++ed by:

1 PAUSE user

Author image Torsten Förtsch


MMapDB - a simple database in shared memory


  use MMapDB qw/:error/;

  # create a database
  my $db=MMapDB->new(filename=>$path, intfmt=>'J');
  $db->start;       # check if the database exists and connect
  $db->begin;       # begin a transaction

  # insert something
  ($id, $pos)=$db->insert([[qw/main_key subkey .../],
                           $sort, $data]);
  # or delete
  $just_deleted=$db->delete_by_id($id, 1);

  # or forget everything

  # make changes visible

  # or forget the transaction

  # use a database
  my $db=MMapDB->new(filename=>$path);

  # tied interface
  ($keys, $sort, $data, $id)=@{$db->main_index->{main_key}->{subkey}};
  @subkeys=keys %$subindex;
  @mainkeys=keys %{$db->main_index};

  # or even
  use Data::Dumper;
  print Dumper($db->main_index); # dumps the whole database

  print Dumper($db->main_index); # dumps only values

  # access by ID
  ($keys, $sort, $data, $id)=@{$db->id_index->{$id}};

  # fast access
  @positions=$db->index_lookup($db->mainidx, $key);
  if( @positions==1 and $positions[0] >= $db->mainidx ) {
    # found another index
    @positions=$db->index_lookup($positions[0], ...);
  } elsif(@positions) {
    # found a data record
    for (@positions) {
      ($keys, $sort, $data, $id)=@{$db->data_record($_)};




  } else {
    # not found

  # access by ID
  ($keys, $sort, $data, $id)=@{$db->data_record($position)};

  # iterate over all valid data records
  for( $it=$db->iterator; $pos=$it->(); ) {
    ($keys, $sort, $data, $id)=@{$db->data_record($pos)};

  # or all invalid data records
  for( $it=$db->iterator(1); $pos=$it->(); ) {
    ($keys, $sort, $data, $id)=@{$db->data_record($pos)};

  # iterate over an index
  for( $it=$db->index_iterator($db->mainidx);
       ($partkey, @positions)=$it->(); ) {

  # and over the ID index
  for( $it=$db->id_index_iterator;
       ($id, $position)=$it->(); ) {

  # disconnect from a database


MMapDB implements a database similar to a hash of hashes of hashes, ..., of arrays of data.

It's main design goals were:

  • very fast read access

  • lock-free read access for massive parallelism

  • minimal memory consumption per accessing process

  • transaction based write access

  • simple backup, compactness, one file

The cost of write access was unimportant and the expected database size was a few thousands to a few hundreds of thousands data records.

Hence come 2 major decisions. Firstly, the database is completely mapped into each process's address space. And secondly, a transaction writes the complete database anew.

Still interested?


The data record

A data record consists of 3-4 fields:

 [[KEY1, KEY2, ..., KEYn], ORDER, DATA, ID]   # ID is optional

All of the KEY1, ..., KEYn, SORT and DATA are arbitrary length octet strings. The key itself is an array of strings showing the way to the data item. The word key in the rest of this text refers to such an array of strings.

Multiple data records can be stored under the same key. So, there is perhaps an less-greater relationship between the data records. That's why there is the ORDER field. If the order field of 2 or more data records are equal (eq, not ==), their order is defined by the stability of perl's sort operation. New data records are always appended to the set of records. So, if sort is stable they appear at the end of a range of records with the same ORDER.

The DATA field contains the data itself.

A data record in the database further owns an ID. The ID uniquely identifies the data record. It is assigned when the record is inserted.

An ID is a fixed size number (32 or 64 bits) except 0. They are allocated from 1 upwards. When the upper boundary is reached the next ID becomes 1 if it is not currently used.

The index record

An index record consists of 2 or more fields:

 (PARTKEY, POS1, POS2, ..., POSn)

PARTKEY is one of the KEYi that form the key of a data record. The POSi are positions in the database file that point to other indices or data records. Think of such a record as an array of data records or the leafes in the hash of hashes tree.

