch with Value Cursors

A value cursor is declared in one of these ways:

  • declare C cursor fetch from call foo(args)
    • foo must be a procedure that returns one row with OUT
  • declare C cursor like select 1 id, "x" name;
  • declare C cursor like X;
    • where X is the name of a table, a view, another cursor, or a procedure that returns a structured result

A value cursor is always automatic; it’s purpose is to hold a row. It doesn’t iterate over anything but it can be re-loaded in a loop.

  • fetch C or fetch C into ... is not valid on such a cursor, because it doesn’t have a source to step through.

The canonical way to load such a cursor is:

  • fetch C from call foo(args);
    • foo must be a procedure that returns one row with OUT
  • fetch C(a,b,c...) from values(x, y, z);

The first form is in some sense the origin of the value cursor. Value cursors were added to the language initially to provide a way to capture the single row OUT statement results, much like result set cursors were added to capture procedure results from OUT UNION. In the first form, the cursor storage (a C struct) is provided by reference as a hidden out parameter to the procedure and the procedure fills it in. The procedure may or may not use the OUT statement in its control flow, as the cursor might not hold a row. You can use if C then ... as before to test for a row.

The second form is more interesting as it allows the cursor to be loaded from arbitrary expressions subject to some rules:

  • you should think of the cursor as a logical row: it’s either fully loaded or it’s not, therefore you must specify enough columns in the column list to ensure that all NOT NULL columns will get a value
  • if not mentioned in the list, NULL will be loaded where possible
  • if insufficient columns are named, an error is generated
  • if the value types specified are not compatible with the column types mentioned, an error is generated
  • later in this chapter, we’ll show that columns can also be filled with dummy data using a seed value

With this form, any possible valid cursor values could be set, but many forms of updates that are common would be awkward. So there are various forms of syntactic sugar that are automatically rewritten into the canonical form. See the examples below:

  • fetch C from values(x, y, z)

    • if no columns are specified this is the same as naming all the columns, in declared order
  • fetch C from arguments

    • the arguments to the procedure in which this statement appears are used as the values, in order
    • in this case C was also rewritten into C(a,b,c,..) etc.
  • fetch C from arguments like C

    • the arguments to the procedure in which this statement appears are used, by name, as the values, using the names of of the indicated shape
    • the order in which the arguments appeared no longer matters, the names that match the columns of C are used if present
    • the formal parameter name may have a single trailing underscore (this is what like C would generate)
    • e.g. if C has columns a and b then there must exist formals named a or a_ and b or b_, in any position
  • fetch C(a,b) from cursor D(a,b)

    • the named columns of D are used as the values
    • in this case the statement becomes: fetch C(a,b) from values(D.a, D.b);

That most recent form doesn’t seem like it saves much, but recall the first rewrite:

  • fetch C from cursor D
    • both cursors are expanded into all their columns, creating a copy from one to the other
    • fetch C from D can be used if the cursors have the exact same column names and types; it also generates slightly better code and is a common case

It is very normal to want to use only some of the columns of a cursor; these LIKE forms do that job. We saw some of these forms in an earlier example.

  • fetch C from cursor D(like C)
    • here D is presumed to be “bigger” than C, in that it has all of the C columns and maybe more. The like C expands into the names of the C columns so C is loaded from the C part of D
    • the expansion might be fetch C(a, b, g) from values (D.a, D.b, D.g)
    • D might have had fields c, d, e, f which were not used because they are not in C.

The symmetric operation, loading some of the columns of a wider cursor can be expressed neatly:

  • fetch C(like D) from cursor D
    • the like D expands into the columns of D causing the cursor to be loaded with what’s in D and NULL (if needed)
    • when expanded, this might look like fetch C(x, y) from values(D.x, D.y)

LIKE can be used in both places, for instance suppose E is a shape that has a subset of the rows of both C and D. You can write a form like this:

  • fetch C(like E) from cursor D(like E)
    • this means take the column names found in E and copy them from D to C.
    • the usual type checking is done

As is mentioned above, the fetch form means “load an entire row into the cursor”. This is important because “half loaded” cursors would be semantically problematic. However there are many cases where you might like to amend the values of an already loaded cursor. You can do this with the update form.

  • update cursor C(a,b,..) from values(1,2,..);
    • the update form is a no-op if the cursor is not already loaded with values (!!)
    • the columns and values are type checked so a valid row is ensured (or no row)
    • all the re-writes above are legal so update cursor C(like D) from D is possible; it is in fact the use-case for which this was designed.

