The RFC Archive
 The RFC Archive   RFC 1913   « Jump to any RFC number directly 
 RFC Home
Full RFC Index
Recent RFCs
RFC Standards
Best Current Practice
RFC Errata
1 April RFC



IETF RFC 1913

Architecture of the Whois++ Index Service

Last modified on Friday, February 16th, 1996

Permanent link to RFC 1913
Search GitHub Wiki for RFC 1913
Show other RFCs mentioning RFC 1913







Network Working Group                                          C. Weider
Request for Comments: 1913                                        Bunyip
Category: Standards Track                                   J. Fullton
                                                                   CNIDR
                                                                S. Spero
                                                                     EIT
                                                           February 1996


               Architecture of the Whois++ Index Service

 Status of this Memo

   This document specifies an Internet standards track protocol for the
   Internet community, and requests discussion and suggestions for
   improvements.  Please refer to the current edition of the "Internet
   Official Protocol Standards" (STD 1) for the standardization state
   and status of this protocol.  Distribution of this memo is unlimited.

 Abstract

   The authors describe an architecture for indexing in distributed
   databases, and apply this to the WHOIS++ protocol.

1. Purpose:

   The WHOIS++ directory service [Deutsch, et al, 1995] is intended to
   provide a simple, extensible directory service predicated on a
   template-based information model and a flexible query language. This
   document describes a general architecture designed for indexing
   distributed databases, and then applys that architecture to link
   together many of these WHOIS++ servers into a distributed, searchable
   wide area directory service.

2. Scope:

   This document details a distributed, easily maintained architecture
   for providing a unified index to a large number of distributed
   WHOIS++ servers. This architecture can be used with systems other
   than WHOIS++ to provide a distributed directory service which is also
   searchable.

3. Motivation and Introduction:

   It seems clear that with the vast amount of directory information
   potentially available on the Internet, it is simply not feasible to
   build a centralized directory to serve all this information. If we
   are to distribute the directory service, the easiest (although not



