7  Mnesia System Information

7 Mnesia System Information

The following topics are included:

  • Database configuration data
  • Core dumps
  • Dumping tables
  • Checkpoints
  • Startup files, log file, and data files
  • Loading tables at startup
  • Recovery from communication failure
  • Recovery of transactions
  • Backup, restore, fallback, and disaster recovery

The following two functions can be used to retrieve system information. For details, see the Reference Manual.

If Mnesia malfunctions, system information is dumped to file MnesiaCore.Node.When. The type of system information contained in this file can also be generated with the function mnesia_lib:coredump(). If a Mnesia system behaves strangely, it is recommended that a Mnesia core dump file is included in the bug report.

Tables of type ram_copies are by definition stored in memory only. However, these tables can be dumped to disc, either at regular intervals or before the system is shut down. The function mnesia:dump_tables(TabList) dumps all replicas of a set of RAM tables to disc. The tables can be accessed while being dumped to disc. To dump the tables to disc, all replicas must have the storage type ram_copies.

The table content is placed in a .DCD file on the disc. When the Mnesia system is started, the RAM table is initially loaded with data from its .DCD file.

A checkpoint is a transaction consistent state that spans over one or more tables. When a checkpoint is activated, the system remembers the current content of the set of tables. The checkpoint retains a transaction consistent state of the tables, allowing the tables to be read and updated while the checkpoint is active. A checkpoint is typically used to back up tables to external media, but they are also used internally in Mnesia for other purposes. Each checkpoint is independent and a table can be involved in several checkpoints simultaneously.

Each table retains its old contents in a checkpoint retainer. For performance critical applications, it can be important to realize the processing overhead associated with checkpoints. In a worst case scenario, the checkpoint retainer consumes more memory than the table itself. Also, each update becomes slightly slower on those nodes where checkpoint retainers are attached to the tables.

For each table, it is possible to choose if there is to be one checkpoint retainer attached to all replicas of the table, or if it is enough to have only one checkpoint retainer attached to a single replica. With a single checkpoint retainer per table, the checkpoint consumes less memory, but it is vulnerable to node crashes. With several redundant checkpoint retainers, the checkpoint survives as long as there is at least one active checkpoint retainer attached to each table.

Checkpoints can be explicitly deactivated with the function mnesia:deactivate_checkpoint(Name), where Name is the name of an active checkpoint. This function returns ok if successful or {error, Reason} if there is an error. All tables in a checkpoint must be attached to at least one checkpoint retainer. The checkpoint is automatically deactivated by Mnesia, when any table lacks a checkpoint retainer. This can occur when a node goes down or when a replica is deleted. Use arguments min and max (described in the following list) to control the degree of checkpoint retainer redundancy.

Checkpoints are activated with the function mnesia:activate_checkpoint(Args), where Args is a list of the following tuples:

  • {name,Name}, where Name specifies a temporary name of the checkpoint. The name can be reused when the checkpoint has been deactivated. If no name is specified, a name is generated automatically.
  • {max,MaxTabs}, where MaxTabs is a list of tables that are to be included in the checkpoint. Default is [] (empty list). For these tables, the redundancy is maximized. The old content of the table is retained in the checkpoint retainer when the main table is updated by the applications. The checkpoint is more fault tolerant if the tables have several replicas. When new replicas are added by the schema manipulation function mnesia:add_table_copy/3 it also attaches a local checkpoint retainer.
  • {min,MinTabs}, where MinTabs is a list of tables that are to be included in the checkpoint. Default is []. For these tables, the redundancy is minimized, and there is to be single checkpoint retainer per table, preferably at the local node.
  • {allow_remote,Bool}, where false means that all checkpoint retainers must be local. If a table does not reside locally, the checkpoint cannot be activated. true allows checkpoint retainers to be allocated on any node. Default is true.
  • {ram_overrides_dump,Bool}. This argument only applies to tables of type ram_copies. Bool specifies if the table state in RAM is to override the table state on disc. true means that the latest committed records in RAM are included in the checkpoint retainer. These are the records that the application accesses. false means that the records on the disc .DAT file are included in the checkpoint retainer. These records are loaded on startup. Default is false.

