Network Working Group                           Chris Weider
INTERNET-DRAFT                                  Microsoft Corp.
                                                John Strassner
                                                Cisco
                                                Bob Huston
                                                Iris Associates
Intended Category: Standards Track              November, 1997

LDAP Multi-Master Replication Protocol
<draft-ietf-asid-ldap-mult-mast-rep-02.txt>

Status of this Memo

This document is an Internet-Draft. Internet-Drafts are working
documents of the Internet Engineering Task Force (IETF), its areas, and
its working groups. Note that other groups may also distribute working
documents as Internet-Drafts.

Internet-Drafts are draft documents, and are valid for a maximum of six
months. They may be updated, replaced, or obsoleted by other documents
at any time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as 'work in progress.'

To learn the current status of any Internet-Draft, please check the
'1id-abstracts.txt' listing contained in the Internet-Drafts Shadow
Directories on ds.internic.net (US East Coast), nic.nordu.net (Europe),
ftp.isi.edu (US West Coast), or munnari.oz.au (Pacific Rim).

This draft expires May 20, 1998.

0: Abstract

This paper defines a multi-master, incremental replication protocol
using the LDAP protocol [LDAPv3]. This protocol uses and builds upon
previous LDAP support protocols, namely the changelog [change] and LDIF
[LDIF] protocols. It defines the use of two types of transport protocols
for replication data, and specifies the schema that must be supported by
a server that wishes to participate in replication activities using this
protocol. In addition, it specifies a conflict resolution mechanism for
integrating updates from multiple servers.

1: Introduction

LDAP is increasing in popularity as a generalized query, access, and
retrieval protocol for directory information. Data replication is key
to effectively distributing and sharing such information. Therefore, it
becomes important to create a replication protocol for use specifically
with LDAP to ensure that heterogeneous directory servers can reliably
exchange information. This document defines a multi-master, incremental
replication protocol for use with LDAP. It does not specifically address

the needs of single-master (i.e., master-slave) systems, though this
document could be used as the basis for such a scheme. In addition, it
defines how to use that replication protocol over two transport
mechanisms: standard email and LDAP. The new replication protocol
requires new data to be entered into the directory for use with this
protocol. Therefore, we must define new schema to hold that information.
Also, the data must be transmitted in a specific format; we will use
the proposed LDIF format [LDIF] for doing this.

2: Protocol Behavior

2.1 A glossary of replication terminology

There are 6 axes along which replication functionality can be provided.
These are:

- single-master vs. multi-master
- full vs partial
- whole vs fractional
- transactional vs loosely consistent
- complete vs. incremental
- synchronous vs. asynchronous

Each of these terms is described below.

A single-master (also known as master-slave) replication model assumes
that only one server (the master) allows write access to the replicated
entries. Changes flow from the master server to all of the replicas. A
multi-master replication model assumes that entries can be written on
multiple servers. Changes must then propagate from all masters to every
replica, which requires additional work for conflict resolution. Here,
an update conflict is defined as updating the same data within a given
same replication interval (i.e., not necessarily at exactly the same
time).


Full replication is where every object in a database or DSA is copied to
the replica. Partial replication is where some subset of the objects is
copied.

Whole and fractional replication refer to the attributes transmitted
during replication. If all attributes of the replicated objects are
copied, this is referred to as whole replication. If only a subset of
the attributes are copied, this is referred to as fractional
replication.

Transactional replication requires that the replica gets and commits all
changes between its copy of the data and the master's copy of the data
before the client is notified that the change was successful. Note that
'commit' is used in the general sense to define the action of writing
changes to a data store and verifying that those changes were written
successfully. Specifically, it does NOT imply two-phased commit as used
in databases. Loosely consistent means that there are times when the
written server has data that the replicas do not, from the client's
point
of view. Note also that a general replication topology may well have a
mix of links that are transactional and loosely consistent.

Complete replication requires the replicating server to send a complete
copy of itself to the replica every time it replicates. Incremental
replication allows the replicating server to only send that data which
has changed.

Synchronous replication updates the replica as soon as the source data
is
changed. Asynchronous replication updates the replica some time after
the
source data has been modified.

