This module provides a query interface to Mnesia, ETS, Dets, and other data structures that provide an iterator style traversal of objects.

Overview

This module provides a query interface to QLC tables. Typical QLC tables are Mnesia, ETS, and Dets tables. Support is also provided for user-defined tables, see section Implementing a QLC Table. A query is expressed using Query List Comprehensions (QLCs). The answers to a query are determined by data in QLC tables that fulfill the constraints expressed by the QLCs of the query. QLCs are similar to ordinary list comprehensions as described in Erlang Reference Manual and Programming Examples, except that variables introduced in patterns cannot be used in list expressions. In the absence of optimizations and options such as cache and unique (see section Common Options, every QLC free of QLC tables evaluates to the same list of answers as the identical ordinary list comprehension.

While ordinary list comprehensions evaluate to lists, calling q/1,2 returns a query handle. To obtain all the answers to a query, eval/1,2 is to be called with the query handle as first argument. Query handles are essentially functional objects (funs) created in the module calling q/1,2. As the funs refer to the module code, be careful not to keep query handles too long if the module code is to be replaced. Code replacement is described in section Compilation and Code Loading in the Erlang Reference Manual. The list of answers can also be traversed in chunks by use of a query cursor. Query cursors are created by calling cursor/1,2 with a query handle as first argument. Query cursors are essentially Erlang processes. One answer at a time is sent from the query cursor process to the process that created the cursor.

Syntax

Syntactically QLCs have the same parts as ordinary list comprehensions:

[Expression || Qualifier1, Qualifier2, ...]

Expression (the template) is any Erlang expression. Qualifiers are either filters or generators. Filters are Erlang expressions returning boolean/0. Generators have the form Pattern <- ListExpression, where ListExpression is an expression evaluating to a query handle or a list. Query handles are returned from append/1,2, keysort/2,3, q/1,2, sort/1,2, string_to_handle/1,2,3, and table/2.

Evaluation

A query handle is evaluated in the following order:

• Inspection of options and the collection of information about tables. As a result, qualifiers are modified during the optimization phase.
• All list expressions are evaluated. If a cursor has been created, evaluation takes place in the cursor process. For list expressions that are QLCs, the list expressions of the generators of the QLCs are evaluated as well. Be careful if list expressions have side effects, as list expressions are evaluated in unspecified order.
• The answers are found by evaluating the qualifiers from left to right, backtracking when some filter returns false, or collecting the template when all filters return true.

Filters that do not return boolean/0 but fail are handled differently depending on their syntax: if the filter is a guard, it returns false, otherwise the query evaluation fails. This behavior makes it possible for the qlc module to do some optimizations without affecting the meaning of a query. For example, when testing some position of a table and one or more constants for equality, only the objects with equal values are candidates for further evaluation. The other objects are guaranteed to make the filter return false, but never fail. The (small) set of candidate objects can often be found by looking up some key values of the table or by traversing the table using a match specification. It is necessary to place the guard filters immediately after the table generator, otherwise the candidate objects are not restricted to a small set. The reason is that objects that could make the query evaluation fail must not be excluded by looking up a key or running a match specification.

Join

The qlc module supports fast join of two query handles. Fast join is possible if some position P1 of one query handler and some position P2 of another query handler are tested for equality. Two fast join methods are provided:

• Lookup join traverses all objects of one query handle and finds objects of the other handle (a QLC table) such that the values at P1 and P2 match or compare equal. The qlc module does not create any indexes but looks up values using the key position and the indexed positions of the QLC table.
• Merge join sorts the objects of each query handle if necessary and filters out objects where the values at P1 and P2 do not compare equal. If many objects with the same value of P2 exist, a temporary file is used for the equivalence classes.

The qlc module warns at compile time if a QLC combines query handles in such a way that more than one join is possible. That is, no query planner is provided that can select a good order between possible join operations. It is up to the user to order the joins by introducing query handles.

The join is to be expressed as a guard filter. The filter must be placed immediately after the two joined generators, possibly after guard filters that use variables from no other generators but the two joined generators. The qlc module inspects the operands of =:=/2, ==/2, is_record/2, element/2, and logical operators (and/2, or/2, andalso/2, orelse/2, xor/2) when determining which joins to consider.