Weider, et al               Standards Track                  PAGE 1 top


RFC 1913 Architecture of the Whois++ Index Service February 1996 necessarily the best) way of building the directory service is to build a hierarchy of directory information collection agents. In this architecture, a directory query is delivered to a certain agent in the tree, and then handed up or down, as appropriate, so that the query is delivered to the agent which holds the information which fills the query. This approach has been tried before, most notably in some implementations of the X.500 standard. However, there are number of major flaws with the approach as it has been taken. This new Index Service is designed to fix these flaws. 3.1. The search problem One of the primary assumptions made by recent implementations of distributed directory services is that every entry resides in some location in a hierarchical name space. While this arrangement is ideal for reading the entry once one knows its location, it is not as good when one is searching for the location in the namespace of those entries which meet some set of criteria. If the only criteria we know about a desired entry are items which do not appear in the namespace, we are forced to do a global query. Whenever we issue a global query (at the root of the namespace), or a query at the top of a given subtree in the namespace, that query is replicated to "all" subtrees of the starting point. The replication of the query to all subtrees is not necessarily a problem; queries are cheap. However, every server to which the query has been replicated must process that query, even if it has no entries which match the specified criteria. This part of the global query processing is quite expensive. A poorly designed namespace or a thin namespace can cause the vast majority of queries to be replicated globally, but a very broad namespace can cause its own navigation problems. Because of these problems, search has been turned off at high levels of the X.500 namespace. 3.2. The location problem With global search turned off, one must know in advance how the name space is laid out so that one can guide a query to a proper location. Also, the layout of the namespace then becomes critical to a user's ability to find the desired information. Thus there are endless battles about how to lay out the name space to best serve a given set of users, and enormous headaches whenever it becomes apparent that the current namespace is unsuited to the current usages and must be changed (as recently happened in X.500). Also, assuming one does impose multiple hierarchies on the entries through use of the namespace, the mechanisms to maintain these multiple hierarchies in X.500 do not exist yet, and it is possible to move entries out from under their pointers. Also, there is as yet no agreement on how the X.500 namespace should look even for the White Pages types of information that is currently installed in the X.500 pilot project. Weider, et al Standards Track PAGE 2 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 3.3. The Yellow Pages problem Current implementations of this hierarchical architecture have also been unsuited to solving the Yellow Pages problem; that is, the problem of easily and flexibly building special-purpose directories (say of molecular biologists) and of automatically maintaining these directories once they have been built. In particular, the attributes appropriate to the new directory must be built into the namespace because that is the only way to segregate related entries into a place where they can be found without a global search. Also, there is a classification problem; how does one adequately specify the proper categories so that people other than the creator of the directory can find the correct subtree? Additionally, there is the problem of actually finding the data to put into the subtree; if one must traverse the hierarchy to find the data, we have to look globally for the proper entries. 3.4. Solutions The problems examined in this section can be addressed by a combination of two new techniques: directory meshes and forward knowledge. 4. Directory meshes and forward knowledge We'll hold off for a moment on describing the actual architecture used in our solution to these problems and concentrate on a high level description of what solutions are provided by our conceptual approach. To begin with, although every entry in WHOIS++ does indeed have a unique identifier (resides in a specific location in the namespace) the navigational algorithms to reach a specific entry do not necessarily depend on the identifier the entry has been assigned. The Index Service gets around the namespace and hierarchy problems by creating a directory mesh on top of the entries. Each layer of the mesh has a set of 'forward knowledge' which indicates the contents of the various servers at the next lower layer of the mesh. Thus when a query is received by a server in a given layer of the mesh, it can prune the search tree and hand the query off to only those lower level servers which have indicated that they might be able to answer it. Thus search becomes feasible at all levels of the mesh. In the current version of this architecture, we have chosen a certain set of information to hand up the mesh as forward knowledge. This may or may not be exactly the set of information required to construct a truly searchable directory, but the protocol itself doesn't restrict the types of information which can be handed