The function mnesia:activate_checkpoint(Args) returns one of the following values:

  • {ok, Name, Nodes}
  • {error, Reason}

Name is the checkpoint name. Nodes are the nodes where the checkpoint is known.

A list of active checkpoints can be obtained with the following functions:

This section describes the internal files that are created and maintained by the Mnesia system. In particular, the workings of the Mnesia log are described.

Start Mnesia states the following prerequisites for starting Mnesia:

  • An Erlang session must be started and a Mnesia directory must be specified for the database.
  • A database schema must be initiated, using the function mnesia:create_schema/1.

The following example shows how these tasks are performed:

Step 1: Start an Erlang session and specify a Mnesia directory for the database:

% erl -sname klacke -mnesia dir '"/ldisc/scratch/klacke"'
Erlang (BEAM) emulator version 4.9
 
Eshell V4.9  (abort with ^G)
(klacke@gin)1> mnesia:create_schema([node()]).
ok
(klacke@gin)2> 
^Z
Suspended

Step 2: You can inspect the Mnesia directory to see what files have been created:

% ls -l /ldisc/scratch/klacke
-rw-rw-r--   1 klacke   staff       247 Aug 12 15:06 FALLBACK.BUP

The response shows that the file FALLBACK.BUP has been created. This is called a backup file, and it contains an initial schema. If more than one node in the function mnesia:create_schema/1 had been specified, identical backup files would have been created on all nodes.

Step 3: Start Mnesia:

(klacke@gin)3>mnesia:start( ).
ok

Step 4: You can see the following listing in the Mnesia directory:

-rw-rw-r--   1 klacke   staff         86 May 26 19:03 LATEST.LOG
-rw-rw-r--   1 klacke   staff      34507 May 26 19:03 schema.DAT

The schema in the backup file FALLBACK.BUP has been used to generate the file schema.DAT. Since there are no other disc resident tables than the schema, no other data files were created. The file FALLBACK.BUP was removed after the successful "restoration". You also see some files that are for internal use by Mnesia.

Step 5: Create a table:

(klacke@gin)4> mnesia:create_table(foo,[{disc_copies, [node()]}]).
{atomic,ok}

Step 6: You can see the following listing in the Mnesia directory:

% ls -l /ldisc/scratch/klacke
-rw-rw-r-- 1 klacke staff    86 May 26 19:07 LATEST.LOG
-rw-rw-r-- 1 klacke staff    94 May 26 19:07 foo.DCD
-rw-rw-r-- 1 klacke staff  6679 May 26 19:07 schema.DAT

The file foo.DCD has been created. This file will eventually store all data that is written into the foo table.

When starting Mnesia, a .LOG file called LATEST.LOG is created and placed in the database directory. This file is used by Mnesia to log disc-based transactions. This includes all transactions that write at least one record in a table that is of storage type disc_copies or disc_only_copies. The file also includes all operations that manipulate the schema itself, such as creating new tables. The log format can vary with different implementations of Mnesia. The Mnesia log is currently implemented in the standard library module disk_log in Kernel.

The log file grows continuously and must be dumped at regular intervals. "Dumping the log file" means that Mnesia performs all the operations listed in the log and place the records in the corresponding .DAT, .DCD, and .DCL data files. For example, if the operation "write record {foo, 4, elvis, 6}" is listed in the log, Mnesia inserts the operation into the file foo.DCL. Later, when Mnesia thinks that the .DCL file is too large, the data is moved to the .DCD file. The dumping operation can be time consuming if the log is large. Notice that the Mnesia system continues to operate during log dumps.

By default Mnesia either dumps the log whenever 1000 records have been written in the log or when three minutes have passed. This is controlled by the two application parameters -mnesia dump_log_write_threshold WriteOperations and -mnesia dump_log_time_threshold MilliSecs.

