FIND Working Group J. Allen Internet Draft Bunyip Information Systems <draft-ietf-find-cip-arch-00.txt> Michael Mealling 9 June 1997 Network Solutions, Inc. Expire in six months The Architecture of the Common Indexing Protocol (CIP) 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 valid for a maximum of six months and 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 ftp.is.co.za (Africa), nic.nordu.net (Europe), munnari.oz.au (Pacific Rim), ds.internic.net (US East Coast), or ftp.isi.edu (US West Coast). Abstract The Common Indexing Protocol (CIP) is used to pass indexing information from server to server in order to facilitate query routing. Query routing is the process of redirecting and replicating queries through a distributed database system towards servers holding the desired results. This document describes the CIP framework, including it's architecture and the protocol specifics of exchanging indices. 1. Introduction 1.1. History and Motivation The Common Indexing Protocol (CIP) is an evolution and refinement of distributed indexing concepts first introduced in the Whois++ Directory Service [RFC1913, RFC1914]. While indexing proved useful in that system to promote query routing, the centroid index object which is passed among Whois++ servers is specifically designed for template-based databases searchable by token-based matching. With alternative index objects, the index-passing technology will prove useful to many more application domains, not simply Directory Services and those applications which can be cast into the form of template collections. The indexing part of Whois++ is integrated with the data access protocol. The goal in designing CIP is to extract the indexing portion of Whois++, while abstracting the index objects to apply more broadly to information retrieval. In addition, another kind of technology reuse has been undertaken by converting the ad-hoc data representations used by Whois++ into structures based on the MIME specification for structured Internet mail. Whois++ used a version number field in centroid objects to facilitate future growth. The initial version was "1". Version 1 of CIP (then embedded in Whois++, and not referred to separately as CIP) had support for only ISO-8895-1 characters, and for only the centroid index object type. Version 2 of the Whois++ centroid was used in the Digger software by Bunyip Information Systems to notify recipients that the centroid carried extra character set information. Digger's centroids can carry UTF-8 encoded 16-bit Unicode characters, or ISO-8859-1 characters, determined by a field in the headers. This specification is for CIP version 3 (CIPv3). Version 3 is a major overhaul to the protocol, though through a short negotiation sequence, CIP version 3 and earlier servers can interoperate in an index-passing mesh. 1.2 CIP's place in the Information Retrieval world CIP facilitates query routing. CIP is a protocol used between servers in a network to pass hints which make data access by clients at a later date more efficient. Query routing is the act of redirecting and replicating queries through a distributed database system towards the servers holding the actual results via reference to indexing information. CIP is a "backend" protocol -- it is implemented in and "spoken" among only network servers. These same servers must also speak some kind of data access protocol to communicate with clients. During query resolution in the native protocol implementation, the server will refer to the indexing information collected by the CIP implementation for guidance on how to route the query. Data access protocols used with CIP must have some provision for control information in the form of a referral. The syntax and semantics of these referrals are outside the scope of this specification. 2. Related Documents This document is one of three documents. This document describes the fundamental concepts and framework of CIP. The document "MIME Object Definitions for the Common Indexing Protocol" [CIP-MIME] describes the MIME objects that make up the items that are passed by the transport system. Requirements and examples of several transport systems are layed out in the "bla" [CIP-TRANSPORT] document. A second set of document describe the various specifications for specific index types. 3. Architecture 3.1 CIP in the Information Retrieval World 3.1.1 Information Retrieval in the Abstract In order to better understand how CIP fits into the information retrieval world, we need to first understand the unifying abstract features of existing information retrieval technology. Next, we discuss why adding indexing technology to this model results in a system capable of query routing, and why query routing is useful. An abstract view of the client/server data retrieval process includes data sets and data access protocols. An individual server is responsible for handling queries over a fixed domain of data. For the purposes of CIP, we call this domain of data the dataset. Clients make searches in the dataset and retrieve parts of it via a data access protocol. There are many data access protocols, each optimized for the data in question. For instance, LDAP and Whois++ are access protocols that reflect the needs of the directory services application domain. Other data access protocols include HTTP and Z39.50. 