If an index record contains more than 1 positions they must all point to data records. This way an ordered array of data records is formed. The position order is determined by the ORDER fields of the data records involved.

If the index record contains only one position it can point to another index or a data record. The distiction is made by the so called main index position. If the position is lower it points to a data record otherwise to an index.

The index

An index is a list of index records ordered by their PARTKEY field. Think of an index as a hash or a twig in the hash of hashes tree. When a key is looked up a binary search in the index is performed.

There are 2 special indices, the main index and the ID index. The positions of both of them are part of the database header. This fact is the only thing that makes the main index special. The ID index is special also because it's keys are IDs rather than strings.

The hash of hashes

To summarize all of the above, the following data structure displays the logical structure of the database:


What cannot be expressed is an index record containing a pointer to a data record and to another subindex, somthing like this:

        {                     # INVALID

Note also, the root element is always a hash. The following is invalid:


Accessing a database

To use a database it must be connected. Once connected a database is readonly. You will always read the same values regardless of other transactions that may have written the database.

The database header contains a flag that says if it is still valid or has been replaced by a transaction. There is a method that checks this flag and reconnects to the new version if necessary. The points when to call this method depend on your application logic.

Accessing data

To access data by a key first thing the main index position is read from the database header. Then a binary search is performed to look up the index record that matches the first partial key. If it could be found and there is no more partial key the position list is returned. If there is another partial key but the position list points to data records the key is not found. If the position list contains another index position and there is another partial key the lookup is repeated on this index with the next partial key until either all partial keys have been found or one of them could not be found.

A method is provided to read a single data record by its position.

The index lookup is written in C for speed. All other parts are perl.


When a transaction is started a new private copy of the database is created. Then new records can be inserted or existing ones deleted. While a transaction is active it is not possible to reconnect to the database. This ensures that the transaction derives only from the data that was valid when the transaction has begun.

Newly written data cannot be read back within the transaction. It becomes visible only when the transaction is committed. So, one will always read the state from the beginning of the transaction.

When a transaction is committed the public version is replaced by the private one. Thus, the old database is deleted. But other processes including the one performing the commit still have the old version still mapped in their address space. At this point in time new processes will see the new version while existing see the old data. Therefore a flag in the old memory mapped database is written signaling that is has become invalid. The atomicity of this operation depends on the atomicity of perls rename operation. On Linux it is atomic in a sense that there is no point in time when a new process won't find the database file.

A lock file can be used to serialize transactions.

The integer format

All numbers and offsets within the database are written in a certain format (byte order plus size). When a database is created the format is written to the database header and cannot be changed afterwards.

Perl's pack command defines several ways of encoding integer numbers. The database format is defined by one of these pack format letters:

  • L

    32 bit in native byte order.

  • N

    32 bit in big endian order. This format should be portable.

  • J

    32 or 64 bit in native byte order.

  • Q

    64 bit in native byte order. But is not always implemented.


Some MMapDB methods return iterators. An iterator is a blessed function reference (also called closure) that if called returns one item at a time. When there are no more items an empty list or undef is returned.

If you haven't heard of this concept yet perl itself has an iterator attached to each hash. The each operation returns a key/value pair until there is no more.

Iterators are mainly used this way:

  $it=$db->...;       # create an iterator
  while( @item=$it->() ) {
    # use @item

As mentioned, iterators are also objects. They have methods to fetch the number of elements they will traverse or to position the iterator to a certain element.

Error Conditions

Some programmers believe exceptions are the right method of error reporting. Other think it is the return value of a function or method.

This modules somehow divides errors in 2 categories. A key that could not be found for example is a normal result not an error. However, a disk full condition or insufficient permissions to create a file are errors. This kind of errors are thrown as exceptions. Normal results are returned through the return value.