Before the log is dumped, the file LATEST.LOG is renamed to PREVIOUS.LOG, and a new LATEST.LOG file is created. Once the log has been successfully dumped, the file PREVIOUS.LOG is deleted.

The log is also dumped at startup and whenever a schema operation is performed.

The directory listing also contains one .DAT file, which contains the schema itself, contained in the schema.DAT file. The DAT files are indexed files, and it is efficient to insert and search for records in these files with a specific key. The .DAT files are used for the schema and for disc_only_copies tables. The Mnesia data files are currently implemented in the standard library module dets in STDLIB.

All operations that can be performed on dets files can also be performed on the Mnesia data files. For example, dets contains the function dets:traverse/2, which can be used to view the contents of a Mnesia DAT file. However, this can only be done when Mnesia is not running. So, to view the schema file, do as follows;

{ok, N} = dets:open_file(schema, [{file, "./schema.DAT"},{repair,false}, 
{keypos, 2}]),
F = fun(X) -> io:format("~p~n", [X]), continue end,
dets:traverse(N, F),
dets:close(N).
Warning

The DAT files must always be opened with option {repair, false}. This ensures that these files are not automatically repaired. Without this option, the database can become inconsistent, because Mnesia can believe that the files were properly closed. For information about configuration parameter auto_repair, see the Reference Manual.

Warning

It is recommended that the data files are not tampered with while Mnesia is running. While not prohibited, the behavior of Mnesia is unpredictable.

The disc_copies tables are stored on disk with .DCL and .DCD files, which are standard disk_log files.

At startup, Mnesia loads tables to make them accessible for its applications. Sometimes Mnesia decides to load all tables that reside locally, and sometimes the tables are not accessible until Mnesia brings a copy of the table from another node.

To understand the behavior of Mnesia at startup, it is essential to understand how Mnesia reacts when it loses contact with Mnesia on another node. At this stage, Mnesia cannot distinguish between a communication failure and a "normal" node-down. When this occurs, Mnesia assumes that the other node is no longer running, whereas, in reality, the communication between the nodes has failed.

To overcome this situation, try to restart the ongoing transactions that are accessing tables on the failing node, and write a mnesia_down entry to a log file.

At startup, notice that all tables residing on nodes without a mnesia_down entry can have fresher replicas. Their replicas can have been updated after the termination of Mnesia on the current node. To catch up with the latest updates, transfer a copy of the table from one of these other "fresh" nodes. If you are unlucky, other nodes can be down and you must wait for the table to be loaded on one of these nodes before receiving a fresh copy of the table.

Before an application makes its first access to a table, mnesia:wait_for_tables(TabList, Timeout) is to be executed to ensure that the table is accessible from the local node. If the function times out, the application can choose to force a load of the local replica with mnesia:force_load_table(Tab) and deliberately lose all updates that can have been performed on the other nodes while the local node was down. If Mnesia has loaded the table on another node already, or intends to do so, copy the table from that node to avoid unnecessary inconsistency.

Warning

Only one table is loaded by mnesia:force_load_table(Tab). Since committed transactions can have caused updates in several tables, the tables can become inconsistent because of the forced load.

The allowed AccessMode of a table can be defined to be read_only or read_write. It can be toggled with the function mnesia:change_table_access_mode(Tab, AccessMode) in runtime. read_only tables and local_content tables are always loaded locally, as there is no need for copying the table from other nodes. Other tables are primarily loaded remotely from active replicas on other nodes if the table has been loaded there already, or if the running Mnesia has decided to load the table there already.

At startup, Mnesia assumes that its local replica is the most recent version and loads the table from disc if either of the following situations is detected:

  • mnesia_down is returned from all other nodes that hold a disc resident replica of the table.
  • All replicas are ram_copies.

This is normally a wise decision, but it can be disastrous if the nodes have been disconnected because of a communication failure, as the Mnesia normal table load mechanism does not cope with communication failures.

