![[Ericsson AB]](min_head.gif)
The earlier chapters of this User Guide described how to get started with Mnesia, and how to build a Mnesia database. In this chapter, we will describe the more advanced features available when building a distributed, fault tolerant Mnesia database. This chapter contains the following sections:
Data retrieval and matching can be performed very efficiently if we know the key for the record. Conversely, if the key is not known, all records in a table must be searched. The larger the table the more time consuming it will become. To remedy this problem Mnesia's indexing capabilities are used to improve data retrieval and matching of records.
The following two functions manipulate indexes on existing tables:
These functions create or delete a table index on field defined by AttributeName. To illustrate this, add an index to the table definition (employee, {emp_no, name, salary, sex, phone, room_no}, which is the example table from the Company database. The function which adds an index on the element salary can be expressed in the following way:
The indexing capabilities of Mnesia are utilized with the following three functions, which retrieve and match records on the basis of index entries in the database.
These functions are further described and exemplified in Chapter 4: Pattern matching.
Mnesia is a distributed, fault tolerant DBMS. It is possible to replicate tables on different Erlang nodes in a variety of ways. The Mnesia programmer does not have to state where the different tables reside, only the names of the different tables are specified in the program code. This is known as "location transparency" and it is an important concept. In particular:
We have previously seen that each table has a number of system attributes, such as index and type.
Table attributes are specified when the table is created. For example, the following function will create a new table with two RAM replicas:
mnesia:create_table(foo,
[{ram_copies, [N1, N2]},
{attributes, record_info(fields, foo)}]).
Tables can also have the following properties, where each attribute has a list of Erlang nodes as its value.
It is also possible to set and change table properties on existing tables. Refer to Chapter 3: Defining the Schema for full details.
There are basically two reasons for using more than one table replica: fault tolerance, or speed. It is worthwhile to note that table replication provides a solution to both of these system requirements.
If we have two active table replicas, all information is still available if one of the replicas fail. This can be a very important property in many applications. Furthermore, if a table replica exists at two specific nodes, applications which execute at either of these nodes can read data from the table without accessing the network. Network operations are considerably slower and consume more resources than local operations.
It can be advantageous to create table replicas for a distributed application which reads data often, but writes data seldom, in order to achieve fast read operations on the local node. The major disadvantage with replication is the increased time to write data. If a table has two replicas, every write operation must access both table replicas. Since one of these write operations must be a network operation, it is considerably more expensive to perform a write operation to a replicated table than to a non-replicated table.
A concept of table fragmentation has been introduced in order to cope with very large tables. The idea is to split a table into several more manageable fragments. Each fragment is implemented as a first class Mnesia table and may be replicated, have indices etc. as any other table. But the tables may neither have local_content nor have the snmp connection activated.
In order to be able to access a record in a fragmented table, Mnesia must determine to which fragment the actual record belongs. This is done by the mnesia_frag module, which implements the mnesia_access callback behaviour. Please, read the documentation about mnesia:activity/4 to see how mnesia_frag can be used as a mnesia_access callback module.
At each record access mnesia_frag first computes a hash value from the record key. Secondly the name of the table fragment is determined from the hash value. And finally the actual table access is performed by the same functions as for non-fragmented tables. When the key is not known beforehand, all fragments are searched for matching records. Note: In ordered_set tables the records will be ordered per fragment, and the the order is undefined in results returned by select and match_object.
The following piece of code illustrates how an existing Mnesia table is converted to be a fragmented table and how more fragments are added later on.
Eshell V4.7.3.3 (abort with ^G)
(a@sam)1> mnesia:start().
ok
(a@sam)2> mnesia:system_info(running_db_nodes).
[b@sam,c@sam,a@sam]
(a@sam)3> Tab = dictionary.
dictionary
(a@sam)4> mnesia:create_table(Tab, [{ram_copies, [a@sam, b@sam]}]).
{atomic,ok}
(a@sam)5> Write = fun(Keys) -> [mnesia:write({Tab,K,-K}) || K <- Keys], ok end.
#Fun<erl_eval>
(a@sam)6> mnesia:activity(sync_dirty, Write, [lists:seq(1, 256)], mnesia_frag).
ok
(a@sam)7> mnesia:change_table_frag(Tab, {activate, []}).