If an exception is thrown in this module $@ will contain a scalar reference. There are several errors defined that can be imported into the using program:

  use MMapDB qw/:error/;  # or :all

The presence of an error is checked this way (after eval):


Human readable error messages are provided by the scalar $@ points to:

  warn ${$@}

List of error conditions


    database is read-only


    attempt to insert the same ID twice


    attempt to begin a transaction while there is one active

  • E_FULL

    no more IDs


    data records cannot be mixed up with subindices

  • E_OPEN

    can't open file

  • E_READ

    can't read from file


    can't write to file


    file could not be closed


    can't rename file

  • E_SEEK

    can't move file pointer


    can't truncate file

  • E_LOCK

    can't lock or unlock


    attempt move an iterator position out of its range


    function not implemented


As of version 0.07 MMapDB supports UTF8 data. What does that mean?

For each string perl maintains a flag that says whether the string is stored in UTF8 or not. MMapDB stores and retrieves this flag along with the string itself.

For example take the string "\320\263\321\200\321\203\321\210\320\260". With the UTF8 flag unset it is just a bunch of octets. But if the flag is set the string suddenly becomes these characters "\x{433}\x{440}\x{443}\x{448}\x{430}" or груша which means pear in Russian.

Note, with the flag unset the length of our string is 10 which is the number of octets. But if the flag is set it is 5 which is the number of characters.

So, although the octet sequences representing both strings (with and without the flag) equal the strings themselves differ.

With MMapDB, if something is stored under an UTF8 key it can be retrieved also only by the UTF8 key.

There is a subtle point here, what if the UTF8 flag is set for a string consisting of ASCII-only characters? In Perl th following expression is true:

 'hello' eq Encode::decode_utf8('hello')

The first hello is a string the the UTF8 flag unset. It is compared with the same sequence of octets with the UTF8 flag set.

MMapDB up to version 0.11 (including) considered those 2 strings different. That bug has been fixed in version 0.12.

The database format

With version 0.07 UTF8 support was introduced into MMapDB. This required a slight change of the database on-disk format. Now the database format is encoded in the magic number of the database file. By now there are 2 formats defined:

  • Format 0 --> magic number: MMDB

  • Format 1 --> magic number: MMDC

MMapDB reads and writes both formats. By default new databases are written in the most up-to-date format and the format of existing ones is not changed.

If you want to write a certain format for a new database or convert an existing database to an other format specify it as parameter to the begin() method.

If you want to write the newest format use -1 as format specifier.


$dh=MMapDB->new( KEY=>VALUE, ... )

$new=$db->new( KEY=>VALUE, ... )

$dh=MMapDB->new( $filename )

$new=$db->new( $filename )

creates or clones a database handle. If there is an active transaction it is rolled back for the clone.

If only one parameter is passed it is taken as the database filename.

Otherwise parameters are passed as (KEY,VALUE) pairs:

  • filename

    specifies the database file

  • lockfile

    specify a lockfile to serialize transactions. If given at start of a transaction the file is locked using flock and unlocked at commit or rollback time. The lockfile is empty but must not be deleted. Best if it is created before first use.

    If lockfile ommitted MMapDB continues to work but it is possible to begin a new transaction while another one is active in another process.

  • intfmt

    specifies the integer format. Valid values are N (the default), L, J and Q (not always implemented). When opening an existing database this value is overwritten by the database format.

  • readonly

    if passed a true value the database file is mapped read-only. That means the accessing process will receive a SEGV signal (on UNIX) if some stray pointer wants to write to this area. At perl level the variable is marked read-only. So any read access at perl level will not generate a segmentation fault but instead a perl exception.


sets the integer format of an existing database handle. Do not call this while the handle is connected to a database.




those are accessors for the attributes passed to the constructor. They can also be assigned to ($db->filename=$new_name) but only before a database is connected to. intfmt must be set using set_intfmt().


(re)connects to the database. This method maps the database file and checks if it is still valid. This method cannot be called while a transaction is active.

If the current database is still valid or the database has been successfully connected or reconnected the database object is returned. undef otherwise.

There are several conditions that make start fail:

  • $db->filename could not be opened

  • the file could not be mapped into the address space

  • the file is empty

  • the magic number of the file does not match

  • the integer format indentifier is not valid


disconnects from a database.



begins a transaction. Returns the database object.

If the $dbformat parameter is ommitted the database format is unchanged.

