20012009 Ericsson AB. All Rights Reserved. The contents of this file are subject to the Erlang Public License, Version 1.1, (the "License"); you may not use this file except in compliance with the License. You should have received a copy of the Erlang Public License along with this software. If not, it can be retrieved online at http://www.erlang.org/. Software distributed under the License is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See the License for the specific language governing rights and limitations under the License. Tables and databases Ingela Anderton 2001-08-07 tablesDatabases.xml
Ets, Dets and Mnesia

Every example using Ets has a corresponding example in Mnesia. In general all Ets examples also apply to Dets tables.

Select/Match operations

Select/Match operations on Ets and Mnesia tables can become very expensive operations. They usually need to scan the complete table. You should try to structure your data so that you minimize the need for select/match operations. However, if you really need a select/match operation, it will still be more efficient than using tab2list. Examples of this and also of ways to avoid select/match will be provided in some of the following sections. The functions ets:select/2 and mnesia:select/3 should be preferred over ets:match/2,ets:match_object/2, and mnesia:match_object/3.

There are exceptions when the complete table is not scanned, for instance if part of the key is bound when searching an ordered_set table, or if it is a Mnesia table and there is a secondary index on the field that is selected/matched. If the key is fully bound there will, of course, be no point in doing a select/match, unless you have a bag table and you are only interested in a sub-set of the elements with the specific key.

When creating a record to be used in a select/match operation you want most of the fields to have the value '_'. The easiest and fastest way to do that is as follows:

#person{age = 42, _ = '_'}. 
Deleting an element

The delete operation is considered successful if the element was not present in the table. Hence all attempts to check that the element is present in the Ets/Mnesia table before deletion are unnecessary. Here follows an example for Ets tables.

DO

...
ets:delete(Tab, Key),
...

DO NOT

...
case ets:lookup(Tab, Key) of
    [] ->
        ok;
    [_|_] ->
        ets:delete(Tab, Key)
end,
...
Data fetching

Do not fetch data that you already have! Consider that you have a module that handles the abstract data type Person. You export the interface function print_person/1 that uses the internal functions print_name/1, print_age/1, print_occupation/1.

If the functions print_name/1 and so on, had been interface functions the matter comes in to a whole new light, as you do not want the user of the interface to know about the internal data representation.

DO

%%% Interface function print_person(PersonId) -> %% Look up the person in the named table person, case ets:lookup(person, PersonId) of [Person] -> print_name(Person), print_age(Person), print_occupation(Person); [] -> io:format("No person with ID = ~p~n", [PersonID]) end. %%% Internal functions print_name(Person) -> io:format("No person ~p~n", [Person#person.name]). print_age(Person) -> io:format("No person ~p~n", [Person#person.age]). print_occupation(Person) -> io:format("No person ~p~n", [Person#person.occupation]).

DO NOT

%%% Interface function print_person(PersonId) -> %% Look up the person in the named table person, case ets:lookup(person, PersonId) of [Person] -> print_name(PersonID), print_age(PersonID), print_occupation(PersonID); [] -> io:format("No person with ID = ~p~n", [PersonID]) end. %%% Internal functionss print_name(PersonID) -> [Person] = ets:lookup(person, PersonId), io:format("No person ~p~n", [Person#person.name]). print_age(PersonID) -> [Person] = ets:lookup(person, PersonId), io:format("No person ~p~n", [Person#person.age]). print_occupation(PersonID) -> [Person] = ets:lookup(person, PersonId), io:format("No person ~p~n", [Person#person.occupation]).
Non-persistent data storage

For non-persistent database storage, prefer Ets tables over Mnesia local_content tables. Even the Mnesia dirty_write operations carry a fixed overhead compared to Ets writes. Mnesia must check if the table is replicated or has indices, this involves at least one Ets lookup for each dirty_write. Thus, Ets writes will always be faster than Mnesia writes.

tab2list

Assume we have an Ets-table, which uses idno as key, and contains:

[#person{idno = 1, name = "Adam",  age = 31, occupation = "mailman"},
 #person{idno = 2, name = "Bryan", age = 31, occupation = "cashier"},
 #person{idno = 3, name = "Bryan", age = 35, occupation = "banker"},
 #person{idno = 4, name = "Carl",  age = 25, occupation = "mailman"}]

If we must return all data stored in the Ets-table we can use ets:tab2list/1. However, usually we are only interested in a subset of the information in which case ets:tab2list/1 is expensive. If we only want to extract one field from each record, e.g., the age of every person, we should use:

DO

...
ets:select(Tab,[{ #person{idno='_', 
                          name='_', 
                          age='$1', 
                          occupation = '_'},
                [],
                ['$1']}]),
...

DO NOT

...
TabList = ets:tab2list(Tab),
lists:map(fun(X) -> X#person.age end, TabList),
...