3.1.2 Indexing Information Facilitates Query Routing The above description reflects a world without indexing, where no server knows about any other server. In some cases (as with X.500 referrals, and HTTP redirects) a server will, as part of it's reply, implicate another server in the process of resolving the query. However, those servers generate replies based solely on their local knowledge. When indexing information is introduced into a server's local database, the server now knows not only answers based on the local dataset, but also answers based on external indices. These indices come from peer servers, via an indexing protocol. CIP is one such indexing protocol. Replies based on index information may not be the complete answer. After all, an index is not a replicated version of the remote dataset, but a possibly reduced version of it. Thus, in addition to giving complete replies from the local dataset, the server may give referrals to other datasets. These referrals are the core feature necessary for effective query routing. When CIP is used to pass indices from server to server, they make a kind of investment. At the cost of some resources to create, transmit and store the indices, query routing becomes possible. Query Routing is the process of replicating and moving a query closer to datasets which can satisfy the query. In some distributed systems, widely distributed searches must be accomplished by replicating the query to all sub-datasets. This approach can be wasteful of resources both in the network, and on the servers, and is thus sometimes explicitly disabled. Using indexing in such a system opens the door to more efficient distributed searching. While CIP-equipped servers provide the referrals necessary to make query routing work, it's always the client's responsibility to collate, filter, and chase the referrals it receives. This gives the end-user (or agent, in the case that there's no human user involved in the search) greatest control over the query resolution process. The cost of the added client complexity is weighed against the benefits of total control over query resolution. In some cases, it may also be possible to decouple the referral chasing from the client by introducing a proxy, allowing existing simple clients to make use of query routing. Such a proxy would transparently resolve referrals into concrete results before returning them to the simple-minded client. 3.1.3 Abstracting the CIP index object As useful as indices seem, the fact remains that not all queries can benefit from the same type of index. For example, say the index consists of a simple list of keywords. With such an index, it is impossible to answer queries about whether two keywords were near one another, or if a keyword was present in a certain context (for instance, in the title). Because of the need for application domain specific indices, CIP index objects are abstract; they must be defined by a separate specification. The basic protocols for moving index objects are widely applicable, but the specific design of the index, and the structure of the mesh of servers which pass a particular type of index is dependent on the application domain. This document describes only the protocols for moving indices among servers. Companion documents describe initial index objects. The requirements that index type specifications must address are specified in the [CIP-MIME] document. 3.2 Architectural Details CIP implements index passing, providing the forward knowledge necessary to generate the referrals used for query routing. The core of the protocol is the index object. In the following sections, the structure of the index objects themselves is presented. Next, how and why indices are passed from server to server is discussed. Finally, the circumstances under which a server may synthesize an index object based on incoming ones are discussed. 3.2.1 The CIP Index Object A CIP index object is composed of two parts, the header and the payload. The header contains metadata necessary to process and make use of the index object being transmitted. The actual index resides in the payload. Three particular headers warrant specific mention at this point. The "type" of the index object selects one of many distinct CIP index object specifications which define exactly how the index blocks are to be created, parsed and used to facilitate query routing. Another header of note is the "DSI", or Dataset Identifier, which uniquely identifies the dataset from which the index was created. Another header that is crucial for generating referrals is the "Base-URI". The URI (or URI's) contained in this header form the basis of any referrals generated based on this index block. ***The URI is also used as input during the index aggregation process to constrain the kinds of aggregation possible, due to multiprotocol constraints.** The exact syntax of these headers will be specified in the CIP MIME specification document [CIP-MIME]. [Editors Note: the next to last sentence needs to be explained more. Can someone who thinks they understand it well enough explain it provide me with some better text?] The payload is opaque to CIP itself. It is defined exclusively by the index object specification associated with the object's type header. Specifications on how to parse and use the payload are published separately as "CIP index object specifications". This abstract definition of the index object forms the basis of CIP's applicability to indexing needs across multiple application domains. A precise definition of the content and form of a CIP index block can be found in the Protocol document [CIP-MIME] 3.2.2 Moving Index Objects: How to Build a Mesh Indices are transmitted among servers participating in a CIP mesh. By distributing this information in anticipation of a query, efficient, accurate query routing is possible at the time a query arrives. A CIP mesh is a set of CIP servers which pass indices of the same type among themselves. Typically, a mesh is arranged in a hierarchical tree fashion, with servers nearer the root of the tree having larger and more comprehensive indices. See Figure 1. However, a CIP mesh is explicitly allowed to have lateral links in it, and there may be more than one part of the mesh that has the properties of a "root". Mesh administrators are encouraged to avoid loops in the system, but they are not obliged to maintain a strict tree structure. Clients wishing to completely resolve all referrals they receive should protect against referral loops while attempting to traverse the mesh to avoid wasting time and network resources. See the section