The Maintenance of Duplicate Databases

Preface:

This RFC is a working paper on the problem of maintaining duplicated databases in an ARPA-like network. It briefly discusses the general duplicate database problem, and then outlines in some detail a solution for a particular type of duplicate database. The concepts developed here were used in the design of the User Identification Database for the TIP user authentication and accounting system. We believe that these concepts are generally applicable to distributed database problems.

Introduction

    There are a number of motivations for maintaining redundant,
duplicate copies of databases in a distributed network environment.  Two
important motivations are:

• to increase reliability of data access.

The accessibility of critical data can be increased by redundantly maintaining it. The database used for TIP login and accounting is redundantly distributed to achieve highly reliable access.

• to insure efficiency of data access.

Data can be more quickly and efficiently accessed when it is "near" the accessing process. A copy of the TIP user ID database is maintained at each site supporting the TIP login service to insure rapid, efficient access. (Reliability considerations dictate that this database be redundantly maintained, and efficiency considerations dictate that a copy be maintained at each authentication site.)

    The design of a system to maintain redundant, duplicate databases is
a challenging task because of the inherent communication delay between
copies of the database, as well as the real world constraints of system
crashes, operator error, communication channel failure, etc.  This paper
discusses some of the problems we encountered in designing such a
system, and outlines a system design for maintaining a particular type
of database which solves those problems.

The Model

    A system for supporting duplicate copies of a database can be
modeled by a group of independent database management processes (DBMPs)
each maintaining its own copy of the database. These processes
communicate with each other over network communication paths. Each DBMP
has complete control over its copy of the database. It handles all
accesses to and modifications of the database in response to requests
from other processes. Though the DBMPs act only upon requests, in the
following they will often be said to be actually causing or originating
the modifications.

    An important design consideration is that the communication paths
between the DBMPs are subject to failures. Thus one DBMP may have its
interactions with other DBMPs interrupted and/or have to wait until
communication paths are re-established before it can communicate with
other DBMPs. An assumption made in this paper about the communication

paths is that messages from one process to another are delivered in the same order that they are sent. This is true of the ARPANET. For networks that make no such guarantee, communication protocols between the DBMPs can be used to correctly order the messages.

    In order to proceed further, it is necessary to be more precise
about the nature of the duplicated database and the operations allowed
on it. A constant, read-only database is at one extreme. The task of the
DBMPs is trivial in this case. They simply respond to value retrieval
requests. At the other extreme is a general shared database where
functional modification requests (such as "X := f(X,Y,Z)") are allowed
and/or where it is necessary to completely restrict access to entries
while they are being modified. In this case all the problems of shared
databases on a single computer system arise (e.g., the need for
synchronization mechanisms and the resulting potential deadlock
situations), as well as those unique to having multiple database copies
distributed among independent computers. For example, a completely
general system must deal with the possibility of communication failures
which cause the network to become partitioned into two or more sub-
networks. Any solution which relies on locking an element of the
database for synchronized modification must cope with the possibility of
processes in non-communicating sub-networks attempting to lock the same
element. Either they both must be allowed to do so (which violates the
lock discipline), or they both must wait till the partition ceases
(which may take arbitrarily long), or some form of centralized or
hierarchical control must be used, with a resulting dependency of some
DBMPs on others for all modifications and perhaps accesses as well.

The Database

    The type of database to be examined in this paper can be represented
as a collection of entries which are (Selector,Value) pairs. Each
selector is unique and the values are atomic entities as far as the
DBMPs are concerned. The mechanisms to be presented may be extended to
handle databases with greater structure - where the values may
themselves be collections of (selector,value) pairs - but this extension
will not be considered further here.

Four operations are to be allowed on the database:

1) Selection - given a selector, the current associated value is returned.

2) Assignment - a selector and a value are given and the given value replaces the old value associated with the selector.

3) Creation - a new selector and an initial value are given and a new (selector,initial value) entry is added to the database.

4) Deletion- a selector is given and the existing (selector,value) entry is removed from the database.

Note that value modification is limited to assignment. Functional modification requests - such as "Change X to be Factorial(X)" - are specifically ruled out. Allowing them would force the use of system wide synchronization interlocks.