If we are only interested in the age of all persons named Bryan, we should:

DO

...
ets:select(Tab,[{ #person{idno='_', 
                          name="Bryan", 
                          age='$1', 
                          occupation = '_'},
                [],
                ['$1']}]),
...

DO NOT

...
TabList = ets:tab2list(Tab),
lists:foldl(fun(X, Acc) -> case X#person.name of
                                "Bryan" ->
                                    [X#person.age|Acc];
                                 _ ->
                                     Acc
                           end
             end, [], TabList),
...

REALLY DO NOT

...
TabList = ets:tab2list(Tab),
BryanList = lists:filter(fun(X) -> X#person.name == "Bryan" end,
                         TabList),
lists:map(fun(X) -> X#person.age end, BryanList),
...

If we need all information stored in the Ets table about persons named Bryan we should:

DO

...
ets:select(Tab, [{#person{idno='_', 
                          name="Bryan", 
                          age='_', 
                          occupation = '_'}, [], ['$_']}]),
...

DO NOT

...
TabList = ets:tab2list(Tab),
lists:filter(fun(X) -> X#person.name == "Bryan" end, TabList),
...
Ordered_set tables

If the data in the table should be accessed so that the order of the keys in the table is significant, the table type ordered_set could be used instead of the more usual set table type. An ordered_set is always traversed in Erlang term order with regard to the key field so that return values from functions such as select, match_object, and foldl are ordered by the key values. Traversing an ordered_set with the first and next operations also returns the keys ordered.

An ordered_set only guarantees that objects are processed in key order. Results from functions as ets:select/2 appear in the key order even if the key is not included in the result.

Ets specific
Utilizing the keys of the Ets table

An Ets table is a single key table (either a hash table or a tree ordered by the key) and should be used as one. In other words, use the key to look up things whenever possible. A lookup by a known key in a set Ets table is constant and for a ordered_set Ets table it is O(logN). A key lookup is always preferable to a call where the whole table has to be scanned. In the examples above, the field idno is the key of the table and all lookups where only the name is known will result in a complete scan of the (possibly large) table for a matching result.

A simple solution would be to use the name field as the key instead of the idno field, but that would cause problems if the names were not unique. A more general solution would be to create a second table with name as key and idno as data, i.e. to index (invert) the table with regards to the name field. The second table would of course have to be kept consistent with the master table. Mnesia could do this for you, but a home brew index table could be very efficient compared to the overhead involved in using Mnesia.

An index table for the table in the previous examples would have to be a bag (as keys would appear more than once) and could have the following contents:

 
[#index_entry{name="Adam", idno=1},
 #index_entry{name="Bryan", idno=2},
 #index_entry{name="Bryan", idno=3},
 #index_entry{name="Carl", idno=4}]

Given this index table a lookup of the age fields for all persons named "Bryan" could be done like this:

...
MatchingIDs = ets:lookup(IndexTable,"Bryan"),
lists:map(fun(#index_entry{idno = ID}) ->
                 [#person{age = Age}] = ets:lookup(PersonTable, ID),
                 Age
          end,
          MatchingIDs),
...

Note that the code above never uses ets:match/2 but instead utilizes the ets:lookup/2 call. The lists:map/2 call is only used to traverse the idnos matching the name "Bryan" in the table; therefore the number of lookups in the master table is minimized.

Keeping an index table introduces some overhead when inserting records in the table, therefore the number of operations gained from the table has to be weighted against the number of operations inserting objects in the table. However, note that the gain when the key can be used to lookup elements is significant.

Mnesia specific
Secondary index

If you frequently do a lookup on a field that is not the key of the table, you will lose performance using "mnesia:select/match_object" as this function will traverse the whole table. You may create a secondary index instead and use "mnesia:index_read" to get faster access, however this will require more memory. Example:

-record(person, {idno, name, age, occupation}).
        ...
{atomic, ok} = 
mnesia:create_table(person, [{index,[#person.age]},
                              {attributes,
                                    record_info(fields, person)}]),
{atomic, ok} = mnesia:add_table_index(person, age), 
...

PersonsAge42 =
     mnesia:dirty_index_read(person, 42, #person.age),
...
Transactions

Transactions is a way to guarantee that the distributed Mnesia database remains consistent, even when many different processes update it in parallel. However if you have real time requirements it is recommended to use dirty operations instead of transactions. When using the dirty operations you lose the consistency guarantee, this is usually solved by only letting one process update the table. Other processes have to send update requests to that process.

...
% Using transaction

Fun = fun() ->
          [mnesia:read({Table, Key}),
           mnesia:read({Table2, Key2})]
      end, 

{atomic, [Result1, Result2]}  = mnesia:transaction(Fun),
...

% Same thing using dirty operations
...

Result1 = mnesia:dirty_read({Table, Key}),
Result2 = mnesia:dirty_read({Table2, Key2}),
...