2.2 Single-master versus multi-master replication

This section provides some additional general information that will help

lay the groundwork for understanding replication.

Replication technology enables the placement of copied and/or shared
data
at different locations distributed throughout an organization. This is
usually done for two reasons: (1) providing 'fast' local access by
eliminating long-distance connections between users and database
servers,
and/or (2) providing corporate-wide high-availability access for
critical
applications, making them more resistant to single system failures.

Replication topologies determine what is to be replicated as well as
what
can be updated when and where. Replication policies define how
replication
is done as well as how update conflicts are resolved. Both of these are
orthogonal to this specification, and so will only be mentioned for
completeness of understanding in this document.

Single-master replication designates one, and only one, copy of the data

as being the 'master' or authoritative source of data. All updates are
done to the master, and the master is responsible for ensuring that all
replicas contain the same data as it does. Single-master schemes require

the definition of both replication topology as well as policy (or
policies),
although since there is by definition no update conflict, these
topologies
and policies are in general simpler than those of their multi-master
counterparts.

Multi-master replication allows updates to different servers, where each
of
the updated servers allow the same naming contexts to be writable. An
update conflict occurs when the same data is updated on multiple servers

within the same replication interval (i.e., not necessarily at the same
time). This means that copies can be temporarily out of sync with each
other. This is acceptable, as long as over time the data converges to
the
same values at all sites. This requires a rich definition of topology
and
policy or policies.

2.3 The basics of multi-master, incremental replication

This specification is aimed primarily at supporting multi-master,
incremental, loosely consistent, asynchronous replication. This
specification could also be used to support single-master replication,
if
one treats only one of the servers in a replication agreement as
writable.
Someof the information required in this specification (i.e., conflict
resolution) is not needed for the single-master case. However, the
ability
to define what has been changed and to ensure that those changes are
propagated throughout the system is common to both single- and
multi-master
replication. Therefore, this specification will address only
multi-master
replication.

To implement multi-master, incremental, loosely consistent, asynchronous

replication, each server which wishes to master data MUST have the
facilities necessary to track changes to the replicate data. In
addition,
each master server MUST have the ability to transmit those changes to
other replicas, and MUST have techniques to implement conflict detection
and resolution. The replication protocol enables servers to transmit
changes over several transport protocols. This document also provides
algorithms for detecting and resolving conflicts.

2.4 The Naming Context (NC)

The Directory Information Base (DIB) is the collection of information
aboutobjects stored in the directory and their relationships. The DIB
may be organized as a hierarchy (or tree), where objects higher in the
hierarchy provide naming resolution for their subordinate objects. This
tree, called the Directory Information Tree (DIT), provides the basis
for using names toquery, access, and retrieve information. The DIT can
in turn be comprised of a set of subtrees.

The basic unit of replication is the NC. A Naming Context consists of a
non-leaf node (called the root of the naming context) and some subset
of its descendants subject to the following restriction: a descendant
cannot be part of a naming context unless all of its ancestors which
are descendants of the naming context root are in the naming context
(e.g. an NC is a complete subtree and cannot have any holes).

Each DSA will have one or more naming contexts. These naming contexts
will be defined and available in the Configuration container pointed to
by the root DSE of the server. The requisite schema are defined in
section 3.

To replicate a given naming context, the only requirement is that the
two servers agree on the contents of every schema entry needed to
defineall the objects in the naming context. The reconciliation of
these entries is beyond the scope of this protocol.

2.4.1 Tracking changes to an NC

Borrowing from the ChangeLog draft [change], each change to a
replicatedNC is logged in its own entry in the changeLog container.
This entry has object class 'changeLogEntry' and holds the trace of the
change, in LDIF format. For more details on the format, see [change].
However, the current ChangeLog draft is designed to provide single
master replication. To provide multi-master, incremental replication,
much more information needs to be kept.

