Metakit Extension for Jim Tcl


The mk extension provides an interface to the Metakit small-footprint embeddable database library ( The underlying library is efficient at manipulating not-so-large amounts of data and takes a different approach to composing database operations than common SQL-based relational databases.

Both the Metakit core library and the mk package can be linked either statically or dynamically and loaded using

package require mk


A database (called a “storage” in Metakit terms) may either reside totally in memory or be backed by a file. To open or create a database, call the storage command with an optional filename parameter:

set db [storage]

The returned handle can be used as a command name to access the database. When you are done, execute the close method, that is, run

$db close

A lost handle won’t be found by GC but will be closed when the interpreter exits. Note that by default Metakit will only record changes to the database when you close the handle. Use the commit method to record the current state of the database to disk.


Views in Metakit are what is called “tables” in conventional databases. A view may several typed properties, or columns, and contains homogenous rows, or records. New properties may be added to a view as needed; however, new properties are not stored in the database file by default. The structure method specifies the stored properties of a view, creating a new view or restructuring an old one as needed:

$db structure viewName description

The view description must be a list of form {propName type propName type ...}. The supported property types include:

A NULL-terminated string, stored as an array of bytes (without any encoding assumptions).
Not yet supported by the mk extension. Blob of binary data that may contain embedded NULLs (zero bytes). Stored as-is. This is more efficient than string when storing large blocks of data (e.g. images) and will adjust the storage strategy as needed.
A signed integer value occupying a maximum of 32 bits. If all values stored in a column can fit in a smaller range (16, 8, or even 4 or 2 bits), they are packed automatically.
Like integer, but is required to fit into 64 bits.
float and double
32-bit and 64-bit IEEE floating-point values respectively.
This type is not usually specified directly; instead, a structure description of a nested view is given. subview properties store complete views as their value, creating hierarchical data structures. When retrieved from a view, a value of a subview property is a normal view handle.

Without a description parameter, the structure method returns the current structure of the named view; without any parameters, it returns a dictionary containing structure descriptions of all views stored in the database.

After specifying the properties you expect to see in the view, call

[$db view $viewName] as viewHandle

to obtain a view handle. These handles are also commands, but are garbage-collected and also destroy themselves after a single method call; the as viewHandle call assigns the view handle to the specified variable and also tells the view not to destroy itself until all the references to it are gone.

View handles may also be made permanent by giving them a global command name, e.g.

rename [$db view data]

However, such view handles are not managed automatically at all and must be destroyed using the destroy method, or by renaming them to "".


The value of a particular property is obtained using

cursor get $cur propName

where $cur is a string of form viewHandle!index. Row indices are zero-based and may also be specified relative to the last row of the view using the end[+-]integer notation.

A dictionary containing all property name and value pairs can be retrieved by omitting the propName argument:

cursor get $cur

Setting property values is also performed either individually, using

cursor set $cur propName value ?propName value ...?

or via a dictionary with

cursor set $cur dictValue

In the first form of the command, property names may also be preceded by a -typeName option. In this case, a new property of the specified type will be created if it doesn’t already exist; note that this will cause all the rows in the view to have the property (but see A NOTE ON NULL below).

If the row index points after the end of the view, an appropriate number of fresh rows will be inserted first. So, for example, you can use end+1 to append a new row. (Note that you then have to set it all at once, though.)

The total number of rows can be obtained using

$viewHandle size

and set manually with

$viewHandle resize newSize

For example, you can use $viewHandle resize 0 to clear a view.


New rows may also be inserted at an arbitrary position in a view with

cursor insert $cur ?count?

This will insert count fresh rows into the view so that $cur points to the first one. The inverse of this operation is

cursor remove $cur ?count?


The real power of Metakit lies in the way existing views are combined to create new ones to obtain a particular perspective on the stored data. A single operation takes one or more views and possibly additional options and produces a new view, usually tracking notifications to the underlying views and sometimes even supporting modification.

Binary operations are left-biased when there are conflicting property values; that is, they always prefer the values from the left view.