around. In addition, the protocols designed to maintain the forward knowledge will also work perfectly well to provide replication of servers for Weider, et al Standards Track PAGE 3 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 redundancy and robustness. In this case, the forward knowledge handed around by the protocols is the entire database of entries held by the replicated server. Another benefit provided by the mesh of index servers is that since the entry identification scheme has been decoupled from the navigation service, multiple hierarchies can be built and easily maintained on top of the existing data. Also, the user does not need to know in advance where in the mesh the entry is contained. Also, the Yellow Pages problem now becomes tractable, as the index servers can pick and choose between information proffered by a given server; because we have an architecture that allows for automatic polling of data, special purpose directories become easy to construct and to maintain. 5. Components of the Index Service: 5.1. WHOIS++ servers The whois++ service is described in [Deutsch, et al, 1995]. As that service specifies only the query language, the information model, and the server responses, whois++ services can be provided by a wide variety of databases and directory services. However, to participate in the Index Service, that underlying database must also be able to generate a 'centroid', or some other type of forward knowledge, for the data it serves. 5.2. Centroids as forward knowledge The centroid of a server is comprised of a list of the templates and attributes used by that server, and a word list for each attribute. The word list for a given attribute contains one occurrence of every word which appears at least once in that attribute in some record in that server's data, and nothing else. A word is any token delimited by blank spaces, newlines, or the '@' character, in the value of an attribute. For example, if a whois++ server contains exactly three records, as follows: Record 1 Record 2 Template: User Template: User First Name: John First Name: Joe Last Name: Smith Last Name: Smith Favourite Drink: Labatt Beer Favourite Drink: Molson Beer Weider, et al Standards Track PAGE 4 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 Record 3 Template: Domain Domain Name: foo.edu Contact Name: Mike Foobar the centroid for this server would be Template: User First Name: Joe John Last Name: Smith Favourite Drink: Beer Labatt Molson Template: Domain Domain Name: foo.edu Contact Name: Mike Foobar It is this information which is handed up the tree to provide forward knowledge. As we mention above, this may not turn out to be the ideal solution for forward knowledge, and we suspect that there may be a number of different sets of forward knowledge used in the Index Service. However, the directory architecture is in a very real sense independent of what types of forward knowledge are handed around, and it is entirely possible to build a unified directory which uses many types of forward knowledge. 5.3. Index servers and Index server Architecture A whois++ index server collects and collates the centroids (or other forward knowledge) of either a number of whois++ servers or of a number of other index servers. An index server must be able to generate a centroid for the information it contains. In addition, an index server can index any other server it wishes, which allows one base level server (or index server) to participate in many hierarchies in the directory mesh. 5.3.1. Queries to index servers An index server will take a query in standard whois++ format, search its collections of centroids and other forward information, determine which servers hold records which may fill that query, and then notifies the user's client of the next servers to contact to submit the query (referral in the X.500 model). An index server can also contain primary data of its own; and thus act a both an index server and a base level server. In this case, the index server's response to Weider, et al Standards Track PAGE 5 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 a query may be a mix of records and referral pointers. 5.3.2. Index server distribution model and centroid propogation The diagram on the next page illustrates how a mesh of index servers might be created for a set of whois++ servers. Although it looks like a hierarchy, the protocols allow (for example) server A to be indexed by both server D and by server H. whois++ index index servers servers servers for for whois++ lower-level servers index servers _______ | | | A |__ |_______| \ _______ \----------| | _______ | D |__ ______ | | /----------|_______| \ | | | B |__/ \----------| | |_______| | F | /----------|______| / _______ _______ / | | | |- | C |--------------| E | |_______| |_______|- \ \ _______ \ ______ | | \----------| | | G |--------------------------------------| H | |_______| |______| Figure 1: Sample layout of the Index Service mesh In the portion of the index tree shown above, whois++ servers A and B hand their centroids up to index server D, whois++ server C hands its centroid up to index server E, and index servers D and E hand their centroids up to index server F. Servers E and G also hand their centroids up to H. The number of levels of index servers, and the number of index servers at each level, will depend on the number of whois++ servers deployed, and the response time of individual layers of the server tree. These numbers will have to be determined