When Mnesia loads many tables, the default load order is used. However, the load order can be affected, by explicitly changing property load_order for the tables, with the function mnesia:change_table_load_order(Tab, LoadOrder). LoadOrder is by default 0 for all tables, but it can be set to any integer. The table with the highest load_order is loaded first. Changing the load order is especially useful for applications that need to ensure early availability of fundamental tables. Large peripheral tables are to have a low load order value, perhaps less than 0

There are several occasions when Mnesia can detect that the network has been partitioned because of a communication failure, for example:

  • Mnesia is operational already and the Erlang nodes gain contact again. Then Mnesia tries to contact Mnesia on the other node to see if it also thinks that the network has been partitioned for a while. If Mnesia on both nodes has logged mnesia_down entries from each other, Mnesia generates a system event, called {inconsistent_database, running_partitioned_network, Node}, which is sent to the Mnesia event handler and other possible subscribers. The default event handler reports an error to the error logger.
  • If Mnesia detects at startup that both the local node and another node received mnesia_down from each other, Mnesia generates an {inconsistent_database, starting_partitioned_network, Node} system event and acts as described in the previous item.

If the application detects that there has been a communication failure that can have caused an inconsistent database, it can use the function mnesia:set_master_nodes(Tab, Nodes) to pinpoint from which nodes each table can be loaded.

At startup, the Mnesia normal table load algorithm is bypassed and the table is loaded from one of the master nodes defined for the table, regardless of potential mnesia_down entries in the log. Nodes can only contain nodes where the table has a replica. If Nodes is empty, the master node recovery mechanism for the particular table is reset and the normal load mechanism is used at the next restart.

The function mnesia:set_master_nodes(Nodes) sets master nodes for all tables. For each table it determines its replica nodes and starts mnesia:set_master_nodes(Tab, TabNodes) with those replica nodes that are included in the Nodes list (that is, TabNodes is the intersection of Nodes and the replica nodes of the table). If the intersection is empty, the master node recovery mechanism for the particular table is reset and the normal load mechanism is used at the next restart.

The functions mnesia:system_info(master_node_tables) and mnesia:table_info(Tab, master_nodes) can be used to obtain information about the potential master nodes.

Determining what data to keep after a communication failure is outside the scope of Mnesia. One approach is to determine which "island" contains most of the nodes. Using option {majority,true} for critical tables can be a way to ensure that nodes that are not part of a "majority island" cannot update those tables. Notice that this constitutes a reduction in service on the minority nodes. This would be a tradeoff in favor of higher consistency guarantees.

The function mnesia:force_load_table(Tab) can be used to force load the table regardless of which table load mechanism that is activated.

A Mnesia table can reside on one or more nodes. When a table is updated, Mnesia ensures that the updates are replicated to all nodes where the table resides. If a replica is inaccessible (for example, because of a temporary node-down), Mnesia performs the replication later.

On the node where the application is started, there is a transaction coordinator process. If the transaction is distributed, there is also a transaction participant process on all the other nodes where commit-work needs to be performed.

Internally Mnesia uses several commit protocols. The selected protocol depends on which table that has been updated in the transaction. If all the involved tables are symmetrically replicated (that is, they all have the same ram_nodes, disc_nodes, and disc_only_nodes currently accessible from the coordinator node), a lightweight transaction commit protocol is used.

The number of messages that the transaction coordinator and its participants need to exchange is few, as the Mnesia table load mechanism takes care of the transaction recovery if the commit protocol gets interrupted. Since all involved tables are replicated symmetrically, the transaction is automatically recovered by loading the involved tables from the same node at startup of a failing node. It does not matter if the transaction was committed or terminated as long as the ACID properties can be ensured. The lightweight commit protocol is non-blocking, that is, the surviving participants and their coordinator finish the transaction, even if any node crashes in the middle of the commit protocol.

If a node goes down in the middle of a dirty operation, the table load mechanism ensures that the update is performed on all replicas, or none. Both asynchronous dirty updates and synchronous dirty updates use the same recovery principle as lightweight transactions.