{atomic,ok}
(a@sam)8> mnesia:table_info(Tab, frag_properties).
[{base_table,dictionary},
{foreign_key,undefined},
{n_doubles,0},
{n_fragments,1},
{next_n_to_split,1},
{node_pool,[a@sam,b@sam,c@sam]}]
(a@sam)9> Info = fun(Item) -> mnesia:table_info(Tab, Item) end.
#Fun<erl_eval>
(a@sam)10> Dist = mnesia:activity(sync_dirty, Info, [frag_dist], mnesia_frag).
[{c@sam,0},{a@sam,1},{b@sam,1}]
(a@sam)11> mnesia:change_table_frag(Tab, {add_frag, Dist}).
{atomic,ok}
(a@sam)12> Dist2 = mnesia:activity(sync_dirty, Info, [frag_dist], mnesia_frag).
[{b@sam,1},{c@sam,1},{a@sam,2}]
(a@sam)13> mnesia:change_table_frag(Tab, {add_frag, Dist2}).
{atomic,ok}
(a@sam)14> Dist3 = mnesia:activity(sync_dirty, Info, [frag_dist], mnesia_frag).
[{a@sam,2},{b@sam,2},{c@sam,2}]
(a@sam)15> mnesia:change_table_frag(Tab, {add_frag, Dist3}).
{atomic,ok}
(a@sam)16> Read = fun(Key) -> mnesia:read({Tab, Key}) end.
#Fun<erl_eval>
(a@sam)17> mnesia:activity(transaction, Read, [12], mnesia_frag).
[{dictionary,12,-12}]
(a@sam)18> mnesia:activity(sync_dirty, Info, [frag_size], mnesia_frag).
[{dictionary,64},
{dictionary_frag2,64},
{dictionary_frag3,64},
{dictionary_frag4,64}]
(a@sam)19>
There is a table property called frag_properties and may be read with mnesia:table_info(Tab, frag_properties). The fragmentation properties is a list of tagged tuples with the arity 2. By default the list is empty, but when it is non-empty it triggers Mnesia to regard the table as fragmented. The fragmentation properties are:
Eshell V4.7.3.3 (abort with ^G)
(a@sam)1> mnesia:start().
ok
(a@sam)2> PrimProps = [{n_fragments, 7}, {node_pool, [node()]}].
[{n_fragments,7},{node_pool,[a@sam]}]
(a@sam)3> mnesia:create_table(prim_dict,
[{frag_properties, PrimProps},
{attributes,[prim_key,prim_val]}]).
{atomic,ok}
(a@sam)4> SecProps = [{foreign_key, {prim_dict, sec_val}}].
[{foreign_key,{prim_dict,sec_val}}]
(a@sam)5> mnesia:create_table(sec_dict,
[{frag_properties, SecProps},
(a@sam)5> {attributes, [sec_key, sec_val]}]).
{atomic,ok}
(a@sam)6> Write = fun(Rec) -> mnesia:write(Rec) end.
#Fun<erl_eval>
(a@sam)7> PrimKey = 11.
11
(a@sam)8> SecKey = 42.
42
(a@sam)9> mnesia:activity(sync_dirty, Write,
[{prim_dict, PrimKey, -11}], mnesia_frag).
ok
(a@sam)10> mnesia:activity(sync_dirty, Write,
[{sec_dict, SecKey, PrimKey}], mnesia_frag).
ok
(a@sam)11> mnesia:change_table_frag(prim_dict, {add_frag, [node()]}).
{atomic,ok}
(a@sam)12> SecRead = fun(PrimKey, SecKey) ->
mnesia:read({sec_dict, PrimKey}, SecKey, read) end.
#Fun<erl_eval>
(a@sam)13> mnesia:activity(transaction, SecRead,
[PrimKey, SecKey], mnesia_frag).
[{sec_dict,42,11}]
(a@sam)14> Info = fun(Tab, Item) -> mnesia:table_info(Tab, Item) end.
#Fun<erl_eval>
(a@sam)15> mnesia:activity(sync_dirty, Info,
[prim_dict, frag_size], mnesia_frag).