in the field. Weider, et al Standards Track PAGE 6 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 5.3.3. Centroid propogation and changes to centroids Centroid propogation is initiated by an authenticated POLL command (sec. 5.2). The format of the POLL command allows the poller to request the centroid of any or all templates and attributes held by the polled server. After the polled server has authenticated the poller, it determines which of the requested centroids the poller is allowed to request, and then issues a CENTROID-CHANGES report (sec. 5.3) to transmit the data. When the poller receives the CENTROID- CHANGES report, it can authenticate the pollee to determine whether to add the centroid changes to its data. Additionally, if a given pollee knows what pollers hold centroids from the pollee, it can signal to those pollers the fact that its centroid has changed by issuing a DATA-CHANGED command. The poller can then determine if and when to issue a new POLL request to get the updated information. The DATA-CHANGED command is included in this protocol to allow 'interactive' updating of critical information. 5.3.4. Centroid propogation and mesh traversal When an index server issues a POLL request, it may indicate to the polled server what relationship it has to the polled. This information can be used to help traverse the directory mesh. Two fields are specified in the current proposal to transmit the relationship information, although it is expected that richer relationship information will be shared in future revisions of this protocol. One field used for this information is the Hierarchy field, and can take on three values. The first is 'topology', which indicates that the indexing server is at a higher level in the network topology (e.g. indexes the whole regional ISP). The second is 'geographical', which indicates that the polling server covers a geographical area subsuming the pollee. The third is 'administrative', which indicates that the indexing server covers an administrative domain subsuming the pollee. The second field used for this information is the Description field, which contains the DESCRIBE record of the polling server. This allows users to obtain richer metainformation for the directory mesh, enabling them to expand queries more effectively. 5.3.5. Query handling and passing algorithms When an index server receives a query, it searches its collection of centroids and determines which servers hold records which may fill that query. As whois++ becomes widely deployed, it is expected that some index servers may specialize in indexing certain whois++ Weider, et al Standards Track PAGE 7 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 templates or perhaps even certain fields within those templates. If an index server obtains a match with the query "for those template fields and attributes the server indexes", it is to be considered a match for the purpose of forwarding the query. 5.3.5.1. Query referral Query referral is the process of informing a client which servers to contact next to resolve a query. The syntax for notifying a client is outlined in section 5.5. 5.3.6 Loop control Since there are no a priori restrictions on which servers may poll which other servers, and since a given server may participate in many sub-meshes, mechanisms must be installed to allow the detection of cycles in the polling relationships. This is accomplished in the current protocol by including a hop-count on polling relationships. Each time a polled server generates forward information, it informs the polling server about its current hopcount, which is the maximum of the hopcounts of all the servers it polls, plus 1. A base level server (one which polls no other servers) will have a hopcount of 0. When a server decides to poll a new server, if its hopcount goes up, then it must information all the other servers which poll it about its new hopcount. A maximum hopcount (8 in the current version) will help the servers detect polling loops. A second approach to loop detection is to do all the work in the client; which would determine which new referrals have already appeared in the referral list, and then simply iterate the referral process until there are no new servers to ask. An algorithm to accomplish this in WHOIS++ is detailed in [Faltstrom 95]. 6. Syntax for operations of the Index Service: The syntax for each protocol componenet is listed below. In addition, each section contains a listing of which of these attributes is required and optional for each of the componenet. All timestamps must be in the format YYYYMMDDHHMM and in GMT. 6.1. Data changed syntax The data changed template look like this: # DATA-CHANGED Version-number: // version number of index service software, used to // insure compatibility. Current value is 1.0 Time-of-latest-centroid-change: // time stamp of latest centroid Weider, et al Standards Track PAGE 8 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 // change, GMT Time-of-message-generation: // time when this message was generated, // GMT Server-handle: // IANA unique identifier for this server Host-Name: // Host name of this server (current name) Host-Port: // Port number of this server (current port) Best-time-to-poll: // For heavily used servers, this will identify // when the server is likely to be lightly // loaded so that response to the poll will be //speedy, GMT Authentication-type: // Type of authentication used by server, or NONE Authentication-data: // data for