If 0 is given the database is written in MMDB format. If 1 is given MMDC format is written.

If -1 is given always the newest format is written.



When a transaction is begun 2 temporary files are created, one for the index and data records the other for the string table. These files are then mapped into the process' address space for faster access.

index_prealloc and stringmap_prealloc define the initial sizes of these files. The files are created as sparse files. The actual space is allocated on demand. By default both these values are set to 10*1024*1024 which is ten megabytes.

When the space in one of the areas becomes insufficient it is extented and remapped in chunks of index_prealloc and stringmap_prealloc respectively.

You should change the default values only for really big databases:




commits a transaction and reconnects to the new version.

Returns the database object it the new version has been connected.

If a true value is passed to this message the old database version is not invalidated. This makes it possible to safely create a copy for backup this way:

  my $db=MMapDB->new(filename=>FILENAME);


  my $db=MMapDB->new(filename=>FILENAME);
  my $backup=$db->new(filename=>BACKUP);


forgets about the current transaction


returns true if a database is connected an valid.


invalidates the current version of the database. This is normally called internally by commit as the last step. Perhaps there are also other uses.


creates a backup of the current database version called BACKUPNAME. It works almost exactly as shown in commit. Note, backup calls $db->start.

If the filename parameter is ommitted the result of appending the .BACKUP extension to the object's filename property is used.


just the opposite of backup. It renames BACKUPNAME to $db->filename and invalidates the current version. So, the backup becomes the current version. For other processes running in parallel this looks just like another transaction being committed.



when a database is newly created it is written in the format passed to the begin() method. When it is connected later via start() dbformat_in() contains this format. So, a database user can check if a database file meets his expectations.

Within a transaction dbformat_out() contains the format of the database currently being written.


contains a number in the range between 0 and 255 that represents the flags byte of the database.

It is written to the database file by the begin() method and read by start().

Using this field is not recommended. It is considered experimental and may disappear in the future.

@positions=$db->index_lookup(INDEXPOS, KEY1, KEY2, ...)

looks up the key [KEY1, KEY2, ...] in the index given by its position INDEXPOS. Returns a list of positions or an empty list if the complete key is not found.

If INDEXPOS is 0 or undef $db->mainidx is used.

To check if the result is a data record array or another index use this code:

  if( @positions==1 and $positions[0] >= $db->mainidx ) {
    # found another index
    # the position can be passed to another index_lookup()
  } elsif(@positions) {
    # found a data record
    # the positions can be passed to data_record()
  } else {
    # not found

($idxpos, $nth)=$db->index_lookup_position(INDEXPOS, KEY1, ...)

This method behaves similar to index_lookup() in that it looks up a key. It differs in

  • It returns the position of the index containing the last key element and the position of the found element within that index. These 2 numbers can be used to create and position an iterator.

  • The last key element may not exist. In this case the value returned as $nth points to the index element before which the key would appear.

  • If an intermediate key element does not exist or is not an index an empty list is returned.

Consider the following database:

   key => {
            aa => ['1'],
            ab => ['2'],
            ad => ['3'],

Now let's define an accessor function:

 sub get {
   my ($subkey)=@_;

         ($db->index_lookup_position($db->mainidx, "key", $subkey))->()


 get "aa";    # returns '1'
 get "ab";    # returns '2'
 get "ac";    # returns '3' although key "ac" does not exist it would
              #             show up between "ab" and "ad". So, the iterator
              #             is positioned on "ad"
 get "aba";   # returns '3' for the same reason
 get "ad";    # returns '3'

 (undef, $nth)=$db->index_lookup_position($db->mainidx, "key", "az");
 # now $nth is 3 because "az" would be inserted after "ad" which is at
 # position 2 within the index.

If INDEXPOS is 0 or undef $db->mainidx is used.


looks up a data record by its ID. Returns the data record's position.


returns true if $pos points to the data record space.

This method is simply a shortcut for

  $pos < $db->mainidx

A true result does not mean it is safe to use $pos in data_record().