Common Options

The following options are accepted by cursor/2, eval/2, fold/4, and info/2:

• {cache_all, Cache}, where Cache is equal to ets or list adds a {cache, Cache} option to every list expression of the query except tables and lists. Defaults to {cache_all, no}. Option cache_all is equivalent to {cache_all, ets}.
• {max_list_size, MaxListSize}, where MaxListSize is the size in bytes of terms on the external format. If the accumulated size of collected objects exceeds MaxListSize, the objects are written onto a temporary file. This option is used by option {cache, list} and by the merge join method. Defaults to 512*1024 bytes.
• {tmpdir_usage, TmpFileUsage} determines the action taken when qlc is about to create temporary files on the directory set by option tmpdir. If the value is not_allowed, an error tuple is returned, otherwise temporary files are created as needed. Default is allowed, which means that no further action is taken. The values info_msg, warning_msg, and error_msg mean that the function with the corresponding name in module error_logger is called for printing some information (currently the stacktrace).
• {tmpdir, TempDirectory} sets the directory used by merge join for temporary files and by option {cache, list}. The option also overrides option tmpdir of keysort/3 and sort/2. Defaults to "", which means that the directory returned by file:get_cwd() is used.
• {unique_all, true} adds a {unique, true} option to every list expression of the query. Defaults to {unique_all, false}. Option unique_all is equivalent to {unique_all, true}.

Getting Started

As mentioned earlier, queries are expressed in the list comprehension syntax as described in section Expressions in Erlang Reference Manual. In the following, some familiarity with list comprehensions is assumed. The examples in section List Comprehensions in Programming Examples can get you started. Notice that list comprehensions do not add any computational power to the language; anything that can be done with list comprehensions can also be done without them. But they add syntax for expressing simple search problems, which is compact and clear once you get used to it.

Many list comprehension expressions can be evaluated by the qlc module. Exceptions are expressions, such that variables introduced in patterns (or filters) are used in some generator later in the list comprehension. As an example, consider an implementation of lists:append(L): [X ||Y <- L, X <- Y]. Y is introduced in the first generator and used in the second. The ordinary list comprehension is normally to be preferred when there is a choice as to which to use. One difference is that eval/1,2 collects answers in a list that is finally reversed, while list comprehensions collect answers on the stack that is finally unwound.

What the qlc module primarily adds to list comprehensions is that data can be read from QLC tables in small chunks. A QLC table is created by calling qlc:table/2. Usually qlc:table/2 is not called directly from the query but through an interface function of some data structure. Erlang/OTP includes a few examples of such functions: mnesia:table/1,2, ets:table/1,2, and dets:table/1,2. For a given data structure, many functions can create QLC tables, but common for these functions is that they return a query handle created by qlc:table/2. Using the QLC tables provided by Erlang/OTP is usually probably sufficient, but for the more advanced user section Implementing a QLC Table describes the implementation of a function calling qlc:table/2.

Besides qlc:table/2, other functions return query handles. They are used more seldom than tables, but are sometimes useful. qlc:append/1,2 traverses objects from many tables or lists after each other. If, for example, you want to traverse all answers to a query QH and then finish off by a term {finished}, you can do that by calling qlc:append(QH, [{finished}]). append/2 first returns all objects of QH, then {finished}. If a tuple {finished} exists among the answers to QH, it is returned twice from append/2.

As another example, consider concatenating the answers to two queries QH1 and QH2 while removing all duplicates. This is accomplished by using option unique:

qlc:q([X || X <- qlc:append(QH1, QH2)], {unique, true})

The cost is substantial: every returned answer is stored in an ETS table. Before returning an answer, it is looked up in the ETS table to check if it has already been returned. Without the unique option, all answers to QH1 would be returned followed by all answers to QH2. The unique option keeps the order between the remaining answers.

If the order of the answers is not important, there is an alternative to the unique option, namely to sort the answers uniquely:

qlc:sort(qlc:q([X || X <- qlc:append(QH1, QH2)], {unique, true})).

This query also removes duplicates but the answers are sorted. If there are many answers, temporary files are used. Notice that to get the first unique answer, all answers must be found and sorted. Both alternatives find duplicates by comparing answers, that is, if A1 and A2 are answers found in that order, then A2 is a removed if A1 == A2.

To return only a few answers, cursors can be used. The following code returns no more than five answers using an ETS table for storing the unique answers:

C = qlc:cursor(qlc:q([X || X <- qlc:append(QH1, QH2)],{unique,true})),
R = qlc:next_answers(C, 5),
ok = qlc:delete_cursor(C),
R.

QLCs are convenient for stating constraints on data from two or more tables. The following example does a natural join on two query handles on position 2:

qlc:q([{X1,X2,X3,Y1} ||
          {X1,X2,X3} <- QH1,
          {Y1,Y2} <- QH2,
          X2 =:= Y2])

The qlc module evaluates this differently depending on the query handles QH1 and QH2. If, for example, X2 is matched against the key of a QLC table, the lookup join method traverses the objects of QH2 while looking up key values in the table. However, if not X2 or Y2 is matched against the key or an indexed position of a QLC table, the merge join method ensures that QH1 and QH2 are both sorted on position 2 and next do the join by traversing the objects one by one.