[{prim_dict,0},
{prim_dict_frag2,0},
{prim_dict_frag3,0},
{prim_dict_frag4,1},
{prim_dict_frag5,0},
{prim_dict_frag6,0},
{prim_dict_frag7,0},
{prim_dict_frag8,0}]
(a@sam)16> mnesia:activity(sync_dirty, Info,
[sec_dict, frag_size], mnesia_frag).
[{sec_dict,0},
{sec_dict_frag2,0},
{sec_dict_frag3,0},
{sec_dict_frag4,1},
{sec_dict_frag5,0},
{sec_dict_frag6,0},
{sec_dict_frag7,0},
{sec_dict_frag8,0}]
(a@sam)17>
The function mnesia:change_table_frag(Tab, Change) is intended to be used for reconfiguration of fragmented tables. The Change argument should have one of the following values:
The function mnesia:create_table/2 is used to create a brand new fragmented table, by setting the table property frag_properties to some proper values.
The function mnesia:delete_table/1 is used to delete a fragmented table including all its fragments. There must however not exist any other fragmented tables which refers to this table in their foreign key.
The function mnesia:table_info/2 now understands the frag_properties item.
If the function mnesia:table_info/2 is invoked in the activity context of the mnesia_frag module, information of several new items may be obtained:
There are several algorithms for distributing records in a fragmented table evenly over a pool of nodes. No one is best, it simply depends of the application needs. Here follows some examples of situations which may need some attention:
permanent change of nodes when a new permanent db_node is introduced or dropped, it may be time to change the pool of nodes and re-distribute the replicas evenly over the new pool of nodes. It may also be time to add or delete a fragment before the replicas are re-distributed.
size/memory threshold when the total size or total memory of a fragmented table (or a single fragment) exceeds some application specific threshold, it may be time to dynamically add a new fragment in order obtain a better distribution of records.
temporary node down when a node temporarily goes down it may be time to compensate some fragments with new replicas in order to keep the desired level of redundancy. When the node comes up again it may be time to remove the superfluous replica.
overload threshold when the load on some node is exceeds some application specific threshold, it may be time to either add or move some fragment replicas to nodes with lesser load. Extra care should be taken if the table has a foreign key relation to some other table. In order to avoid severe performance penalties, the same re-distribution must be performed for all of the related tables.
Use mnesia:change_table_frag/2 to add new fragments and apply the usual schema manipulation functions (such as mnesia:add_table_copy/3, mnesia:del_table_copy/2 and mnesia:change_table_copy_type/2) on each fragment to perform the actual re-distribution.
Replicated tables have the same content on all nodes where they are replicated. However, it is sometimes advantageous to have tables but different content on different nodes.
If we specify the attribute {local_content, true} when we create the table, the table will reside on the nodes where we specify that the table shall exist, but the write operations on the table will only be performed on the local copy.
Furthermore, when the table is initialized at start-up, the table will only be initialized locally, and the table content will not be copied from another node.
It is possible to run Mnesia on nodes that do not have a disc. It is of course not possible to have replicas of neither disc_copies, nor disc_only_copies on such nodes. This especially troublesome for the schema table since Mnesia need the schema in order to initialize itself.
The schema table may, as other tables, reside on one or more nodes. The storage type of the schema table may either be disc_copies or ram_copies (not disc_only_copies). At start-up Mnesia uses its schema to determine with which nodes it should try to establish contact. If any of the other nodes are already started, the starting node merges its table definitions with the table definitions brought from the other nodes. This also applies to the definition of the schema table itself. The application parameter extra_db_nodes contains a list of nodes which Mnesia also should establish contact with besides the ones found in the schema. The default value is the empty list [].
Hence, when a disc-less node needs to find the schema definitions from a remote node on the network, we need to supply this information through the application parameter -mnesia extra_db_nodes NodeList. Without this configuration parameter set, Mnesia will start as a single node system. It is also possible to use mnesia:change_config/2 to assign a value to 'extra_db_nodes' and force a connection after mnesia have been started, i.e. mnesia:change_config(extra_db_nodes, NodeList).