authentication # END // This line must be used to terminate the data changed // message Required/optional table Version-Number REQUIRED Time-of-latest-centroid-change REQUIRED Time-of-message-generation REQUIRED Server-handle REQUIRED Host-Name REQUIRED Host-Port REQUIRED Best-time-to-poll OPTIONAL Authentication-type OPTIONAL Authentication-data OPTIONAL 6.2. Polling syntax # POLL: Version-number: // version number of poller's index software, used to // insure compatibility Type-of-poll: // type of forward data requested. CENTROID or QUERY // are the only one currently defined Poll-scope: // Selects bounds within which data will be returned. // See note. Start-time: // give me all the centroid changes starting at this // time, GMT End-time: // ending at this time, GMT Template: // a standard whois++ template name, or the keyword ALL, // for a full update. Field: // used to limit centroid update information to specific // fields, is either a specific field name, a list of field // names, or the keyword ALL Server-handle: // IANA unique identifier for the polling server. // this handle may optionally be cached by the polled // server to announce future changes Host-Name: // Host name of the polling server. Weider, et al Standards Track PAGE 9 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 Host-Port: // Port number of the polling server. Hierarchy: // This field indicates the relationship which the poller // bears to the pollee. Typical values might include // 'Topology', 'Geographical", or "Administrative" Description: // This field contains the DESCRIBE record of the // polling server Authentication-type: // Type of authentication used by poller, or NONE Authentication-data: // Data for authentication # END // This line must by used to terminate the poll message Note: For poll type CENTROID, the allowable values for Poll Scope are FULL and RELATIVE. Support of the FULL value is required, this provides a complete listing of the centroid or other forward information. RELATIVE indicates that these are the relative changes in the centroid since the last report to the polling server. For poll type QUERY, the allowable values for Poll Scope are a blank line, which indicates that all records are to be returned, or a valid WHOIS++ query, which indicates that just those records which satisfy the query are to be returned. N.B. Security considerations may require additional authentication for successful response to the Blank Line Poll Scope. This value has been included for server replication. A polling server may wish to index different types of information than the polled server has collected. The POLLED-FOR command will indicate which servers the polled server has contacted. Required/Optional Table Version-Number REQUIRED, value is 1.0 Type-Of-Poll REQUIRED, values CENTROID and QUERY are required Poll-scope REQUIRED If Type-of-poll is CENTROID, FULL is required, RELATIVE is optional If Type-of-poll is QUERY, Blank line is required, and WHOIS++-type queries are required Start-time OPTIONAL End-Time OPTIONAL Template REQUIRED Field REQUIRED Server-handle REQUIRED Host-Name REQUIRED Host-Port REQUIRED Hierarchy OPTIONAL Description OPTIONAL Authentication-Type: OPTIONAL Authentication-data: OPTIONAL Weider, et al Standards Track PAGE 10 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 Example of a POLL command: # POLL: Version-number: 1.0 Type-of-poll: CENTROID Poll-scope: FULL Start-time: 199501281030+0100 Template: ALL Field: ALL Server-handle: BUNYIP01 Host-Name: services.bunyip.com Host-Port: 7070 Hierarchy: Geographical # END 6.3. Centroid change report As the centroid change report contains nested multiply-occuring blocks, each multiply occurring block is surrounded *in this paper* by curly braces '{', '}'. These curly braces are NOT part of the syntax, they are for identification purposes only. The syntax of a Data: item is either a list of words, one word per line, or the keyword: ANY The keyword ANY as the only item of a Data: list means that any value for this field should be treated as a hit by the indexing server. The field Any-field: needs more explanation than can be given in the body of the syntax description below. It can take two values, True or False. If the value is True, the pollee is indicating that there are fields in this template which are not being exported to the polling server, but wishes to treat as a hit. Thus, when the polling server gets a query which has a term requesting a field not in this list for this template, the polling server will treat that term as a 'hit'. If the value is False, the pollee is indicating that there are no other fields for this template which should be treated as a hit. This field is required because the basic model for the WHOIS++ query syntax requires that the results of each search term be 'and'ed together. This field allows polled servers to export data only for non-sensitive fields, yet still get referrals of queries which contain sensitive terms. IMPORTANT: The data listed in the centroid must be in the ISO-8859-1 character set in this version of the indexing protocol. Use of any other character set is a violation of the protocol. Note that the base-level server is also specified to use ISO-8859-1 [Deutsch, et Weider, et al Standards Track PAGE 11 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 al, 1995]. # CENTROID-CHANGES Version-number: // version number of pollee's index software, used to // insure compatibility Start-time: // change list starting time, GMT