Calling Procedures with Argument Bundles

It’s often desirable to treat bundles of arguments as a unit, or cursors as a unit, especially when calling other procedures. The shape patterns above are very helpful for moving data between cursors, and the database. These can be rounded out with similar constructs for procedure definitions and procedure calls as follows.

First we’ll define some shapes to use in the examples. Note that we made U using T.

create table T(x integer not null, y integer not null,  z integer not null);
create table U(like T, a integer not null, b integer not null);

We haven’t mentioned this before but the implication of the above is that you can use the LIKE construct inside a table definition to add columns from a shape.

We can also use the LIKE construct to create procedure arguments. To avoid conflicts with column names, when used this way the procedure arguments all get a trailing underscore appended to them. The arguments will be x_, y_, and z_ as we can see if the following:

create proc p1(like T)
begin
  call printf("%d %d %d\n", x_, y_, z_);
end;

Shapes can also be used in a procedure call, as showed below. This next example is obviously contrived, but of course it generalizes. It is exactly equivalent to the above.

create proc p2(like T)
begin
  call printf("%d %d %d\n", from arguments);
end;

Now we might want to chain these things together. This next example uses a cursor to call p1.

create proc q1()
begin
 declare C cursor for select * from T;
 loop fetch C
 begin
   /* this is the same as call p(C.x, C.y, C.z) */
   call p1(from C);
 end;
end;

The LIKE construct allows you to select some of the arguments, or some of a cursor to use as arguments. This next procedure has more arguments than just T. The arguments will be x_, y_, z_, a_, b_. But the call will still have the T arguments x_, y_, and z_.

create proc q2(like U)
begin
  /* just the args that match T: so this is still call p(x_, y_, z_) */
  call p1(from arguments like T);
end;

Or similarly, using a cursor.

create proc q3(like U)
begin
 declare C cursor for select * from U;
 loop fetch C
 begin
  /* just the columns that match T so this is still call p(C.x, C.y, C.z) */
  call p1(from C like T);
 end;
end;

Note that the from argument forms do not have to be all the arguments. For instance you can get columns from two cursors like so:

  call something(from C, from D)

All the varieties can be combined but of course the procedure signature must match. And all these forms work in function expressions as well as procedure calls.

e.g.

  set x := a_function(from C);

Since these forms are simply syntatic sugar, they can also appear inside of function calls that are in SQL statements. The variables mentioned will be expanded and become bound variables just like any other variable that appears in a SQL statement.

Note the form x IN (from arguments) is not supported at this time, though this would be a relatively easy addition.

Using Named Argument Bundles

There are many cases where stored procedures require complex arguments using data shapes that come from the schema, or from other procedures. As we have seen the LIKE construct for arguments can help with this, but it has some limitations. Let’s consider a specific example to study:

create table Person (
  id text primary key,
  name text not null,
  address text not null,
  birthday real
);

To manage this table we might need something like this:

create proc insert_person(like Person)
begin
  insert into Person from arguments;
end;

As we have seen, the above expands into:

create proc insert_person(
  id_ text not null,
  name_ text not null,
  address_ text not null,
  birthday_ real)
begin
  insert into Person(id, name, address, birthday)
    values(id_, name_, address_, birthday_);
end;

It’s clear that the sugared version is a lot easier to reason about than the fully expanded version, and much less prone to errors as well.

This much is already helpful, but just those forms aren’t general enough to handle the usual mix of situations. For instance, what if we need a procedure that works with two people? A hypothetical insert_two_people procedure cannot be written with the forms we have so far.

To generalize this the language adds the notion of named argument bundles. The idea here is to name the bundles which provides a useful scoping. Example:

create proc insert_two_people(p1 like Person, p2 like Person)
begin
  -- using a procedure that takes a Person args
  call insert_person(from p1);
  call insert_person(from p2);
end;

or alternatively

create proc insert_two_people(p1 like Person, p2 like Person)
begin
  -- inserting a Person directly
  insert into Person from p1;
  insert into Person from p2;
end;

The above expands into:

create proc insert_two_people(
  p1_id text not null,
  p1_name text not null,
  p1_address text not null,
  p1_birthday real,
  p2_id text not null,
  p2_name text not null,
  p2_address text not null,
  p2_birthday real)
begin
  insert into Person(id, name, address, birthday)
    values(p1_id, p1_name, p1_address, p1_birthday);
  insert into Person(id, name, address, birthday)
    values(p2_id, p2_name, p2_address, p2_birthday);
end;

Or course different named bundles can have different types – you can create and name shapes of your choice. The language allows you to use an argument bundle name in all the places that a cursor was previously a valid source. That includes insert, fetch, update cursor, and procedure calls. You can refer to the arguments by their expanded name such as p1_address or alternatively p1.address – they mean the same thing.