If a transaction involves updates of asymmetrically replicated tables or updates of the schema table, a heavyweight commit protocol is used. This protocol can finish the transaction regardless of how the tables are replicated. The typical use of a heavyweight transaction is when a replica is to be moved from one node to another. Then ensure that the replica either is entirely moved or left as it was. Do never end up in a situation with replicas on both nodes, or on no node at all. Even if a node crashes in the middle of the commit protocol, the transaction must be guaranteed to be atomic. The heavyweight commit protocol involves more messages between the transaction coordinator and its participants than a lightweight protocol, and it performs recovery work at startup to finish the terminating or commit work.

The heavyweight commit protocol is also non-blocking, which allows the surviving participants and their coordinator to finish the transaction regardless (even if a node crashes in the middle of the commit protocol). When a node fails at startup, Mnesia determines the outcome of the transaction and recovers it. Lightweight protocols, heavyweight protocols, and dirty updates, are dependent on other nodes to be operational to make the correct heavyweight transaction recovery decision.

If Mnesia has not started on some of the nodes that are involved in the transaction and neither the local node nor any of the already running nodes know the outcome of the transaction, Mnesia waits for one, by default. In the worst case scenario, all other involved nodes must start before Mnesia can make the correct decision about the transaction and finish its startup.

Thus, Mnesia (on one node) can hang if a double fault occurs, that is, when two nodes crash simultaneously and one attempts to start when the other refuses to start, for example, because of a hardware error.

The maximum time that Mnesia waits for other nodes to respond with a transaction recovery decision can be specified. The configuration parameter max_wait_for_decision defaults to infinity, which can cause the indefinite hanging as mentioned earlier. However, if the parameter is set to a definite time period (for example, three minutes), Mnesia then enforces a transaction recovery decision, if needed, to allow Mnesia to continue with its startup procedure.

The downside of an enforced transaction recovery decision is that the decision can be incorrect, because of insufficient information about the recovery decisions from the other nodes. This can result in an inconsistent database where Mnesia has committed the transaction on some nodes but terminated it on others.

In fortunate cases, the inconsistency is only visible in tables belonging to a specific application. However, if a schema transaction is inconsistently recovered because of the enforced transaction recovery decision, the effects of the inconsistency can be fatal. However, if the higher priority is availability rather than consistency, it can be worth the risk.

If Mnesia detects an inconsistent transaction decision, an {inconsistent_database, bad_decision, Node} system event is generated to give the application a chance to install a fallback or other appropriate measures to resolve the inconsistency. The default behavior of the Mnesia event handler is the same as if the database became inconsistent as a result of partitioned network (as described earlier).

The following functions are used to back up data, to install a backup as fallback, and for disaster recovery:

These functions are explained in the following sections. See also Checkpoints, which describes the two functions used to activate and deactivate checkpoints.

Backup operation are performed with the following functions:

By default, the actual access to the backup media is performed through module mnesia_backup for both read and write. Currently mnesia_backup is implemented with the standard library module disc_log. However, you can write your own module with the same interface as mnesia_backup and configure Mnesia so that the alternative module performs the actual accesses to the backup media. The user can therefore put the backup on a media that Mnesia does not know about, possibly on hosts where Erlang is not running. Use configuration parameter -mnesia backup_module <module> for this purpose.

The source for a backup is an activated checkpoint. The backup function mnesia:backup_checkpoint(Name, Opaque,[Mod]) is most commonly used and returns ok or {error,Reason}. It has the following arguments:

  • Name is the name of an activated checkpoint. For details on how to include table names in checkpoints, see the function mnesia:activate_checkpoint(ArgList) in Checkpoints.
  • Opaque. Mnesia does not interpret this argument, but it is forwarded to the backup module. The Mnesia default backup module mnesia_backup interprets this argument as a local filename.
  • Mod is the name of an alternative backup module.

The function mnesia:backup(Opaque [,Mod]) activates a new checkpoint that covers all Mnesia tables with maximum degree of redundancy and performs a backup. Maximum redundancy means that each table replica has a checkpoint retainer. Tables with property local_contents are backed up as they look on the current node.