Option join can be used to force the qlc module to use a certain join method. For the rest of this section it is assumed that the excessively slow join method called "nested loop" has been chosen:

qlc:q([{X1,X2,X3,Y1} ||
          {X1,X2,X3} <- QH1,
          {Y1,Y2} <- QH2,
          X2 =:= Y2],
      {join, nested_loop})

In this case the filter is applied to every possible pair of answers to QH1 and QH2, one at a time. If there are M answers to QH1 and N answers to QH2, the filter is run M*N times.

If QH2 is a call to the function for gb_trees, as defined in section Implementing a QLC Table, then gb_table:table/1 , the iterator for the gb-tree is initiated for each answer to QH1. The objects of the gb-tree are then returned one by one. This is probably the most efficient way of traversing the table in that case, as it takes minimal computational power to get the following object. But if QH2 is not a table but a more complicated QLC, it can be more efficient to use some RAM memory for collecting the answers in a cache, particularly if there are only a few answers. It must then be assumed that evaluating QH2 has no side effects so that the meaning of the query does not change if QH2 is evaluated only once. One way of caching the answers is to evaluate QH2 first of all and substitute the list of answers for QH2 in the query. Another way is to use option cache. It is expressed like this:

QH2' = qlc:q([X || X <- QH2], {cache, ets})

or only

QH2' = qlc:q([X || X <- QH2], cache)

The effect of option cache is that when generator QH2' is run the first time, every answer is stored in an ETS table. When the next answer of QH1 is tried, answers to QH2' are copied from the ETS table, which is very fast. As for option unique the cost is a possibly substantial amount of RAM memory.

Option {cache, list} offers the possibility to store the answers in a list on the process heap. This has the potential of being faster than ETS tables, as there is no need to copy answers from the table. However, it can often result in slower evaluation because of more garbage collections of the process heap and increased RAM memory consumption because of larger heaps. Another drawback with cache lists is that if the list size exceeds a limit, a temporary file is used. Reading the answers from a file is much slower than copying them from an ETS table. But if the available RAM memory is scarce, setting the limit to some low value is an alternative.

Option cache_all can be set to ets or list when evaluating a query. It adds a cache or {cache, list} option to every list expression except QLC tables and lists on all levels of the query. This can be used for testing if caching would improve efficiency at all. If the answer is yes, further testing is needed to pinpoint the generators that are to be cached.

Implementing a QLC Table

As an example of how to use function table/2, the implementation of a QLC table for the gb_trees module is given:

-module(gb_table).

-export([table/1]).

table(T) ->
    TF = fun() -> qlc_next(gb_trees:next(gb_trees:iterator(T))) end,
    InfoFun = fun(num_of_objects) -> gb_trees:size(T);
                 (keypos) -> 1;
                 (is_sorted_key) -> true;
                 (is_unique_objects) -> true;
                 (_) -> undefined
              end,
    LookupFun =
        fun(1, Ks) ->
                lists:flatmap(fun(K) ->
                                      case gb_trees:lookup(K, T) of
                                          {value, V} -> [{K,V}];
                                          none -> []
                                      end
                              end, Ks)
        end,
    FormatFun =
        fun({all, NElements, ElementFun}) ->
                ValsS = io_lib:format("gb_trees:from_orddict(~w)",
                                      [gb_nodes(T, NElements, ElementFun)]),
                io_lib:format("gb_table:table(~s)", [ValsS]);
           ({lookup, 1, KeyValues, _NElements, ElementFun}) ->
                ValsS = io_lib:format("gb_trees:from_orddict(~w)",
                                      [gb_nodes(T, infinity, ElementFun)]),
                io_lib:format("lists:flatmap(fun(K) -> "
                              "case gb_trees:lookup(K, ~s) of "
                              "{value, V} -> [{K,V}];none -> [] end "
                              "end, ~w)",
                              [ValsS, [ElementFun(KV) || KV <- KeyValues]])
        end,
    qlc:table(TF, [{info_fun, InfoFun}, {format_fun, FormatFun},
                   {lookup_fun, LookupFun},{key_equality,'=='}]).

qlc_next({X, V, S}) ->
    [{X,V} | fun() -> qlc_next(gb_trees:next(S)) end];
qlc_next(none) ->
    [].

gb_nodes(T, infinity, ElementFun) ->
    gb_nodes(T, -1, ElementFun);
gb_nodes(T, NElements, ElementFun) ->
    gb_iter(gb_trees:iterator(T), NElements, ElementFun).

gb_iter(_I, 0, _EFun) ->
    '...';
gb_iter(I0, N, EFun) ->
    case gb_trees:next(I0) of
        {X, V, I} ->
            [EFun({X,V}) | gb_iter(I, N-1, EFun)];
        none ->
            []
    end.