Here’s another example showing a silly but illustrative thing you could do:

create proc insert_lotsa_people(P like Person)
begin
  -- make a cursor to hold the arguments
  declare C cursor like P;

  -- convert arguments to a cursor
  fetch C from P;

  -- set up to patch the cursor and use it 20 times
  let i := 0;
  while i < 20
  begin
    update cursor C(id) from values(printf("id_%d", i));
    insert into Person from C;
    set i := i + 1;
  end;
end;

The above shows that you can use a bundle as the source of a shape, and you can use a bundle as a source of data to load a cursor. After which you can do all the usual value cursor things. Of course in this case the value cursor was redundant, we could just as easily have done something like this:

  set P_id := printf("id_%d", i);
  insert into Person from P;
  set i := i + 1;

NOTE: the CQL JSON output includes extra information about procedure arguments if they originated as part of a shape bundle do identify the shape source for tools that might need that information.

The @COLUMNS construct in the SELECT statement

The select list of a SELECT statement already has complex syntax and functionality, but it is a very interesting place to use shapes. To make it possible to use shape notations and not confuse that notation with standard SQL the @COLUMNS construct was added to the language. This allows for a sugared syntax for extracting columns in bulk.

The @COLUMNS clause is like of a generalization of the select T.* with shape slicing and type-checking. The forms are discussed below:

Columns from a join table or tables

This is the simplest form, it’s just like T.*:

-- same as A.*
select @columns(A) from ...;

-- same as A.*, B.*
select @columns(A, B) from ...;

Columns from a particular joined table that match a shape

This allows you to choose some of the columns of one table of the FROM clause.

-- the columns of A that match the shape Foo
select @columns(A like Foo) from ...;

-- get the Foo shape from A and the Bar shape from B
select @columns(A like Foo, B like Bar) from ...;

Columns from any that match a shape, from anywhere in the FROM

Here we do not specify a particular table that contains the columns, the could come from any of the tables in the FROM clause.

--- get the Foo shape from anywhere in the join
select @columns(like Foo) from ...;

-- get the Foo and Bar shapes, from anywhere in the join
select @columns(like Foo, like Bar) from ...;

Subsets of Columns from shapes

This pattern can be helpful for getting part of a shape.

-- get the a and b from the Foo shape only
select @columns(like Foo(a,b))

This pattern is great for getting almost all of a shape (e.g. everything but the pk).

-- get the Foo shape except the a and b columns
select @columns(like Foo(-a, -b))

Specific columns

This form allows you to slice out a few columns without defining a shape, you simply list the exact columns you want.

-- T1.x and T2.y plus the Foo shape
select @columns(T1.x, T2.y, like Foo) from ...;

Distinct columns

Its often the case that there are duplicate column names in the FROM clause. For instance, you could join A to B with both having a column pk. The final result set can only have one column named pk, the distinct clause helps you to get distinct column names. In this context distinct is about column names, not values.

-- removes duplicate column names
-- e.g. there will be one copy of 'pk'
select @columns(distinct A, B) from A join B using(pk);

-- if both Foo and Bar have an (e.g.) 'id' field you only get one copy
select @columns(distinct like Foo, like Bar) from ...;

If a specific column is mentioned it is always included, but later expressions that are not a specific column will avoid that column name.

-- if F or B has an x it won't appear again, just T.x
select @columns(distinct T.x, F like Foo, B like Bar) from F, B ..;

Of course this is all just sugar, so it all compiles to a column list with table qualifications – but the syntax is very powerful. You can easily narrow a wide table, or fuse joins that share common keys without creating conflicts.