The application parameter schema_location controls where Mnesia will search for its schema. The parameter may be one of the following atoms:
When the schema_location is set to opt_disc the function mnesia:change_table_copy_type/3 may be used to change the storage type of the schema. This is illustrated below:
1> mnesia:start().
ok
2> mnesia:change_table_copy_type(schema, node(), disc_copies).
{atomic, ok}
Assuming that the call to mnesia:start did not find any schema to read on the disc, then Mnesia has started as a disc-less node, and then changed it to a node that utilizes the disc to locally store the schema.
It is possible to add and remove nodes from a Mnesia system. This can be done by adding a copy of the schema to those nodes.
The functions mnesia:add_table_copy/3 and mnesia:del_table_copy/2 may be used to add and delete replicas of the schema table. Adding a node to the list of nodes where the schema is replicated will affect two things. First it allows other tables to be replicated to this node. Secondly it will cause Mnesia to try to contact the node at start-up of disc-full nodes.
The function call mnesia:del_table_copy(schema, mynode@host) deletes the node 'mynode@host' from the Mnesia system. The call fails if mnesia is running on 'mynode@host'. The other mnesia nodes will never try to connect to that node again. Note, if there is a disc resident schema on the node 'mynode@host', the entire mnesia directory should be deleted. This can be done with mnesia:delete_schema/1. If mnesia is started again on the the node 'mynode@host' and the directory has not been cleared, mnesia's behaviour is undefined.
If the storage type of the schema is ram_copies, i.e, we have disc-less node, Mnesia will not use the disc on that particular node. The disc usage is enabled by changing the storage type of the table schema to disc_copies.
New schemas are created explicitly with mnesia:create_schema/1 or implicitly by starting Mnesia without a disc resident schema. Whenever a table (including the schema table) is created it is assigned its own unique cookie. The schema table is not created with mnesia:create_table/2 as normal tables.
At start-up Mnesia connects different nodes to each other, then they exchange table definitions with each other and the table definitions are merged. During the merge procedure Mnesia performs a sanity test to ensure that the table definitions are compatible with each other. If a table exists on several nodes the cookie must be the same, otherwise Mnesia will shutdown one of the nodes. This unfortunate situation will occur if a table has been created on two nodes independently of each other while they were disconnected. To solve the problem, one of the tables must be deleted (as the cookies differ we regard it to be two different tables even if they happen to have the same name).
Merging different versions of the schema table, does not always require the cookies to be the same. If the storage type of the schema table is disc_copies, the cookie is immutable, and all other db_nodes must have the same cookie. When the schema is stored as type ram_copies, its cookie can be replaced with a cookie from another node (ram_copies or disc_copies). The cookie replacement (during merge of the schema table definition) is performed each time a RAM node connects to another node.
mnesia:system_info(schema_location) and mnesia:system_info(extra_db_nodes) may be used to determine the actual values of schema_location and extra_db_nodes respectively. mnesia:system_info(use_dir) may be used to determine whether Mnesia is actually using the Mnesia directory. use_dir may be determined even before Mnesia is started. The function mnesia:info/0 may now be used to printout some system information even before Mnesia is started. When Mnesia is started the function prints out more information.
Transactions which update the definition of a table, requires that Mnesia is started on all nodes where the storage type of the schema is disc_copies. All replicas of the table on these nodes must also be loaded. There are a few exceptions to these availability rules. Tables may be created and new replicas may be added without starting all of the disc-full nodes. New replicas may be added before all other replicas of the table have been loaded, it will suffice when one other replica is active.
System events and table events are the two categories of events that Mnesia will generate in various situations.
It is possible for user process to subscribe on the events generated by Mnesia. We have the following two functions:
Event-Category may either be the atom system, or one of the tuples {table, Tab, simple}, {table, Tab, detailed}. The old event-category {table, Tab} is the same event-category as {table, Tab, simple}. The subscribe functions activate a subscription of events. The events are delivered as messages to the process evaluating the mnesia:subscribe/1 function. The syntax of system events is {mnesia_system_event, Event} and {mnesia_table_event, Event} for table events. What system events and table events means is described below.