End-time: // change list ending time, GMT Server-handle: // IANA unique identifier of the responding server Case-sensitive: // states whether data is case sensitive or case // insensitive. values are TRUE or FALSE Authentication-type: // Type of authentication used by pollee, or NONE Authentication-data: // Data for authentication Compression-type: // Type of compression used on the data, or NONE Size-of-compressed-data: // size of compressed data if compression // is used Operation: // One of 3 keywords: ADD, DELETE, FULL // ADD - add these entries to the centroid for this server // DELETE - delete these entries from the centroid of this // server // FULL - the full centroid as of end-time follows { // The multiply occurring template block starts here # BEGIN TEMPLATE Template: // a standard whois++ template name Any-field: // TRUE or FALSE. See beginning of 6.3 for explanation. { // the template contains multiple field blocks # BEGIN FIELD Field: // a field name within that template Data: // Either the keyword *ANY*, or // the word list itself, one per line, cr/lf terminated, // each line starting with a dash character ('-'). # END FIELD } // the field ends with END FIELD # END TEMPLATE } // the template block ends with END TEMPLATE # END CENTROID-CHANGES // This line must be used to terminate the // centroid change report For each template, all fields must be listed, or queries will not be referred correctly. Required/Optional table Version-number REQUIRED, value is 1.0 Start-time REQUIRED (even if the centroid type is FULL) End-time REQUIRED (even if the centroid type is FULL) Server-handle REQUIRED Case-Sensitive OPTIONAL Authentication-Type OPTIONAL Weider, et al Standards Track PAGE 12 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 Authentication-Data OPTIONAL Compression-type OPTIONAL Size-of-compressed-data OPTIONAL (even if compression is used) Operation OPTIONAL, if used, upport for all three values is required Tokenization-type OPTIONAL #BEGIN TEMPLATE REQUIRED Template REQUIRED Any-field REQUIRED #BEGIN FIELD REQUIRED Field REQUIRED Data REQUIRED #END FIELD REQUIRED #END TEMPLATE REQUIRED #END CENTROID-CHANGES REQUIRED Example: # CENTROID-CHANGES Version-number: 1.0 Start-time: 197001010000 End-time: 199503012336 Server-handle: BUNYIP01 # BEGIN TEMPLATE Template: USER Any-field: TRUE # BEGIN FIELD Field: Name Data: Patrik -Faltstrom -Malin -Linnerborg #END FIELD #BEGIN FIELD Field: Email Data: paf@bunyip.com -malin.linnerborg@paf.se # END FIELD # END TEMPLATE # END CENTROID-CHANGES 6.4 QUERY and POLLEES responses The response to a QUERY command is done in WHOIS++ format. Weider, et al Standards Track PAGE 13 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 6.5. Query referral When referrals are included in the body of a response to a query, each referral is listed in a separate SERVER-TO-ASK block as shown below. # SERVER-TO-ASK Version-number: // version number of index software, used to insure // compatibility Body-of-Query: // the original query goes here Server-Handle: // WHOIS++ handle of the referred server Host-Name: // DNS name or IP address of the referred server Port-Number: // Port number to which to connect, if different from the // WHOIS++ port number # END Required/Optional table Version-number REQUIRED, value should be 1.0 Body-of-query OPTIONAL Server-Handle REQUIRED Host-Name REQUIRED Port-Number OPTIONAL, must be used if different from port 63 Example: # SERVER-TO-ASK Version-Number: 1.0 Server-Handle: SUNETSE01 Host-Name: sunic.sunet.se Port-Number: 63 # END 7: Reply Codes In addition to the reply codes listed in [Deutsch 95] for the basic WHOIS++ client/server interaction, the following reply codes are used in version 1.0 of this protocol. 113 Requested method not available Unable to provide a requested compression method. Contacted server will send requested data in different format. 227 Update request acknowledged A DATA-CHANGED transmission has been accepted and logged for further action. Weider, et al Standards Track PAGE 14 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 503 Required attribute missing A REQUIRED attribute is missing in an interaction. 504 Desired server unreachable The desired server is unreachable. 505 Desired server unavailable The desired server fails to respond to requests, but host is still reachable. 8. References [Deutsch 95] Deutsch, et al., "Architecture of the WHOIS++ service", RFC 1835, August 1995. [Faltstrom 95] Faltstrom, P., et al., "How to Interact with a WHOIS++ Mesh, RFC 1914, February 1996. 9. Security Considerations Security issues are not discussed in this memo. Weider, et al Standards Track PAGE 15 top

RFC 1913 Architecture of the Whois++ Index Service February 1996 10. Authors' Addresses Chris Weider Bunyip Information Systems, Inc. 310 St. Catherine St. West Montreal, PQ H2X 2A1 CANADA Phone: +1-514-875-8611 Fax: +1-514-875-6134 EMail: clw@bunyip.com Jim Fullton MCNC Center for Communications Post Office Box 12889 3021 Cornwallis Road Research Triangle Park North Carolina 27709-2889 Phone: 410-795-5422 Fax: 410-795-5422 EMail: fullton@cnidr.org Simon Spero EMail: ses@eit.com Weider, et al Standards Track PAGE 16 top

Architecture of the Whois++ Index Service RFC TOTAL SIZE: 33743 bytes PUBLICATION DATE: Friday, February 16th, 1996 LEGAL RIGHTS: The IETF Trust (see BCP 78)


RFC-ARCHIVE.ORG

© RFC 1913: The IETF Trust, Friday, February 16th, 1996
© the RFC Archive, 2024, RFC-Archive.org
Maintainer: J. Tunnissen

Privacy Statement