Consistency

    The extent to which the copies of the database can be kept
"identical" must be examined. Because of the inherent delay in
communications between DBMPs, it is impossible to guarantee that the
data bases are identical at all times. Rather, our goal is to guarantee
that the copies are "consistent" with each other. By this we mean that
given a cessation of update activity to any entry, and enough time for
each DBMP to communicate with all other DBMPs, then the state of that
entry (its existence and value) will be identical in all copies of the
database.

Timestamps

    We permit modifications to the database to originate at any of the
DBMPs maintaining it. These changes must, of course, be communicated to
the other DBMPs. To insure consistency, all of the DBMPs must make the
same decision as to which modification to a particular entry is to be
considered "final". It is desirable to select the "most recent" change.
However, since there is no way to absolutely determine the time sequence
of events in a distributed system without a universal, always accessible
sequence number generator (a network time standard should suffice),
"most recent" can only be approximated. We accomplish the approximation
by associating a timestamp with each modification and with each entry,
the latter being the timestamp of the modification which set its current
value.(1) Since the uniqueness of timestamps given out at more than one
----------
(1) Time is useful in this context because it has the desired properties
of being monotonically increasing, and of being available with a
reasonable degree of accuracy. Any other sequence numbering scheme with
these properties can be used, "time-of-day" was chosen because it is
simple to obtain. Its main faults are that it is often manually set (and
thus prone to error), and it may stop during system service

interruptions. As computer systems learn to adapt to a network environment, the use of an independent time source should become more common. This is beginning to happen with the TENEX sites on the ARPANET.

DBMP can not be guaranteed, a "DBMP of origin" is included as part of
each timestamp. By (arbitrarily) ordering the DBMPs, we thus have a
means of uniquely ordering timestamps. Each
 timestamp is a pair (T,D): T is a time, D is a DBMP identifier. For two
timestamps (T1,D1) and (T2,D2) we have the following:

    (T1,D1)>(T2,D2) <=> (T1>T2) or (T1=T2 and D1>D2)
    (T1,D1) is said to be "more recent" than (T2,D2)

If D1=D2 and T1=T2 it is assumed that the two modifications are really two copies of the same modification request.

    In order to insure the uniqueness of timestamps, it is necessary
that each individual DBMP associate different times with different
modifications. This is certainly possible to do, though the fineness of
the unit of time may restrict the frequency of modifications at a single
DBMP site.

Each entry in the data base is now a triple:

E ::= (S,V,T), where
S is the selector
V is the associated value
T is the timestamp (a Time,DBMP pair) of the last change to the

        entry

    The task of each DBMP is to keep its copy of the database up-to-
date, given the information on modifications that it has received so
far. At the same time it must insure that each of its entries stays
consistent with those of all the other DBMPs. This must be done despite
the fact that the order in which it receives modifications may be very
different from the order in which they are received by other DBMPs. In
the remainder of this paper we examine the allowable database operations
with respect to their effect on DBMP operation.

Assignment

    Consider the case of assignment to an existing entry. When the
assignment is initiated (by a person or process) the DBMP involved makes
the change locally, and creates a copy of the modified entry and an
associated list of DBMPs to which the change must be sent. As the
modification is delivered to the other DBMPs, they are removed from the
list until no DBMPs remain. The copy of the modification is then
deleted. This distribution mechanism must be error free (i.e., receipt

of a modification must be positively confirmed before removing a DBMP

from the list of recipients).(2)

    When a DBMP receives an assignment modification (either locally or
from another DBMP) it compares the timestamp of the modification with
the timestamp of the copy of the entry in its database and keeps
whichever is more "recent" as defined by the ordering given above.  Thus
when all existing assignments to a given entry have been distributed to
all the DBMPs, they are guaranteed have the same "latest" value
associated with that entry.

Creation

    Creation and deletion of entries pose more of a problem. Note that
the ability to create new, previously unknown entries requires that a
DBMP be able to handle assignments to unknown entries. For example,
consider the case of an entry XYZ created by DBMP A, and the following
sequence of events: DBMP A tells DBMP B about the new entry, and
subsequently B assigns a new value to XYZ; DBMP B then tells DBMP C
about the assignment before C has heard from A about the creation.  DBMP
C must either save the assignment to XYZ until it hears about the
creation, or simply assume the creation will be coming and use the "new"
entry right away. The latter is more in the spirit of trying to keep the
database as "up-to-date" as possible and leads to no inconsistencies.