@recs=$db->data_record(POS, ...)

given a list of positions fetches the data records. Each @recs element is an array reference with the following structure:

  [[KEY1, KEY2, ...], SORT, DATA, ID]

All of the KEYs, SORT and DATA are read-only strings.

@values=$db->data_value(POS, ...)

similar to data_record() but returns only the DATA fields of the records. Faster than data_record().

See "The MMapDB::Data class" for an example.

@sorts=$db->data_sort(POS, ...)

similar to data_record() but returns only the SORT fields of the records. Faster than data_record().

@records=$db->index_lookup_records(INDEXPOS, KEY1, KEY2, ...)

a more effective shortcut of

 @records=$db->data_record($db->index_lookup((INDEXPOS, KEY1, KEY2, ...)))

@values=$db->index_lookup_values(INDEXPOS, KEY1, KEY2, ...)

a more effective shortcut of

 @records=$db->data_value($db->index_lookup((INDEXPOS, KEY1, KEY2, ...)))

@sorts=$db->index_lookup_sorts(INDEXPOS, KEY1, KEY2, ...)

a more effective shortcut of

 @records=$db->data_sort($db->index_lookup((INDEXPOS, KEY1, KEY2, ...)))



perhaps you want to iterate over all data records in a database. The iterator returns a data record position:


If a true value is passed as parameter only deleted records are found otherwise only valid ones.

$it is an MMapDB::Iterator object. Invoking any method on this iterator results in a E_NOT_IMPLEMENTED exception.

$it=$db->index_iterator(POS, NTH)

($it, $nitems)=$db->index_iterator(POS, NTH)

iterate over an index given by its position. The iterator returns a partial key and a position list:

  ($partkey, @positions)=$it->()

If called in array context the iterator and the number of items it will iterate is returned.

The optional NTH parameter initially positions the iterator within the index as MMapDB::Iterator->nth does.

index_iterator() can be used in combination with index_lookup_position() to create and position an iterator:

 $it=$db->index_iterator($db->index_lookup_position($db->mainidx, qw/key .../));

$it is an MMapDB::Iterator object.


($it, $nitems)=$db->id_index_iterator

iterate over the ID index. The iterator returns 2 elements, the ID and the data record position:

  ($id, $position)=$it->()

If called in array context the iterator and the number of items it will iterate is returned.

$it is an MMapDB::Iterator object.

($id, $pos)=$db->insert([[KEY1, KEY2, ....], SORT, DATA])

($id, $pos)=$db->insert([[KEY1, KEY2, ....], SORT, DATA, ID])

insert a new data record into the database. The ID parameter is optional and should be ommitted in most cases. If it is ommitted the next available ID is allocated and bound to the record. The ID and the position in the new database version are returned. Note, that the position cannot be used as parameter to the data_record method until the transaction is committed.


($keys, $sort, $data, $id)=@{$db->delete_by_id(ID, 1)}

delete an data record by its ID. If the last parameter is true the deleted record is returned if there is one. Otherwise only a status is returned whether a record has been deleted by the ID or not.


deletes all data records from the database.


returns the main index position. You probably need this as parameter to index_lookup().


This is an alternative way to access the database via tied hashes.


returns almost the same as

  $db->data_record( ($db->index_lookup($db->mainidx, KEY1, KEY2))[0] )

provided [KEY1, KEY2] point to an array of data records.

While it is easier to access the database this way it is also much slower.

See also "The MMapDB::Index and MMapDB::IDIndex classes"

$db->datamode=$mode or $mode=$db->datamode

Set/Get the datamode for %{$db->main_index}.

See also "The MMapDB::Index and MMapDB::IDIndex classes"


The same works for indices:


returns almost the same as

  $db->data_record( $db->id_index_lookup(42) )

See also "The MMapDB::Index and MMapDB::IDIndex classes"

$db->id_datamode=$mode or $mode=$db->id_datamode

Set/Get the datamode for %{$db->id_index}.

See also "The MMapDB::Index and MMapDB::IDIndex classes"

The MMapDB::Index and MMapDB::IDIndex classes

$db->main_index and $db->id_index elements are initialized when $db->start is called. They simply point to empty anonymous hashes. These hashes are then tied to these classes.