All system events are subscribed by Mnesia's gen_event handler. The default gen_event handler is mnesia_event. But it may be changed by using the application parameter event_module. The value of this parameter must be the name of a module implementing a complete handler as specified by the gen_event module in STDLIB. mnesia:system_info(subscribers) and mnesia:table_info(Tab, subscribers) may be used to determine which processes are subscribed to various events.
The system events are detailed below:
Another category of events are table events, which are events related to table updates. There are two types of table events simple and detailed.
The simple table events are tuples looking like this: {Oper, Record, ActivityId}. Where Oper is the operation performed. Record is the record involved in the operation and ActivityId is the identity of the transaction performing the operation. Note that the name of the record is the table name even when the record_name has another setting. The various table related events that may occur are:
The detailed table events are tuples looking like this: {Oper, Table, Data, [OldRecs], ActivityId}. Where Oper is the operation performed. Table is the table involved in the operation, Data is the record/oid written/deleted. OldRecs is the contents before the operation. and ActivityId is the identity of the transaction performing the operation. The various table related events that may occur are:
Debugging a Mnesia application can be difficult due to a number of reasons, primarily related to difficulties in understanding how the transaction and table load mechanisms work. An other source of confusion may be the semantics of nested transactions.
We may set the debug level of Mnesia by calling:
Where the parameter Level is:
The debug level of Mnesia itself, is also an application parameter, thereby making it possible to start an Erlang system in order to turn on Mnesia debug in the initial start-up phase by using the following code:
% erl -mnesia debug verbose
Programming concurrent Erlang systems is the subject of a separate book. However, it is worthwhile to draw attention to the following features, which permit concurrent processes to exist in a Mnesia system.
A group of functions or processes can be called within a transaction. A transaction may include statements that read, write or delete data from the DBMS. A large number of such transactions can run concurrently, and the programmer does not have to explicitly synchronize the processes which manipulate the data. All programs accessing the database through the transaction system may be written as if they had sole access to the data. This is a very desirable property since all synchronization is taken care of by the transaction handler. If a program reads or writes data, the system ensures that no other program tries to manipulate the same data at the same time.
It is possible to move tables, delete tables or reconfigure the layout of a table in various ways. An important aspect of the actual implementation of these functions is that it is possible for user programs to continue to use a table while it is being reconfigured. For example, it is possible to simultaneously move a table and perform write operations to the table . This is important for many applications that require continuously available services. Refer to Chapter 4: Transactions and other access contexts for more information.
If and when we decide that we would like to start and manipulate Mnesia, it is often easier to write the definitions and data into an ordinary text file. Initially, no tables and no data exist, or which tables are required. At the initial stages of prototyping it is prudent write all data into one file, process that file and have the data in the file inserted into the database. It is possible to initialize Mnesia with data read from a text file. We have the following two functions to work with text files.
These functions are of course much slower than the ordinary store and load functions of Mnesia. However, this is mainly intended for minor experiments and initial prototyping. The major advantages of these functions is that they are very easy to use.
The format of the text file is:
{tables, [{Typename, [Options]},
{Typename2 ......}]}.
{Typename, Attribute1, Atrribute2 ....}.
{Typename, Attribute1, Atrribute2 ....}.
Options is a list of {Key,Value} tuples conforming to the options we could give to mnesia:create_table/2.
For example, if we want to start playing with a small database for healthy foods, we enter then following data into the file FRUITS.
{tables,
[{fruit, [{attributes, [name, color, taste]}]},
{vegetable, [{attributes, [name, color, taste, price]}]}]}.
{fruit, orange, orange, sweet}.
{fruit, apple, green, sweet}.
{vegetable, carrot, orange, carrotish, 2.55}.
{vegetable, potato, yellow, none, 0.45}.
The following session with the Erlang shell then shows how to load the fruits database.
% erl
Erlang (BEAM) emulator version 4.9
Eshell V4.9 (abort with ^G)
1> mnesia:load_textfile("FRUITS").
New table fruit
New table vegetable
{atomic,ok}
2> mnesia:info().