on "Navigating the Mesh" for a discussion of this. base level index index directory servers servers servers for for base level lower-level servers index servers _______ | | | A |__ |_______| \ _______ \---CIP----| | _______ | D |__ | | /---CIP----|_______| \ ------ | B |__/ \--CIP------| | |_______| | F | /--CIP------|______| / _______ _______ / | | | |- | C |-------CIP----| E | |_______| |_______|- | \ r \ _______ e \ ______ | | f \--CIP-----| | | G |-------CIP---------e------------------| H | |_______| r |______| \--referral---| r --referral-/ | a | | l | \ 3 | 2 | 1 \--------/ | | | client | | | -------- Figure 1: Sample layout of the Index Service mesh All indices passed in a given mesh are assumed, as of this writing, to be of the same type (i.e. governed by the same CIP index object specification). It may be possible to create gateways between meshes carrying different index objects, but at this time that process is undefined and declared to be outside the scope of this specification. [ Note to reviewers: What to do if you get an index object you don't understand? Is it OK to drop it? Is it OK to ignore it for referral generation, but pass it through with the outgoing indices? ] Experience suggests that this index passing activity should take place among CIP servers as a parallel (and possibly lower-priority) job to their primary job of answering queries. Index objects travel among CIP servers by protocol exchanges explicitly defined in this document, not via the server's native protocol. This distinction is important, and bears repeating: Queries are answered (and referrals are sent) via the native data access protocol. Index objects are transferred via alternative means, as defined by this document. When two servers cooperate to move indexing information, the pair are said to be in a "polling relationship". The server that holds the data of interest, and generates the index is called the "polled server". The other server, which is the one that collects the generated index, is the "polling server". In a polling relationship, the polled server is responsible for notifying the polling server when it has a new index that the polling server might be interested in. In response, the polling server may immediately pick up the index object, or it may schedule a job to pick up a copy of the new index at a more convenient time. But, a polling server is not required to wait on the polled server to notify it of changes. The polling server can request a new index at any time. Independent of the symmetric polling relationship, there's another way that servers can pass indices using CIP. In an "index pushing" relationship, a CIP server simply sends the index to a peer whenever necessary, and allows the receiver to handle the index object as it chooses. The receiving server may refuse it, may accept is, then silently discard it, may accept only portions of it (by accepting it as is, then filtering it), or may accept it without question. The index pushing relationship is intended for use by dumb leaf nodes which simply want to make their index available to the global mesh of servers, but have no interest in implementing the complete CIP transaction protocol. It lowers the barriers to entry for CIP leaf nodes. For more information on participating in a CIP mesh in this restricted manner, see the section below on "Protocol Conformance". CIP index passing operations take place across a reliable transport mechanisms, including both TCP connections, and Internet mail messages. The precise mechanisms are described in the Transport document [Transport] 3.2.3 Index Object Synthesis From the preceding discussion, it should be clear that indexing servers read and write index objects as they pass them around the mesh. However, a CIP server need not simply pass the inbound indices through as the outbound ones. While it's always permissible to pass an index object through to other servers, a server may choose to aggregate two or more of them, thereby reducing redundancy in the index, at the cost of longer referral chains. A basic premise of index passing is that even while collapsing a body of data into an index by lossy compression methods, hints useful to routing queries will survive in the resulting index. Since the index is not a complete copy of the original dataset, it contains less information. Index objects can be passed along unchanged, but as more and more information collects in the resulting index object, redundancy will creep in again, and it may prove useful to apply the compression again, by aggregating two or more index objects into one. This kind of aggregation should be performed without compromising the ability to correctly route queries while avoiding excessive numbers of missed results. The acceptable likelihood of false negatives must be established on a per-application-domain basis, and is controlled by the granularity of the index and the aggregation rules defined for it by the particular specification. However, when CIP is used in a multi-protocol application domain, such as a Directory Service (with contenders including Whois++, LDAP, and Ph), things get significantly trickier. The fundamental problem is to avoid forcing a referral chain to pass through part of the mesh which does not support the protocol by which that client made the query. If this ever happens, the client loses access to any hits beyond that point in the referral chain, since it cannot resolve the referral in its native data access protocol. This is a failure of query routing, which should be avoided. In addition to multi-protocol considerations, server managers may choose not to allow index object aggregation for performance reasons. As referral chains lengthen, a client needs to perform more transactions to resolve a query. As the number of transactions increases, so do the user-perceived delays, the system loads, and the global bandwidth demands. In general, there's a tradeoff between aggressive aggregation (which