Now if a hash value is accessed the FETCH function of the tied object checks whether the element points to another index (that means another hash) or to a list of data records. The $db->id_index hash elements are always data records. So, there is nothing to decide here.

If an index is found the reference of another anonymous hash is returned. This hash itself is again tied to a MMapDB::Index object. If a list of data records is found the reference of an anonymous array is returned. The array itself is tied to a MMapDB::Data object.

All 3 classes, MMapDB::Index, MMapDB::IDIndex and MMapDB::Data have a property called datamode. When a MMapDB::Index object creates a new MMapDB::Index or MMapDB::Data object it passes its datamode on.

The datamode itself is either DATAMODE_NORMAL or DATAMODE_SIMPLE. It affects only the behavior of a MMapDB::Data object. But since it is passed on setting it for say tied(%{$db->main_index}) affects all MMapDB::Data leaves created afterwards.

The MMapDB::Data class

When a list of data records is found by a MMapDB::Index or MMapDB::IDIndex object it is tied to an instance of MMapDB::Data.

The individual data records can be read in 2 modes DATAMODE_NORMAL and DATAMODE_SIMPLE. In normal mode the record is read with $db->data_record, in simple mode with $db->data_value.

The following example shows the difference:

Create a database with a few data items:

 my $db=MMapDB->new("mmdb");
 $db->start; $db->begin;
 $db->insert([["k1", "k2"], "0", "data1"]);
 $db->insert([["k1", "k2"], "1", "data2"]);
 $db->insert([["k2"], "0", "data3"]);

Now, in DATAMODE_NORMAL it looks like:

   k1 => {
           k2 => [
                   [['k1', 'k2'], '0', 'data1', 1],
                   [['k1', 'k2'], '1', 'data2', 2],
   k2 => [
           [['k2'], '0', 'data3', 3],

Now, we set $db->datamode=DATAMODE_SIMPLE and it looks like:

   k1 => {
           k2 => ['data1', 'data2'],
   k2 => ['data3'],

The MMapDB::Iterator class

To this class belong all iterators documented here. An iterator is mainly simply called as function reference to fetch the next element:


But sometimes one wants to reposition an iterator or ask it a question. Here this class comes into play.

Some iterators don't implement all of the methods. In this case an E_NOT_IMPLEMENTED exception is thrown.

The index and id-index iterators implement all methods.



If called in void context the iterator position is moved to the $n'th element. The element read by the next $it->() call will be this one.

If called in other context the iterator position is moved to the $n'th element which then is returned. The next $it->() call will return the ($n+1)'th element.

An attempt to move the position outside its boundaries causes an E_RANGE exception.


returns the current iterator position as an integer starting at 0 and counting upwards by 1 for each element.


returns the number of elements that the iterator will traverse.


An MMapDB object is internally an array. If another module want to inherit from MMapDB and needs to add other member data it can add elements to the array. @MMapDB::attributes contains all attribute names that MMapDB uses.

To add to this list the following scheme is recommended:

 package MMapDB::XX;
 use MMapDB;
 our @ISA=('MMapDB');
 our @attributes;

 # add attribute accessors
   @attributes=(@MMapDB::attributes, qw/attribute1 attribute2 .../);
   for( my $i=@MMapDB::attributes; $i<@attributes; $i++ ) {
     my $method_num=$i;
     no strict 'refs';
        sub : lvalue {$_[0]->[$method_num]};

Now an MMapDB::XX object can call $obj->filename to get the MMapDB filename attribute or $obj->attribute1 to get its own attribute1.

If another module then wants to inherit from MMapDB::XX it uses

   @attributes=(@MMapDB::XX::attributes, qw/attribute1 attribute2 .../);
   for( my $i=@MMapDB::XX::attributes; $i<@attributes; $i++ ) {

Multiple inheritance is really tricky this way.


None by default.

Error constants are imported by the :error tag.

DATAMODE_SIMPLE is imported by the :mode tag.

All constants are imported by the :all tag.


