This question is raised many times, in different forms. The most common, however, is: Why doesn't OpenLDAP drop Berkeley DB and use a relational database management system (RDBMS) instead? In general, expecting that the sophisticated algorithms implemented by commercial-grade RDBMS would make OpenLDAP be faster or somehow better and, at the same time, permitting sharing of data with other applications.
The short answer is that use of an embedded database and custom indexing system allows OpenLDAP to provide greater performance and scalability without loss of reliability. OpenLDAP uses Berkeley DB concurrent / transactional database software. This is the same software used by leading commercial directory software.
Now for the long answer. We are all confronted all the time with the choice RDBMSes vs. directories. It is a hard choice and no simple answer exists.
It is tempting to think that having a RDBMS backend to the directory solves all problems. However, it is a pig. This is because the data models are very different. Representing directory data with a relational database is going to require splitting data into multiple tables.
Think for a moment about the person objectclass. Its definition requires attribute types objectclass, sn and cn and allows attribute types userPassword, telephoneNumber, seeAlso and description. All of these attributes are multivalued, so a normalization requires putting each attribute type in a separate table.
Now you have to decide on appropriate keys for those tables. The primary key might be a combination of the DN, but this becomes rather inefficient on most database implementations.
The big problem now is that accessing data from one entry requires seeking on different disk areas. On some applications this may be OK but in many applications performance suffers.
The only attribute types that can be put in the main table entry are those that are mandatory and single-value. You may add also the optional single-valued attributes and set them to NULL or something if not present.
But wait, the entry can have multiple objectclasses and they are organized in an inheritance hierarchy. An entry of objectclass organizationalPerson now has the attributes from person plus a few others and some formerly optional attribute types are now mandatory.
What to do? Should we have different tables for the different objectclasses? This way the person would have an entry on the person table, another on organizationalPerson, etc. Or should we get rid of person and put everything on the second table?
But what do we do with a filter like (cn=*) where cn is an attribute type that appears in many, many objectclasses. Should we search all possible tables for matching entries? Not very attractive.
Once this point is reached, three approaches come to mind. One is to do full normalization so that each attribute type, no matter what, has its own separate table. The simplistic approach where the DN is part of the primary key is extremely wasteful, and calls for an approach where the entry has a unique numeric id that is used instead for the keys and a main table that maps DNs to ids. The approach, anyway, is very inefficient when several attribute types from one or more entries are requested. Such a database, though cumbersomely, can be managed from SQL applications.
The second approach is to put the whole entry as a blob in a table shared by all entries regardless of the objectclass and have additional tables that act as indices for the first table. Index tables are not database indices, but are fully managed by the LDAP server-side implementation. However, the database becomes unusable from SQL. And, thus, a fully fledged database system provides little or no advantage. The full generality of the database is unneeded. Much better to use something light and fast, like Berkeley DB.
A completely different way to see this is to give up any hopes of implementing the directory data model. In this case, LDAP is used as an access protocol to data that provides only superficially the directory data model. For instance, it may be read only or, where updates are allowed, restrictions are applied, such as making single-value attribute types that would allow for multiple values. Or the impossibility to add new objectclasses to an existing entry or remove one of those present. The restrictions span the range from allowed restrictions (that might be elsewhere the result of access control) to outright violations of the data model. It can be, however, a method to provide LDAP access to preexisting data that is used by other applications. But in the understanding that we don't really have a "directory".
Existing commercial LDAP server implementations that use a relational database are either from the first kind or the third. I don't know of any implementation that uses a relational database to do inefficiently what BDB does efficiently. For those who are interested in "third way" (exposing EXISTING data from RDBMS as LDAP tree, having some limitations compared to classic LDAP model, but making it possible to interoperate between LDAP and SQL applications):
OpenLDAP includes back-sql - the backend that makes it possible. It uses ODBC + additional metainformation about translating LDAP queries to SQL queries in your RDBMS schema, providing different levels of access - from read-only to full access depending on RDBMS you use, and your schema.
For more information on concept and limitations, see slapd-sql(5) man page, or the Backends section. There are also several examples for several RDBMSes in back-sql/rdbms_depend/* subdirectories.