Deletion

    Deletion of entries is the main problem for this database system.
If deletion is taken to mean immediate removal from the database, then
problems arise. Consider the following scenario:
        XYZ is an entry known by all DBMPs.
        XYZ is deleted at DBMP A.
        XYZ is modified at DBMP B (before B is notified of the deletion
            by A).
Now, consider a third DBMP, C. C may hear from DBMP B before DBMP A, in
which case XYZ ends up deleted at DBMP C. C may however hear from DBMP A

----------
(2) This same process (local modification and queuing for remote
distribution) is, of course, performed for the other possible operations
on the database. The details of how the local modification is done
safely, how the messages are queued, how confirmation of delivery is
done, etc., though important, will not be discussed here.  The use of an
addressee list attached to the modification to be delivered is
conceptually easy to work with and not difficult to implement in
practice.

before DBMP B. In this case, if C removes XYZ from its database, then the assignment to XYZ initiated by DBMP B will result in the re-creation of XYZ at DBMP C. To prevent this C must remember that XYZ has been deleted until it is "safe" to completely forget about XYZ.

    One approach to this problem is to never actually remove an entry
from the database. Deletion just marks the entry as being deleted by
setting a "deleted" flag associated each entry. However, the problem of
receiving assignments to deleted entries still exists. For example, DBMP
A may receive an assignment from DBMP B to an entry which A has marked
as deleted. DBMP A can not tell whether B has not heard about the
deletion, or has heard about it but has also received a re-creation
request for the entry, which hasn't reached DBMP A. To enable A to
resolve such situations we include another timestamp in all entries: the
timestamp of the entry's creation. Thus in the above example, DBMP A can
compare the creation timestamps of the assignment and the deleted entry.
The one with the later creation timestamp is kept. Indeed whenever a
modification with an old creation timestamp is received it can be
ignored.

We now have a 5-tuple for entries and modifications:

E ::= (S,V,F,CT,T)
S is the selector
V is the associated value
F is the deleted/not-deleted flag
CT is the timestamp of creation
T is the timestamp of this (last) modification

    Note that the values of the F, CT, and T components of a
modification uniquely specify the type of modification. Thus only the
5-tuple to become the new entry for a selector, not the type of
modification, need be communicated:
        F = not deleted, CT = T => creation
        F = not deleted, CT < T => assignment
        F = deleted => deletion

    The mechanism described above handles all the desired operations on
the distributed database in a way that guarantees the consistency of all
copies. A modification to the database will take effect at each DBMP as
soon as it receives the request from the DBMP originating the change.

    A deficiency with this scheme is that deleted entries are never
removed from the database. A method which permits "garbage collection"
of deleted entries is discussed below.

Removal of Deleted Entries

    The basic constraint is that a DBMP should not remove a deleted
entry until it will never receive any assignments with the same selector
(S) and the same or older create time (CT). If it fails to do this, then
it will be unable to distinguish these "out of date" assignments from
assignments to a newly created entry for the same S.  To be able do
this, each DBMP must know for each deleted entry whether all other DBMPs
have heard about the deletion. To accomplish this, each DBMP could
notify the other DBMPs whenever it hears about a deletion. If these
notifications are transmitted in order with the "normal" sequence of
modifications, then upon receipt of such a notification a DBMP can be
sure that the sending DBMP has delivered any outstanding assignments to
the deleted entry, has marked it as deleted, and will not generate any
new assignments to it. Keeping track of, and exchanging messages about,
each individual deleted entry in this manner is, however, somewhat more
elaborate than necessary.