In addition to the information required by the ChangeLog draft, servers
MUST also keep track of the following information and MUST write it to
the changeLog entry:

- a version number for each property of every entry
- a timestamp for the time each property is changed,
- the attributes that were changed in this particular entry
- the object classes of this particular entry
- the naming context in which a given entry resides
- a unique identifier for each entry, which is NOT the DN or RDN of the
  entry

In addition, servers MUST also keep track of the following information
and MAY write it to the changeLog entry:

- a unique identifier for each entry's parent, which is NOT the DN or
  RDN of the parent, when the operation performed on this entry is a
  modifyDN.

2.4.2 Discussion of the required new changeLog information

The version number and timestamp are required for conflict resolution
in multi-master replication.

The attribute and object class tracking are useful for directory
synchronization with special-purpose directories. The actual changes
themselves are stored in a single binary blob in the changeLog entry.
This allows special-purpose directories (such as mail server
directories) to extract only the changes they need.

The NC is required for conflict resolution in multi-master replication.
The NC in which a given entry resides allows efficient replication of a
given naming context. While this may in principle be derivable from the
DN of the changed entry, adding this information allows much easier
retrieval of the appropriate entries.

The unique identifier is required to handle modifyDN conflicts
correctly.

In addition, the server MUST write the entry's parentUniqueID to the
changeLog entry during tracking of a modifyDN operation. This is
required by the reconciliation algorithms defined below.

The new attributes are defined in section 3.

2.5 Defining the replication topology

Each server replicating a given set of naming contexts needs to have
information about that naming context, including information on how to
replicate it. However, this information is orthogonal to the
replication protocol and as such is beyond the scope of this document.

2.6 Replication conflict resolution policies

This section will describe a simple, yet powerful, policy for
reconciling conflicts in a multi-master replication environment. This
policy is one implementation of resolving conflicts. However, some
applications might require more granular control, where different
policies are used for different parts of the DIT or even at different
times under different circumstances. However, a detailed analysis of
different replication policies is beyond the scope of this document.

2.6.1 Using ChangeLog to implement a replication conflict resolution
      policy

In a multi-master environment, conflict resolution between
incompatible updates is crucial. Since each change listed in the
ChangeLog includes the version number of the attribute, every attribute
received in a replication update is reconciled with the local version
of the attribute in the following way:

A. If the version numbers are different, the higher version is favored
B. If the version numbers are the same, the version with the more recent
   time stamp is favored
C. If both the version and time-stamp match, the values themselves are
   compared and the one with the lowest value is favored. This
   guarantees that the system will quiesce consistently.
D. If all three of these match, the values are identical.

If an object is deleted, a server implementing this replication protocol
MUST keep a 'tombstone' of the deleted object. This is essentially a
copy of the deleted object that can be used to restore it. This document

does not specify the length of time that such tombstones must be kept
(this is part of the replication policy that is implemented in a set of
replicated servers). When an object is deleted and there are replication

changes that affect that object, there are some special rules that must
be applied.

E: Deletions are allowed only on objects which have no children. If a
   deletion is received for an object that has a child, the
   reconciliation is to simply ignore the deletion. The server MAY flag
   this as an error and issue an error to the administrator, who is
   then responsible for correcting the problem.

F: If an incoming replication change is to create a new object under an
   already deleted object, then the tombstones of all the ancestors of
   the already deleted object are reanimated and the new object is
   inserted in the correct place. This reanimation must minimally
   restore the RDN and object class attributes of the ancestor.

A modifyDN operation is not considered, for purposes of replication, to
be a combination of a delete and an add operation unless such an
operation would move the object to a new naming context.

In the case where the operation does not cross NC boundaries, it is a
single operation that essentially modifies an entry's parentUniqueID.
Since this attribute is treated as an attribute of the entry itself,
the standard reconciliation logic applies.

In the case where the operation does cross the NC boundaries, it must
be treated as a delete and add combination. A server conforming to
thisspecification will in addition treat the delete and add
combination as an atomic operation.

In addition, a modifyDN or modifyRDN operation may cause two objects
to have the same DN. In that case, the replication system MUST
algorithmically change the RDN of one or both of the objects. The
algorithmically generated RDN is propagated so that the system will
still reach a consistent state. The easiest way to guarantee a non-
conflicting RDN is to use the object's UID as the new RDN.