TF is the traversal function. The qlc module requires that there is a way of traversing all objects of the data structure. gb_trees has an iterator function suitable for that purpose. Notice that for each object returned, a new fun is created. As long as the list is not terminated by [], it is assumed that the tail of the list is a nullary function and that calling the function returns further objects (and functions).

The lookup function is optional. It is assumed that the lookup function always finds values much faster than it would take to traverse the table. The first argument is the position of the key. As qlc_next/1 returns the objects as {Key, Value} pairs, the position is 1. Notice that the lookup function is to return {Key, Value} pairs, as the traversal function does.

The format function is also optional. It is called by info/1,2 to give feedback at runtime of how the query is to be evaluated. Try to give as good feedback as possible without showing too much details. In the example, at most seven objects of the table are shown. The format function handles two cases: all means that all objects of the table are traversed; {lookup, 1, KeyValues} means that the lookup function is used for looking up key values.

Whether the whole table is traversed or only some keys looked up depends on how the query is expressed. If the query has the form

qlc:q([T || P <- LE, F])

and P is a tuple, the qlc module analyzes P and F in compile time to find positions of tuple P that are tested for equality to constants. If such a position at runtime turns out to be the key position, the lookup function can be used, otherwise all objects of the table must be traversed. The info function InfoFun returns the key position. There can be indexed positions as well, also returned by the info function. An index is an extra table that makes lookup on some position fast. Mnesia maintains indexes upon request, and introduces so called secondary keys. The qlc module prefers to look up objects using the key before secondary keys regardless of the number of constants to look up.

Key Equality

Erlang/OTP has two operators for testing term equality: ==/2 and =:=/2. The difference is all about the integers that can be represented by floats. For example, 2 == 2.0 evaluates to true while 2 =:= 2.0 evaluates to false. Normally this is a minor issue, but the qlc module cannot ignore the difference, which affects the user's choice of operators in QLCs.

If the qlc module at compile time can determine that some constant is free of integers, it does not matter which one of ==/2 or =:=/2 is used:

1> E1 = ets:new(t, [set]), % uses =:=/2 for key equality
Q1 = qlc:q([K ||
{K} <- ets:table(E1),
K == 2.71 orelse K == a]),
io:format("~s~n", [qlc:info(Q1)]).
ets:match_spec_run(
       lists:flatmap(fun(V) ->
			    ets:lookup(#Ref<0.3098908599.2283929601.256025>,
				       V)
		     end,
		     [a, 2.71]),
       ets:match_spec_compile([{{'$1'}, [], ['$1']}]))

In the example, operator ==/2 has been handled exactly as =:=/2 would have been handled. However, if it cannot be determined at compile time that some constant is free of integers, and the table uses =:=/2 when comparing keys for equality (see option key_equality), then the qlc module does not try to look up the constant. The reason is that there is in the general case no upper limit on the number of key values that can compare equal to such a constant; every combination of integers and floats must be looked up:

2> E2 = ets:new(t, [set]),
true = ets:insert(E2, [{{2,2},a},{{2,2.0},b},{{2.0,2},c}]),
F2 = fun(I) ->
qlc:q([V || {K,V} <- ets:table(E2), K == I])
end,
Q2 = F2({2,2}),
io:format("~s~n", [qlc:info(Q2)]).
ets:table(#Ref<0.3098908599.2283929601.256125>,
          [{traverse,
            {select,
             [{{'$1', '$2'}, [{'==', '$1', {const, {2, 2}}}], ['$2']}]}}])
3> lists:sort(qlc:e(Q2)).
[a,b,c]

Looking up only {2,2} would not return b and c.

If the table uses ==/2 when comparing keys for equality, the qlc module looks up the constant regardless of which operator is used in the QLC. However, ==/2 is to be preferred:

4> E3 = ets:new(t, [ordered_set]), % uses ==/2 for key equality
true = ets:insert(E3, [{{2,2.0},b}]),
F3 = fun(I) ->
qlc:q([V || {K,V} <- ets:table(E3), K == I])
end,
Q3 = F3({2,2}),
io:format("~s~n", [qlc:info(Q3)]).
ets:match_spec_run(ets:lookup(#Ref<0.3098908599.2283929601.256211>,
                              {2, 2}),
                   ets:match_spec_compile([{{'$1', '$2'}, [], ['$2']}]))
5> qlc:e(Q3).
[b]

Lookup join is handled analogously to lookup of constants in a table: if the join operator is ==/2, and the table where constants are to be looked up uses =:=/2 when testing keys for equality, then the qlc module does not consider lookup join for that table.

See Also

dets, erl_eval, erlang, error_logger, ets, file, file_sorter, mnesia, shell, Erlang Reference Manual, Programming Examples