You can iterate over a backup, either to transform it into a new backup, or only read it. The function mnesia:traverse_backup(Source, [SourceMod,] Target, [TargetMod,] Fun, Acc), which normally returns {ok, LastAcc}, is used for both of these purposes.

Before the traversal starts, the source backup media is opened with SourceMod:open_read(Source), and the target backup media is opened with TargetMod:open_write(Target). The arguments are as follows:

  • SourceMod and TargetMod are module names.
  • Source and Target are opaque data used exclusively by the modules SourceMod and TargetMod for initializing the backup media.
  • Acc is an initial accumulator value.
  • Fun(BackupItems, Acc) is applied to each item in the backup. The Fun must return a tuple {ValGoodBackupItems, NewAcc}, where ValidBackupItems is a list of valid backup items. NewAcc is a new accumulator value. The ValidBackupItems are written to the target backup with the function TargetMod:write/2.
  • LastAcc is the last accumulator value, that is, the last NewAcc value that was returned by Fun.

Also, a read-only traversal of the source backup can be performed without updating a target backup. If TargetMod==read_only, no target backup is accessed.

By setting SourceMod and TargetMod to different modules, a backup can be copied from one backup media to another.

Valid BackupItems are the following tuples:

  • {schema, Tab} specifies a table to be deleted.
  • {schema, Tab, CreateList} specifies a table to be created. For more information about CreateList, see mnesia:create_table/2.
  • {Tab, Key} specifies the full identity of a record to be deleted.
  • {Record} specifies a record to be inserted. It can be a tuple with Tab as first field. Notice that the record name is set to the table name regardless of what record_name is set to.

The backup data is divided into two sections. The first section contains information related to the schema. All schema-related items are tuples where the first field equals the atom schema. The second section is the record section. Schema records cannot be mixed with other records and all schema records must be located first in the backup.

The schema itself is a table and is possibly included in the backup. Each node where the schema table resides is regarded as a db_node.

The following example shows how mnesia:traverse_backup can be used to rename a db_node in a backup file:

change_node_name(Mod, From, To, Source, Target) ->
    Switch =
        fun(Node) when Node == From -> To;
           (Node) when Node == To -> throw({error, already_exists});
           (Node) -> Node
        end,
    Convert =
        fun({schema, db_nodes, Nodes}, Acc) ->
                {[{schema, db_nodes, lists:map(Switch,Nodes)}], Acc};
           ({schema, version, Version}, Acc) ->
                {[{schema, version, Version}], Acc};
           ({schema, cookie, Cookie}, Acc) ->
                {[{schema, cookie, Cookie}], Acc};
           ({schema, Tab, CreateList}, Acc) ->
                Keys = [ram_copies, disc_copies, disc_only_copies],
                OptSwitch =
                    fun({Key, Val}) ->
                            case lists:member(Key, Keys) of
                                true -> {Key, lists:map(Switch, Val)};
                                false-> {Key, Val}
                            end
                    end,
                {[{schema, Tab, lists:map(OptSwitch, CreateList)}], Acc};
           (Other, Acc) ->
                {[Other], Acc}
        end,
    mnesia:traverse_backup(Source, Mod, Target, Mod, Convert, switched).

view(Source, Mod) ->
    View = fun(Item, Acc) ->
                   io:format("~p.~n",[Item]),
                   {[Item], Acc + 1}
           end,
    mnesia:traverse_backup(Source, Mod, dummy, read_only, View, 0).

Tables can be restored online from a backup without restarting Mnesia. A restore is performed with the function mnesia:restore(Opaque, Args), where Args can contain the following tuples:

  • {module,Mod}. The backup module Mod is used to access the backup media. If omitted, the default backup module is used.
  • {skip_tables, TableList}, where TableList is a list of tables, which is not to be read from the backup.
  • {clear_tables, TableList}, where TableList is a list of tables, which is to be cleared before the records from the backup are inserted. That is, all records in the tables are deleted before the tables are restored. Schema information about the tables is not cleared or read from the backup.
  • {keep_tables, TableList}, where TableList is a list of tables, which is not to be cleared before the records from the backup are inserted. That is, the records in the backup are added to the records in the table. Schema information about the tables is not cleared or read from the backup.
  • {recreate_tables, TableList}, where TableList is a list of tables, which is to be recreated before the records from the backup are inserted. The tables are first deleted and then created with the schema information from the backup. All the nodes in the backup need to be operational.
  • {default_op, Operation}, where Operation is one of the operations skip_tables, clear_tables, keep_tables, or recreate_tables. The default operation specifies which operation is to be used on tables from the backup that are not specified in any of the previous lists. If omitted, the operation clear_tables is used.