---> Processes holding locks <---
---> Processes waiting for locks <---
---> Pending (remote) transactions <---
---> Active (local) transactions <---
---> Uncertain transactions <---
---> Active tables <---
vegetable : with 2 records occuping 299 words of mem
fruit : with 2 records occuping 291 words of mem
schema : with 3 records occuping 401 words of mem
===> System info in version "1.1", debug level = none <===
opt_disc. Directory "/var/tmp/Mnesia.nonode@nohost" is used.
use fallback at restart = false
running db nodes = [nonode@nohost]
stopped db nodes = []
remote = []
ram_copies = [fruit,vegetable]
disc_copies = [schema]
disc_only_copies = []
[{nonode@nohost,disc_copies}] = [schema]
[{nonode@nohost,ram_copies}] = [fruit,vegetable]
3 transactions committed, 0 aborted, 0 restarted, 2 logged to disc
0 held locks, 0 in queue; 0 local transactions, 0 remote
0 transactions waits for other nodes: []
ok
3>
Where we can see that the DBMS was initiated from a regular text file.
The Company database introduced in Chapter 2 has three tables which store records (employee, dept, project), and three tables which store relationships (manager, at_dep, in_proj). This is a normalized data model, which has some advantages over a non-normalized data model.
It is more efficient to do a generalized search in a normalized database. Some operations are also easier to perform on a normalized data model. For example, we can easily remove one project, as the following example illustrates:
remove_proj(ProjName) ->
F = fun() ->
Ip = qlc:e(qlc:q([X || X <- mnesia:table(in_proj),
X#in_proj.proj_name == ProjName]
)),
mnesia:delete({project, ProjName}),
del_in_projs(Ip)
end,
mnesia:transaction(F).
del_in_projs([Ip|Tail]) ->
mnesia:delete_object(Ip),
del_in_projs(Tail);
del_in_projs([]) ->
done.
In reality, data models are seldom fully normalized. A realistic alternative to a normalized database model would be a data model which is not even in first normal form. Mnesia is very suitable for applications such as telecommunications, because it is easy to organize data in a very flexible manner. A Mnesia database is always organized as a set of tables. Each table is filled with rows/objects/records. What sets Mnesia apart is that individual fields in a record can contain any type of compound data structures. An individual field in a record can contain lists, tuples, functions, and even record code.
Many telecommunications applications have unique requirements on lookup times for certain types of records. If our Company database had been a part of a telecommunications system, then it could be that the lookup time of an employee together with a list of the projects the employee is working on, should be minimized. If this was the case, we might choose a drastically different data model which has no direct relationships. We would only have the records themselves, and different records could contain either direct references to other records, or they could contain other records which are not part of the Mnesia schema.
We could create the following record definitions:
-record(employee, {emp_no,
name,
salary,
sex,
phone,
room_no,
dept,
projects,
manager}).
-record(dept, {id,
name}).
-record(project, {name,
number,
location}).
An record which describes an employee might look like this:
Me = #employee{emp_no= 104732,
name = klacke,
salary = 7,
sex = male,
phone = 99586,
room_no = {221, 015},
dept = 'B/SFR',
projects = [erlang, mnesia, otp],
manager = 114872},
This model only has three different tables, and the employee records contain references to other records. We have the following references in the record.
We could also use the Mnesia record identifiers ({Tab, Key}) as references. In this case, the dept attribute would be set to the value {dept, 'B/SFR'} instead of 'B/SFR'.
With this data model, some operations execute considerably faster than they do with the normalized data model in our Company database. On the other hand, some other operations become much more complicated. In particular, it becomes more difficult to ensure that records do not contain dangling pointers to other non-existent, or deleted, records.
The following code exemplifies a search with a non-normalized data model. To find all employees at department Dep with a salary higher than Salary, use the following code:
get_emps(Salary, Dep) ->
Q = qlc:q(
[E || E <- mnesia:table(employee),
E#employee.salary > Salary,
E#employee.dept == Dep]
),
F = fun() -> qlc:e(Q) end,
transaction(F).
This code is not only easier to write and to understand, but it also executes much faster.
It is easy to show examples of code which executes faster if we use a non-normalized data model, instead of a normalized model. The main reason for this is that fewer tables are required. For this reason, we can more easily combine data from different tables in join operations. In the above example, the get_emps/2 function was transformed from a join operation into a simple query which consists of a selection and a projection on one single table.