openldap/servers/slapd/back-ldap/TODO.proxy

back-proxy

A proxy that handles a pool of URI associated to a unique suffix.
Each request is spread over the different URIs and results are
masqueraded to appear as coming from a unique server.

Suppose a company has two branches, whose existing DS have URIs

"ldap://ldap.branch1.com/o=Branch 1, c=US"
"ldap://ldap.branch2.it/o=Branch 2, c=IT"

and it wants to propose to the outer world as a unique URI

"ldap://ldap.company.net/dc=company, dc=net"

It could do some rewriting to map everything that comes in with a base dn
of "o=Branch 1, dc=company, dc=net" as the URI of the Branch 1, and
everything that comes in with a base dn of "o=Branch 2, dc=company, dc=net"
as the URI of Branch 2, and by rewriting all the dns back to the new, uniform 
base. Everything that comes in with a base dn of "dc=company, dc=net" should 
be handled locally and propagated to the two branch URIs if a subtree 
(or at least onelevel) search is required.

Operations:

- bind
- unbind
- search
- compare
- add
- modify
- modrdn
- delete
- abandon

The input of each operation may be related to:

		exact dn	exact parent	ancestor
-------------------------------------------------------------
bind		x
unbind
search		x		x		x
compare		x
add				x
modify		x
modrdn		x
delete		x
abandon

The backend must rely on a dn fetching mechanism. Each operation requires
to determine as early as possible which URI will be able to satisfy it.
Apart from searches, which by definition are usually allowed to return
multiple results, and apart from unbind and abandon, which do not return any
result, all the remaining operations require the related entry to be unique.

A major problem isposed by the uniqueness of the dns. As far as the suffixes
are masqueraded by a common suffix, tyhe dns are no longer guaranteed to be
unique. This backend relies on the assumption that the uniqueness of the
dns is guaranteed.

Two layers of depth in dn fetching are envisaged.
The first layer is provided by a backend-side cache made of previously
retrieved entries. The cache relates each rdn (i.e. the dn apart from the
common suffix) to the pool of URIs that are expected to contain a subset
of its children.

The second layer is provided by a fetching function that spawns a search for
each URI in the pool determined by the cache if the correct URI has not been 
directly determined.

Note that, as the remote servers may have been updated by some direct
operation, this mechanism does not guarantee the uniqueness of the result.
So write operations will require to skip the cache search and to perform
the exaustive search of all the URIs unless some hint mechanism is provided
to the backend (e.g. a server is read-only).

Again, the lag between the fetching of the required dn and the actual
read/write may result in a failure; however, this applies to any LDAP 
operation AFAIK.

- bind
if updates are to be strictly honored, a bind operation is performed against
each URI; otherwise, it is performed against the URIs resulting from a
cache-level dn fetch.

- unbind
nothing to say; all the open handles related to the connection are reset.

- search
if updates are to be strictly honored, a search operation is performed agaist
each URI. Note that this needs be performed also when the backend suffix
is used as base. In case the base is stricter, the URI pool may be restricted 
by performing a cache dn fetch of the base first.

- compare
the same applies to the compare dn.

- add
this operation is delicate. Unless the dn up to the top-level part excluded
can be uniquely associated to a URI, and unless its uniqueness can be trusted,
no add operation should be allowed.