Unary operations

view unique
Derived view with duplicate rows removed.
view sort crit ?crit …?
Derived view sorted on the specified criteria, in order. A single crit is either a property name or a property name preceded by a dash; the latter specifies that the sorting is to be performed in reverse order.

Binary operations

The operations taking set arguments require that the given views have no duplicate rows. The unique method can be used to ensure this.

view1 concat view2
Vertical concatenation; that is, all the rows of view1 and then all rows of view2.
view1 pair view2
Pairing, or horizontal concatenation: every row in view1 is matched with a row with the same index in view2; the result has all the properties of view1 and all the properties of view2.
view1 product view2
Cartesian product: each row in view1 horizontally concatenated with every row in view2.
set1 union set2
Set union. Unlike concat, this operation removes duplicates from the result. A row is in the result if it is in set1 or in set2.
set1 intersect set2
Set intersection. A row is in the result if it is in set1 and in set2.
set1 different set2
Symmetric difference. A row is in the result if it is in set1 xor in set2, that is, in set1 or in set2, but not in both.
set1 minus set2
Set minus. A row is in the result if it is in set1 and not in set2.

Relational operations

view1 join view2 ?-outer? prop ?prop …?
Relational join on the specified properties: the rows from view1 and view2 with all the specified properties equal are concatenated to form a new row. If the -outer option is specified, the rows from view1 that do not have a corresponding one in view2 are also left in the view, with the properties existing only in view2 filled with default values.
view group subviewName prop ?prop …?
Groups the rows with all the specified properties equal; moves all the remaining properties into a newly created subview property called subviewName.
view flatten subviewProp
The inverse of group.

Projections and selections

view project prop ?prop …?
Projection: a derived view with only the specified properties left.
view without prop ?prop …?
The opposite of project: a derived view with the specified properties removed.

view range start end ?step? A slice or a segment of view: rows at start, start+step, and so on, until the row number becomes larger than end. The usual end[+-]integer notation is supported, but the indices don’t change if the underlying view is resized.

(!) select etc. should go here

Search and storage optimization

view blocked
Invokes an optimization designed for storing large amounts of data. view must have a single subview property called _B with the desired structure inside. This additional level of indirection is used by blocked to create a view that looks like a usual one, but can store much more data efficiently. As a result, indexing into the view becomes a bit slower. Once this method is invoked, all access to view must go through the returned view.
view ordered prop ?prop …?
Does not transform the structure of the view in any way, but signals that the view should be considered ordered on a unique key consisting of the specified properties, enabling some optimizations. Note that duplicate keys are not allowed in an ordered view.

(!) TODO: hash, indexed(?) – these make no sense until searches are implemented


Because constructs like [[view op1 ...] op2 ...] op3 ... tend to be common in programs using Metakit, a shorthand syntax is introduced: such expressions may also be written as view op1 ... | op2 ... | op3 ....

Note though that this syntax is not in any way magically wired into the interpreter: it is understood only by the view handles and the two commands that can possibly return a view: $db view and cursor get. If you want to support this syntax in Tcl procedures, you’ll need to do this yourself, or you may want to create a custom view method and have the view handle work out the syntax for you (see USER-DEFINED METHODS below).


view copy
Creates a copy of view with the same data.
view clone
Creates a view with the same structure, but no data.
view pin
Specifies that the view should not be destroyed after a single method call. Returns view.
view as varName
In addition to the actions performed by pin, assigns the view handle to the variable named varName in the caller’s scope.
view properties
Returns the names of all properties in the view.
view type prop
Returns the type of the specified property.


Note that Metakit does not have a special NULL value like conventional relational databases do. Instead, it defines default property values: "" for string and binary types, 0 for all numeric types and a view with no rows for subviews. These defaults are used when a fresh row is inserted and when a new property is added to the view to fill in the missing values.


The storage and view handles support custom methods defined in Tcl: to define methodName on every storage or view handle, create a procedure called { methodName} or {mk.view methodName} respectively. These procedures will receive the handle as the first argument and all the remaining arguments. Remember to pin the view handle in view methods if you call more than one method of it!

Custom cursor subcommands may also be defined by creating a procedure called {cursor methodName}. These receive all the arguments without any modifications.