    Having each DBMP deliver all its own modifications in sequential
order (by timestamp) allows the following simplification. We have all
DBMPs maintain a table of the timestamps of the last modification
received from each other DBMP. Any DBMP, say A, is then guaranteed to
have received all modifications originating at another DBMP, say B.
which have timestamps earlier than (or equal to) the entry for B in A s
copy of this table. If this table is exchanged between DBMPs, then all
DBMPs would have a second N*N (N= number of DBMPs) table where entry
(I,J) would be the timestamp of the last modification received by DBMP I
from DBMP J. Thus DBMP A can remove a deleted entry whose deletion
originated at DBMP K when all entries in the Kth column of this table at
DBMP A are later than or equal to the timestamp of the deleted entry.
When a DBMP receives a deletion modification, in addition to acting on
it and acknowledging receipt of it, the DBMP should also send its table
of last timestamps received to all other DBMPs. This is sent in a
timestamped message which is queued with the "normal" modification
messages.

    A refinement to this approach, which reduces the amount of
information exchanged and the size of the tables, is to have the DBMPs
exchange only the oldest of the entries in the first table (of
timestamps of last modifications received from other DBMPs). These would
then be saved in a 1*N table, replacing the N*N table. A DBMP receiving
a modification with a timestamp equal to or older than the oldest
timestamp in its second table knows that the modification has been
confirmed as being received by all other DBMPs. A deleted entry can thus
be removed when its timestamp satisfies this condition. A DBMP would,
upon receipt of a deletion modification, queue up a message with this
"timestamp of oldest last modification received" for delivery to all
other DBMPs.

Summary of solution:

An entry in the database is a 5-tuple:

        (S,V,F,CT,T) where
        S is an selector used in all references to this entry.
        V is its present value.
        F is a deleted/undeleted flag.
        CT is the timestamp of the creation of this entry.
        T is the timestamp of the modification which set the current V
          and/or F of the entry.
        A timestamp is a pair (time,DBMP) where the DBMP identifies the
          site at which the time was generated, and the DBMPs are
          (arbitrarily) ordered, so that timestamps are completely
          ordered.

A modification is a pair (Set-of-DBMPs,Entry) where Set-of-DBMPs is the set of DBMPs to which the Entry has yet to be delivered.

An ordered (by timestamp) list of modifications is kept at each DBMP. The DBMP periodically attempts to deliver modification requests to those DBMPs which remain in the Set-of-DBMPs associated with each modification. Modification entries are removed from this list when they have been delivered to all DBMPs.

When a DBMP receives a modification request from another DBMP, it compares the timestamps of the request with the timestamps of the corresponding entry (if any) in its database, and acts upon or disregards the new entry accordingly.

Each DBMP keeps a vector of the timestamp (T) of the last modification received by it from each other DBMP.

When a DBMP receives a deletion modification, it sends a timestamped message to all other DBMPs containing the oldest timestamp in its vector of timestamps of last modification received. Each DBMP keeps a second vector of the last of these timestamps received from each other DBMP.

A deleted entry may be removed from the database when its timestamp (T) is older than all the timestamps in this second vector of timestamps received from other DBMPs.

Conclusion

    This paper has presented techniques by which a number of loosely
coupled processes can maintain duplicate copies of a database, despite
the unreliability of their only means of communication. The copies of
the database can be kept "consistent'. However it is possible for
seemingly anomalous behavior to occur. For example a user may assign a
value to a selector using one DBMP, later use another DBMP and assign a
new value, and still later find that the first value is the one that
remains in the database. This can occur if the clocks used by the two
DBMPs for their timestamps are sufficiently out of synchrony that the
second assignment is timestamped as having taken place before the first
assignment. To the extent that the communication paths can be made
reliable, and the clocks used by the processes kept close to synchrony,
the probability of seemingly strange behavior can be made very small.
However, the distributed nature of the system dictates that this
probability can never be zero.

    The major innovation presented here is the explicit representation
of the time sequence of modifications through timestamps for both
modifications and entries. This enables the each DBMP to select the same
"most-recent" modification of an entry. In addition, timestamps enable
the DBMPs to decide when a deleted entry is no longer referenced (i.e.,
still active anywhere) and can be deallocated. These techniques should
have broader utility in building and modeling other systems of
concurrent, cooperating processes.

       [ This RFC was put into machine readable form for entry ]
       [ into the online RFC archives by Alex McKenzie with    ]
       [ support from GTE, formerly BBN Corp.            12/99 ]