2.6.2 Loading data

In a replicated environment, the problem of loading multiple remote
systems in a coordinated fashion is much more complex. There are
three different methods to instantiate data at participating
locations in a replicated environment: (1) pre-instantiation,
(2)on-line instantiation, and (3) off-line instantiation.

Pre-instantiation copies all data to all locations before
beginning the actual configuration of the replication environment.
This guarantees identical data copies before allowing the
replication process to begin. This requires that no replication
activity take place against the data until all locations are up
and configured.

On-line instantiation is used when there is an initial location that
has already been populated with data that will be used to replicate
data to a set of remote locations. Note that in a general replication
topology, there may be several authoritative servers that master data
to different sets of replicas. The replication policy that is used in
this case should ensure that there are no duplicate replica sets for
the initial loading of data. This has the advantage of guaranteeing
that all locations have the same data and gives the administrator a
single point of control. However, there can be significant delay while
data are copied over the network.

In off-line instantiation , an initial location is configured as a
master. Replication agreements are then set up between it and other
locations to replicate data at a later time. The initial location stores
changes destined for future locations and pushes them to the other
locations as they come on line. This enables other locations to be
loaded with data and synchronized with the initial location when they
need new data.

On-line instantiation is enabled by the use of the FullReplicaControl
control, discussed in section 4.1.


3: Schema

This section defines new attributes used in this protocol. Object
classes and attributes which are not defined in this document can be
found in [LSPA] or in [change].

3.1 Changes to the ChangeLog document

As noted above, multi-master replication requires a substantial number
of changes to the changeLog document. Here are the new object class and
attributes.

Note that commonName, namingContexts, and description are all defined
in other documents

3.1.1 Changes to changeLogEntry

( 2.16.840.1.113730.3.2.1
   NAME 'changeLogEntry'
   SUP 'top'
   STRUCTURAL
   MUST (
      ChangeNumber $ targetDN $ changeType $ changes $
      changedAttribute $ entryObjectClass $ namingContext $
      uniqueIdentifier )
   MAY  (
      ParentUniqueIdentifier $ NewRDN $ deleteOldRDN $ newSuperior
   )
)

3.1.2 Changed attributes

( 2.16.840.1.113730.3.1.5
   NAME 'changeNumber'
   DESC 'a 64 bit number which uniquely identifies a change made
      to a Directory entry'
   SYNTAX 'Integer'
)

3.1.3 New attributes

(1.2.840.113556.1.4.AAO
   NAME changedAttribute
   DESC 'OID of changed attribute'
   SYNTAX 'DirectoryString'
)

( 1.2.840.113556.1.4.AAR
   NAME 'entryObjectClass'
   DESC 'object class this entry participates in'
   SYNTAX 'DirectoryString'
)

(1.2.840.113556.1.4.AAS
   NAME 'parentUniqueIdentifier'
   DESC 'Unique identifier of the entry's parent'
   SYNTAX 'DirectoryString'
)

3.4 Changes to the LDIF document

To allow incremental efficient multi-master replication, we required two
pieces of information for each attribute to be transmitted that must
appear on a per-attribute basis; version number and timestamp. This
should be transmitted in the LDIF format as qualifiers on the
appropriate attribute: i.e. 'commonName;2,19970308133106Z: Fred Foobar'.
The version number is always the second to last qualifier, the timestamp
is always the last qualifier. Note that this information is formatted
this way for transmission purposes only.

4: LDAP transport

One of the two methods used to transport replication data is by using
the LDAP protocol itself. The target server sets up an ordinary LDAP
session with the source server, binding to the source DSA as the target
server (remember that the root naming context has been replicated
everywhere and so every server participating in a given replication
topology knows about all the other servers) and issues a search with
the new 'replicate' extended control. The target server will specify
the changeLog container as the base of the search, and will use a
filter that states that all records with changeNumber greater than the
current high update number, that reside in one of the replicated naming
contexts, will be given back. The source server MUST then order the
results in such a way so that when they are applied to the replica in
that order, the replica will be synced with the source server at the
time that the replication snapshot was taken. This ordering of the
changes is imperative. One possible way to provide such an ordering
would be to sort the results on changeNumber. There will be a number of
LDAP implementations which may not wish to provide a general sort
facility for search results; however, a conformant implementation of the
replicate control MUST order the results into a correct order.