The argument Opaque is forwarded to the backup module. It returns {atomic, TabList} if successful, or the tuple {aborted, Reason} if there is an error. TabList is a list of the restored tables. Tables that are restored are write-locked during the restore operation. However, regardless of any lock conflict caused by this, applications can continue to do their work during the restore operation.

The restoration is performed as a single transaction. If the database is large, it cannot always be restored online. The old database must then be restored by installing a fallback, followed by a restart.

The function mnesia:install_fallback(Opaque, [Mod]) installs a backup as fallback. It uses the backup module Mod, or the default backup module, to access the backup media. The function returns ok if successful, or {error, Reason} if there is an error.

Installing a fallback is a distributed operation, which is only performed on all db_nodes. The fallback restores the database the next time the system is started. If a Mnesia node with a fallback installed detects that Mnesia on another node has died, it unconditionally terminates itself.

A fallback is typically used when a system upgrade is performed. A system typically involves the installation of new software versions, and Mnesia tables are often transformed into new layouts. If the system crashes during an upgrade, it is highly probable that reinstallation of the old applications is required, and restoration of the database to its previous state. This can be done if a backup is performed and installed as a fallback before the system upgrade begins.

If the system upgrade fails, Mnesia must be restarted on all db_nodes to restore the old database. The fallback is automatically deinstalled after a successful startup. The function mnesia:uninstall_fallback() can also be used to deinstall the fallback after a successful system upgrade. Again, this is a distributed operation that is either performed on all db_nodes or none. Both the installation and deinstallation of fallbacks require Erlang to be operational on all db_nodes, but it does not matter if Mnesia is running or not.

The system can become inconsistent as a result of a power failure. The UNIX feature fsck can possibly repair the file system, but there is no guarantee that the file content is consistent.

If Mnesia detects that a file has not been properly closed, possibly as a result of a power failure, it tries to repair the bad file in a similar manner. Data can be lost, but Mnesia can be restarted even if the data is inconsistent. Configuration parameter -mnesia auto_repair <bool> can be used to control the behavior of Mnesia at startup. If <bool> has the value true, Mnesia tries to repair the file. If <bool> has the value false, Mnesia does not restart if it detects a suspect file. This configuration parameter affects the repair behavior of log files, DAT files, and the default backup media.

Configuration parameter -mnesia dump_log_update_in_place <bool> controls the safety level of the function mnesia:dump_log() By default, Mnesia dumps the transaction log directly into the DAT files. If a power failure occurs during the dump, this can cause the randomly accessed DAT files to become corrupt. If the parameter is set to false, Mnesia copies the DAT files and target the dump to the new temporary files. If the dump is successful, the temporary files are renamed to their normal DAT suffixes. The possibility for unrecoverable inconsistencies in the data files becomes much smaller with this strategy. However, the actual dumping of the transaction log becomes considerably slower. The system designer must decide whether speed or safety is the higher priority.

Replicas of type disc_only_copies are only affected by this parameter during the initial dump of the log file at startup. When designing applications with very high requirements, it can be appropriate not to use disc_only_copies tables at all. The reason for this is the random access nature of normal operating system files. If a node goes down for a reason such as a power failure, these files can be corrupted because they are not properly closed. The DAT files for disc_only_copies are updated on a per transaction basis.

If a disaster occurs and the Mnesia database is corrupted, it can be reconstructed from a backup. Regard this as a last resort, as the backup contains old data. The data is hopefully consistent, but data is definitely lost when an old backup is used to restore the database.