-- just the Foo columns
select @columns(like Foo) from Superset_Of_Foo_From_Many_Joins_Even;

-- only one copy of 'pk'
select @columns(distinct A,B,C) from
  A join B using (pk) join C using (pk);

And of course you can define shapes however you like and then use them to slice off column chucks of your choice. There are many ways to build up shapes from other shapes. For instance, you can declare procedures that return the shape you want and never actually create the procedure – a pattern is very much like a shape “typedef”. E.g.

declare proc shape1() (x integer, y real, z text);
declare proc shape2() (like shape1, u bool, v bool);

With this combination you can easily define common column shapes and slice them out of complex queries without having to type the columns names over and over.

Missing Data Columns, Nulls and Dummy Data

What follows are the rules for columns that are missing in an INSERT, or FETCH statement. As with many of the other things discussed here, the forms result in automatic rewriting of the code to include the specified dummy data. So SQLite will never see these forms.

Two things to note: First, the dummy data options described below are really only interesting in test code, it’s hard to imagine them being useful in production code. Second, none of what follows applies to the update cursor statement because its purpose is to do partial updates on exactly the specified columns and we’re about to talk about what happens with the columns that were not specified.

When fetching a row all the columns must come from somewhere; if the column is mentioned or mentioned by rewrite then it must have a value mentioned, or a value mentioned by rewrite. For columns that are not mentioned, a NULL value is used if it is legal to do so. For example, fetch C(a) from values(1) might turn into fetch C(a,b,c,d) from values (1, NULL, NULL, NULL)

In addition to the automatic NULL you may add the annotation @dummy_seed([long integer expression]). If this annotation is present then:

  • the expression is evaluated and stored in the hidden variable seed
  • all integers, and long integers get seed as their value (possibly truncated)
  • booleans get 1 if and only if seed is non-zero
  • strings get the name of the string column an underscore and the value as text (e.g. “myText_7” if seed is 7)
  • blobs are not currently supported for dummy data (CQL is missing blob conversions which are needed first)

This construct is hugely powerful in a loop to create many complete rows with very little effort, even if the schema change over time.

declare i integer not null;
declare C like my_table;
set i := 0;
while (i < 20)
begin
   fetch C(id) from values(i+10000) @dummy_seed(i);
   insert into my_table from cursor C;
end;

Now in this example we don’t need to know anything about my_table other than that it has a column named id. This example shows several things:

  • we got the shape of the cursor from the table we were inserting into
  • you can do your own computation for some of the columns (those named) and leave the unnamed values to be defaulted
  • the rewrites mentioned above work for the insert statement as well as fetch
  • in fact insert into my_table(id) values(i+10000) @dummy_seed(i) would have worked too with no cursor at all
    • bonus, dummy blob data does work in insert statements because SQLite can do the string conversion easily
    • the dummy value for a blob is a blob that holds the text of the column name and the text of the seed just like a string column

The @dummy_seed form can be modified with @dummy_nullables, this indicates that rather than using NULL for any nullable value that is missing, CQL should use the seed value. This overrides the default behavior of using NULL where columns are needed. Note the NULL filling works a little differently on insert statements. Since SQLite will provide a NULL if one is legal the column doesn’t have to be added to the list with a NULL value during rewriting, it can simply be omitted, making the statement smaller.

Finally for insert statement only, SQLite will normally use the default value of a column if it has one, so there is no need to add missing columns with default values to the insert statement. However if you specify @dummy_defaults then columns with a default value will instead be rewritten and they will get _seed_ as their value.

Some examples. Suppose columns a, b, c are not null; m, n are nullable; and x, y have defaults.

-- as written
insert into my_table(a) values(7) @dummy_seed(1000)
-- rewrites to
insert into my_table(a, b, c) values(7, 1000, 1000);
-- as written
insert into my_table(a) values(7) @dummy_seed(1000) @dummy_nullables
-- rewrites to
insert into my_table(a, b, c, m, n) values(7, 1000, 1000, 1000, 1000);
-- as written
insert into my_table(a) values(7) @dummy_seed(1000) @dummy_nullables @dummy_defaults
-- rewrites to
insert into my_table(a, b, c, m, n, x, y) values(7, 1000, 1000, 1000, 1000, 1000, 1000);

The sugar features on fetch, insert, and update cursor are as symmetric as possible, but again, dummy data is generally only interesting in test code. Dummy data will continue to give you valid test rows even if columns are added or removed from the tables in question.