leads to reductions in the indexing overhead) and aggressive referral chain optimization. This tradeoff, which is also sensitive to the particular application domain, needs to be explored more in actual operational situations. Conceptually, a CIP index server has several index objects on hand at any given time. If it holds data in addition to indexing information, the server has an index object formed from its own data, called the "local index". It may have one or more indices from remote servers which it has collected via the index passing mechanisms. These are called "inbound indices". Implementor's Note: It may not be necessary to keep all of these structures intact and distinct in the local database. It is also not required to keep the outbound index (or indices) built and ready to distribute at all times. The previous paragraph merely introduces a useful model for expressing the aggregation rules. Implementors are free to model index objects internally however they see fit. The following two rules control how a CIP server formulates it's outgoing indices: 1. An index server may pass any of the index objects in its local index and its inbound indices through unchanged to polling servers. 2. If and only if the following three conditions are true, an index server can aggregate two or more index objects into a single new index object, to be added to the set of outbound indices. a. Each index object to be aggregated covers exactly the same set of protocols, as defined by the scheme component of the Base-URI's in each index object. b. The index server supports every one of the data access protocols represented by the Base-URI's in the index objects to be aggregated. c. The specification for the index object type specified by the type header of the index objects explicitly defines the aggregation operation. The resulting index object must have Base-URI's characteristic of the local server for each protocol it supports. The outgoing objects should have the DSI of the local server. 4. Navigating the mesh With the CIP infrastructure in place to manage index objects, the only problem remaining is how to successfully use the indexing information to do efficient searches. CIP facilitates query routing, which is essentially a client activity. A client connects to one server, which redirects the query to servers "closer to" the answer. This redirection message is called a referral. 4.1 The Referral The concept of a referral and the mechanism for deciding when they should be issued is described by CIP. However, the referral itself must be transferred to the client in the native protocol, so its syntax is not directly a CIP issue. The mechanism for deciding that a referral needs to be made and generating that referral resides in the CIP implementation in the server. The mechanism for sending the referral to the client resides in the server's native protocol implementation. A referral is made when a search against the index objects held by the server shows that there may be hits available in one of the datasets represented by those index objects. If more that one index object indicates that a referral must be generated to a given dataset, the server should generate only one referral to the given dataset, as the client may not be able to detect duplicates. Though the format of the referral is dependent on the native protocol(s) of the CIP server, the baseline contents of the referral are constant across all protocols. At the least, a DSI and a URI must be returned. The DSI is the DSI associated with the dataset which caused the hit. This must be presented to the client so that it can avoid referral loops. The Base-URI parameter which travels along with index objects is used to provide the other required part of a referral. The additional information in the Base-URI may be necessary for the server receiving the referred query to correctly handle it. A good example of this is an LDAP server, which needs a base X.500 distinguished name from which to search. When an LDAP server sends a centroid-format index object up to a CIP indexing server, it sends a Base-URI along with the name of the X.500 subtree for which the index was made. When a referral is made, the Base-URI is passed back to the client so that it can pass it to the original LDAP server. As usual, in addition to sending the DSI, a DSI-Description header can be optionally sent. Because a client may attempt to check with the user before chasing the referral, and because this string is the friendliest representation of the DSI that CIP has to offer, it should be included in referrals when available (i.e. when it was sent along with the index object). 4.2 Cross-protocol Mappings Each data access protocol which uses CIP will need a clearly defined set of rules to map queries in the native protocol to searches against an index object. These rules will vary according to the data domain. In principle, this could create a bit of a scaling difficulty; for N protocols and M data domains, there would be N x M mappings required. In practice, this should not be the case, since some access protocols will be wholly unsuited to some data domains. Consider for example, a LDAP server trying to make a search in an index object composed from unorganized text based pages. What would the results be? How would the client make sense of the results? However, as pre-existing protocols are connected to CIP, and as new ones are developed to work with CIP, this issue must be examined. In the case of Whois++ and the CENTROID index type, there is an extremely close mapping, since the two were designed together. When hooking LDAP to the CENTROID index type, it will be necessary to map the attribute names used in the LDAP system to attribute names which are already being used in the CENTROID mesh. It will also be necessary to tokenize the LDAP queries under the same rules as the CENTROID indexing policy, so that searches will take place correctly. These application- and protocol-specific actions must be specified in the index object specification, as discussed in the CIP-MIME [CIP-MIME] document. 