The t/002-benchmark.t test as the name suggests is mainly about benchmarking. It is run only if the BENCHMARK environment variable is set. E.g. on a bash command line:

 BENCHMARK=1 make test

It creates a database with 10000 data records in a 2-level hash of hashes structure. Then it finds the 2nd level hash with the largest number of elements and looks for one of the keys there.

This is done for each database format using both index_lookup and the tied interface. For comparison 2 perl hash lookups are also measured:

 sub {(sub {scalar @{$c->{$_[0]}->{$_[1]}}})->($k1, $k2)};
 sub {scalar @{$c->{$k1}->{$k2}}};

As result you can expect something like this:

             Rate mmdb_L mmdb_N mmdb_Q mmdb_J hash1 idxl_N idxl_Q idxl_J idxl_L hash2
 mmdb_L   41489/s     --    -0%    -1%    -1%  -89%   -92%   -93%   -93%   -93%  -99%
 mmdb_N   41696/s     1%     --    -0%    -1%  -89%   -92%   -93%   -93%   -93%  -99%
 mmdb_Q   41755/s     1%     0%     --    -1%  -89%   -92%   -93%   -93%   -93%  -99%
 mmdb_J   42011/s     1%     1%     1%     --  -89%   -92%   -93%   -93%   -93%  -99%
 hash1   380075/s   816%   812%   810%   805%    --   -31%   -32%   -33%   -33%  -87%
 idxl_N  548741/s  1223%  1216%  1214%  1206%   44%     --    -2%    -3%    -4%  -81%
 idxl_Q  560469/s  1251%  1244%  1242%  1234%   47%     2%     --    -1%    -2%  -81%
 idxl_J  568030/s  1269%  1262%  1260%  1252%   49%     4%     1%     --    -0%  -81%
 idxl_L  570717/s  1276%  1269%  1267%  1258%   50%     4%     2%     0%     --  -81%
 hash2  2963100/s  7042%  7006%  6996%  6953%  680%   440%   429%   422%   419%    --

The mmdb tests use the tied interface. idxl means index_lookup. hash1 is the first of the 2 perl hash lookups above, hash2 the 2nd.

hash2 is naturally by far the fastest. But add one anonymous function level and idxl becomes similar. Further, the N format on this machine requires byte rearrangement. So, it is expected to be slower. But it differs only in a few percents from the other idxl.


The database consists only of one file. So a backup is in principle a simple copy operation. But there is a subtle pitfall.

If there are writers active while the copy is in progress it may become invalid between the opening of the database file and the read of the first block.

So, a better way is this:

 perl -MMMapDB -e 'MMapDB->new(filename=>shift)->backup' DATABASENAME

To restore a database use this one:

 perl -MMMapDB -e 'MMapDB->new(filename=>shift)->restore' DATABASENAME

See also backup() and restore()


In this paragraph the letter S is used to designate a number of 4 or 8 bytes according to the intfmt. The letter F is used as a variable holding that format. So, when you see F/a* think N/a* or Q/a* etc. according to intfmt.

If an integer is used somewhere in the database it is aligned correctly. So, 4-byte integer values are located at positions divisible by 4 and 8-byte integers are at positions divisible by 8. This also requires strings to be padded up to the next divisible by 4 or 8 position.

Strings are always packed as F/a* and padded up to the next S byte boundary. With pack this can be achieved with F/a* x!S where F is one of N, L, J or Q and S either 4 or 8.

The database can be divided into a few major sections:

  • the database header

  • the data records

  • the indices

  • the string table

Differences between database format 0 and 1

Format 0 and 1 are in great parts equal. Only the database header and the string table entry slightly differ.

The database header

Database Format 0

At start of each database comes a descriptor:

 | MAGIC NUMBER (4 bytes) == 'MMDB' |
 | FORMAT (1 byte) + 3 bytes resrvd |
 | MAIN INDEX POSITION (S bytes)    |
 | ID INDEX POSITION (S bytes)      |
 | NEXT AVAILABLE ID (S bytes)      |

The magic number always contains the string MMDB. The FORMAT byte contains the pack format letter that describes the integer format of the database. It corresponds to the intfmt property. When a database becomes invalid a NULL byte is written at this location.