Generalized Cursor Lifetimes aka Cursor “Boxing”

Generalized Cursor Lifetime refers to capturing a Statement Cursor in an object so that it can used more flexibly. Wrapping something in an object is often called “boxing”. Since Generalized Cursor Lifetime is a mouthful we’ll refer to it as “boxing” from here forward. The symmetric operation “unboxing” refers to converting the boxed object back into a cursor.

The normal cursor usage pattern is by far the most common, a cursor is created directly with something like these forms:

declare C cursor for select * from shape_source;

declare D cursor for call proc_that_returns_a_shape();

At this point the cursor can be used normally as follows:

loop fetch C
begin
  -- do stuff with C
end;

Those are the usual patterns and they allow statement cursors to be consumed sort of “up” the call chain from where the cursor was created. But what if you want some worker procedures that consume a cursor? There is no way to pass your cursor down again with these normal patterns alone.

To generalize the patterns, allowing, for instance, a cursor to be returned as an out parameter or accepted as an in parameter you first need to declare an object variable that can hold the cursor and has a type indicating the shape of the cursor.

To make an object that can hold a cursor:

declare obj object<T cursor>;

Where T is the name of a shape. It can be a table name, or a view name, or it can be the name of the canonical procedure that returns the result. T should be some kind of global name, something that could be accessed with #include in various places. Referring to the examples above, choices for T might be shape_source the table or proc_that_returns_a_shape the procedure.

NOTE: it’s always possible make a fake procedure that returns a result to sort of “typedef” a shape name. e.g.

declare proc my_shape() (id integer not null, name text);

The procedure here my_shape doesn’t have to actually ever be created, in fact it’s better if it isn’t. It won’t ever be called; its hypothetical result is just being as a shape. This can be useful if you have several procedures like proc_that_returns_a_shape that all return results with the columns of my_shape.

To create the boxed cursor, first declare the object variable that will hold it and then set object from the cursor. Note that in the following example the cursor C must have the shape defined by my_shape or an error is produced. The type of the object is crucial because, as we’ll see, during unboxing that type defines the shape of the unboxed cursor.

-- recap: declare the box that holds the cursor (T changed to my_shape for this example)
declare box_obj object<my_shape cursor>;

-- box the cursor into the object (the cursor shape must match the box shape)
set box_obj from cursor C;

The variable box_obj can now be passed around as usual. It could be stored in a suitable out variable or it could be passed to a procedure as an in parameter. Then, later, you can “unbox” box_obj to get a cursor back. Like so:

-- unboxing a cursor from an object, the type of box_obj defines the type of the created cursor
declare D cursor for box_obj;

These primitives will allow cursors to be passed around with general purpose lifetime.

Example:

-- consumes a cursor
create proc cursor_user(box_obj object<my_shape cursor>)
begin
   declare C cursor for box_obj;  -- the cursors shape will be my_shape matching box
   loop fetch C
   begin
      -- do something with C
   end;
end;

-- captures a cursor and passes it on
create proc cursor_boxer()
begin
   declare C cursor for select * from something_like_my_shape;
   declare box_obj object<my_shape cursor>
   set box from cursor C; -- produces error if shape doesn't match
   call cursor_user(box_obj);
end;

Importantly, once you box a cursor the underlying SQLite statement’s lifetime is managed by the box object with normal retain/release semantics. The box and underlying statement can be released simply by setting all references to it to null as usual.

With this pattern it’s possible to, for instance, create a cursor, box it, consume some of the rows in one procedure, do some other stuff, and then consume the rest of the rows in another different procedure.

Important Notes:

  • the underlying SQLite statement is shared by all references to it. Unboxing does not reset the cursor’s position. It is possible, even desirable, to have different procedures advancing the same cursor
  • there is no operation for “peeking” at a cursor without advancing it; if your code requires that you inspect the row and then delegate it, you can do this simply by passing the cursor data as a value rather than the cursor statement. Boxing and unboxing are for cases where you need to stream data out of the cursor in helper procedures
  • durably storing a boxed cursor (e.g. in a global) could lead to all manner of problems – it SQLite terms is exactly like holding on to a sqlite3_stmt * for a long time with all the same problems, because that is exactly is happening

Summarizing, the main reason for using the boxing patterns is to allow for standard helper procedures that can get a cursor from a variety of places and process it. Boxing isn’t the usual pattern at all and returning cursors in a box, while possible, should be avoided in favor of the simpler patterns, if only because then then lifetime management is very simple in all those cases.