Once the target starts receiving entries, it then applies each of the
changeLogEntries to its own database, in the same order in which the
entries were sorted, incrementing its highUpdateNumber attribute for
that server appropriately. If the source server has indicated that it
has more entries, the target server can then reissue the search with the
new highUpdateNumber. In an environment with a rapidly changing
directory, the source directory may at its discretion return a maximum
highUpdateNumber indicating the highest number used by the server at the
start of the session. The target server should then use that number as
an
additional term on the filter on subsequent search requests to allow a
'snapshot' of the data to be replicated. Otherwise, the target server
might never close the connection to the source server, which would
impact
source server performance and available bandwidth.

The replicate control is included in the searchRequest and
searchResultDone messages as part of the controls field of the
LDAPMessage, as defined in Section 4.1.12 of [LDAPv3]. The structure of
this control is as follows:

replicateControl ::= SEQUENCE {
    controlType         1.2.840.113556.1.4.617
    criticality         BOOLEAN DEFAULT TRUE
    controlValue        INTEGER (1..2^64-1)
)

The replicateControl controlValue is used by the source server to return
a maximum highUpdateNumber if it wishes to allow the target server to
take a snapshot of the replication data.

4.1: Full updates


If the target server wishes to retrieve a full listing of the current
contents of the DSA, it must issue a fullReplicaControl. This control is
used on a search operation, just like replicateControl. The structure of

the control is as follows:

fullReplicaControl ::=SEQUENCE {
    ControlType 1.2.840.113556.1.4.618
    Criticality         BOOLEAN DEFAULT TRUE
    ControlValue        INTEGER (1..2^64-1)

The fullReplicaControl controlValue is used by the source server on the
first response message to indicate how many entries it will be sending.
If the stream of entries is interrupted, the target server MUST flush
the entries received so far and issue the fullReplicaControl control
again. When the source server sends the last entry, it must set the
ControlValue on the message to the correct highUpdateNumber so that
subsequent replication operations can retrieve only the data that has
been changed.

5: Mail transport

The other method of transporting replication data is by using an email
protocol. In this case, the target server mails the search command with
the replicate extended control or the fullReplicaControl control to the
source server, and then the source server mails the results of the
replication command back to the target server, in LDIF format as
modified
above [LDIF]. When the target server receives the changes, it processes
them as appropriate. The actual mail transport protocol used is not
covered in this document; it needs to be established as a bilateral
agreement between the two servers. The security on this transaction is
enabled by the security of the underlying mail protocol chosen.

6: Security Considerations

Replication requires secure connections and the ability to secure the
change information stored in the directory. Securing the change
information is covered in [change]. Standard LDAP security should be
applied to the LDAP transmission of data. Standard mail security should
be applied to the mail transmission of data. The information necessary
to secure these connections will be stored as part of the URLs defining
the connection points.

7: References

[change] Good, Gordon, Definition of an Object Class to Hold LDAP Change
Records, Internet Draft, July 1997. Available as draft-ietf-asid-
changelog-01.txt

[LDIF] Good, Gordon, The LDAP Data Interchange Format (LDIF) -
Technical Specification, Internet Draft, July 1997. Available as
draft-ietf-asid-ldif-02.txt.

[LSPA] Wahl, M. A Summary of the Pilot X.500 Schema for use in LDAPv3,
Internet Draft, March 1997. Available as draft-ietf-asid-schema-
pilot-00.txt

8: Author's addresses

Chris Weider
Cweider@microsoft.com
 Microsoft Way
Redmond, WA 98052
+1-206-703-2947

John Strassner
Johns@cisco.com
 West Tasman Drive
San Jose, CA 95134
+1-408-527-1069

Bob Huston
bhuston@iris.com
Iris Associates
 Technology Park.
Westford, MA 01886
+1-978-392-5203