Main Index position is the file position just after all data records where the main index resides.

ID Index is the last index written to the file. This field contains its position.

The NEXT AVAILABLE ID field contains the next ID to be allocated.

The STRING TABLE POSITION keeps the offset of the string table at the end of the file.

Database Format 1

At start of each database comes a descriptor:

 | MAGIC NUMBER (4 bytes) == 'MMDC' |
 | FORMAT (1 byte)                  |
 | FLAGS  (1 byte)                  |
 | 2 bytes reserved                 |
 | MAIN INDEX POSITION (S bytes)    |
 | ID INDEX POSITION (S bytes)      |
 | NEXT AVAILABLE ID (S bytes)      |

In format 1 the database header sightly differs. The magic number is now MMDC instead of MMDB. One of the reserved bytes following the format indicator is used as flags field.

All other fields remain the same.

Data Records

Just after the descriptor follows an arbitrary umber of data records. In the following diagrams the pack format is shown after most of the fields. Each record is laid out this way:

 | VALID FLAG (S bytes)             |
 | ID (S bytes)                     |
 | NKEYS (S bytes)                  |
 ¦                                  ¦
 ¦ KEY POSITIONS (NKEYS * S bytes)  ¦
 ¦                                  ¦
 | SORT POSITION                    |
 | DATA POSITION                    |

A data record consists of a certain number of keys, an ID, a sorting field and of course the data itself. The key, sort and data positions are offsets from the start of the string table.

The valid flag is 1 if the data is active or 0 if it has been deleted. In the latter case the next transaction will purge the database.

The Index

Just after the data section follows the main index. Its starting position is part of the header. After the main index an arbitrary number of subindices may follow. The last index in the database is the ID index.

An index is mainly an ordered list of strings each of which points to a list of positions. If an index element points to another index this list contains 1 element, the position of the index. If it points to data records the position list is ordered according to the sorting field of the data items.

An index starts with a short header consisting of 2 numbers followed by constant length index records:

 | NRECORDS (S bytes)               |
 | RECORDLEN (S bytes)              |  in units of integers
 ¦                                  ¦
 ¦ constant length records          ¦
 ¦                                  ¦

The record length is a property of the index itself. It is the length of the longest index record constituting the index. It is expressed in units of S bytes.

Each record looks like this:

 | KEY POSITION (S bytes)           |
 | NPOS (S bytes)                   |
 ¦                                  ¦
 ¦ POSITION LIST (NPOS * S bytes)   ¦
 ¦                                  ¦
 ¦                                  ¦
 ¦ padding up to RECORDLEN          ¦
 ¦                                  ¦

The KEY POSITION contains the offset of the record's partial key from the start of the string table. NPOS is the number of elements of the subsequent position list. Each position list element is the starting point of a data record or another index relativ to the start of the file.

The string table

Database Format 0

The string table is located at the end of the database file that so far consists only of integers. There is no structure in the string table. All strings are simply padded to the next integer boundary and concatenated.

Each string is encoded with the F/a* x!S pack-format:

 | LENGTH (S bytes)                 |
 ¦                                  ¦
 ¦ OCTETS (LENGTH bytes)            ¦
 ¦                                  ¦
 ¦                                  ¦
 ¦ padding                          ¦
 ¦                                  ¦

Database Format 1

In database format 1 a byte representing the utf8-ness is appended to each string. The pack-format F/a*C x!S is used:

 | LENGTH (S bytes)                 |
 ¦                                  ¦
 ¦ OCTETS (LENGTH bytes)            ¦
 ¦                                  ¦
 ¦ UTF8 INDICATOR (1 byte)          ¦
 ¦                                  ¦
 ¦ padding                          ¦
 ¦                                  ¦


Torsten Förtsch, <torsten.foertsch@gmx.net>


Copyright (C) 2009-2010 by Torsten Förtsch

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself, either Perl version 5.10.0 or, at your option, any later version of Perl 5 you may have available.