4.3 Moving through the mesh From a client's point of view, CIP simply pushes all the "hard work" onto its shoulders. After all, it's the client which needs to track down the real data. While this is true, it's very misleading. Because the client has control over the query routing process, the client has total control over the size of the result set, the speed with which the query progresses, and the depth of the search. The simplest client implementation simply provides referrals to the user in a raw, ready-to-reuse form, without attempting to follow them. For instance, one Whois++ client, which interacts with the user via a Web-based form, simply makes referrals into HTML hypertext links. Encoded in the link via the HTML forms interface GET encoding rules is the data of the referral: the hostname, port, and query. If a user chooses to follow the referral link, they execute a new search on the new host. A more savvy client might present the referrals to the user and ask which should be followed. And, assuming appropriate limits were placed on search time, and bandwidth usage, it might be reasonable to program a client to follow all referral automatically. When following all referrals, a client must show a bit of intelligence. Remember that the mesh is defined as an interconnected graph of CIP servers. This graph may have cycles, which could cause an infinite loop of referrals, wasting the servers' time and the client's too. When faced with the job of tacking down all referrals, a client must use some form of a mesh traversal algorithm. Such an algorithm has been documented for use with Whois++ in RFC-1914. The same algorithm can be easily used with this version of CIP. In Whois++ the equivalent of a DSI is called a handle. With this substitution, the Whois++ mesh traversal algorithm works unchanged with CIP. Finally, the mesh entry point (i.e. the first server queried) can have an impact on the success of the query. To avoid scaling issues, it is not acceptable to use a single "root" node, and force all clients to connect to it. Instead, clients should connect to a reasonably well connected (with respect to the CIP mesh, not the Internet infrastructure) local server. If no match can be made from this entry point, the client can expand the search by asking the original server who polls it. In general, those servers will have a better "vantage point" on the mesh, and will turn up answers that the initial search didn't. The mechanism for dynamically determining the mesh structure like this exists, but it not documented here for brevity. See RFC-1913 for more information on the POLLED-BY and POLLED-FOR commands. It still should be noted that, while these mesh operations are important to optimizing the searches that a client should make, the client still speaks its native protocol. This information must be communicated to the client without causing the client to have to understand CIP. 5. Security Considerations In this section, we discuss the security considerations necessary when making use of this specification. There are at least two levels at which security considerations come into play. Indexing information can leak undesirable amounts of proprietary information, unless carefully controlled. At a more fundamental level, the CIP protocol itself requires external security services to operate in a safe manner. Both topics are covered below. 5.1 Secure Indexing CIP is designed to index all kinds of data. Some of this data might be considered valuable, proprietary, or even highly sensitive by the data maintainer. Take, for example, a human resources database. Certain public bits of data, in moderation, can be very helpful for a company to make public. However, the database in its entirety is a very valuable asset, which the company must protect. Much experience has been gained in the directory service community over the years as to how best to walk this fine line between completely revealing the database and making useful pieces of it available. Another example where security becomes a problem is for a data publisher who'd like to participate in a CIP mesh. The data that publisher creates and manages is the prime asset of the company. There is a financial incentive to participate in a CIP mesh, since exporting indices of the data will make it more likely that people will search your database. (Making profit off of the search activity is left as an exercise to the entrepreneur.) Once again, the index must be designed carefully to protect the database while providing a useful synopsis of the data. One of the basic premises of CIP is that data providers will be willing to provide indices of their data to peer indexing servers. Unless they are carefully constructed, these indices could constitute a threat to the security of the database. Thus, security of the data must be a prime consideration when developing a new index object type. The risk of reverse engineering a database based only on the index exported from it must be kept to a level consistent with the value of the data and the need for fine-grained indexing. Acknowledgments Thanks to the many helpful members of the FIND working group for discussions leading to this specification. Specific acknowledgment is given to Jeff Allen formerly of Bunyip Information Systems. His original version of these documents helped enormously in crystallizing the debate and consensus. Most of the actual text in this document was originally authored by Jeff. Author's Address Jeff R. Allen Michael Mealling Bunyip Information Systems, Inc. Network Solutions, Inc. 310 Ste-Catherine West, Suite 300 505 Huntmar Park Drive Montreal, Quebec H2X 2A1 Herndon, VA 22070 Canada Phone: +1-514-875-8611 Phone: (703) 742-0400 EMail: jeff@bunyip.com Email: michael.mealling@RWhois.net References [RFC1913] Weider, C., Fullton, J., S. Spero, "Architecture of the Whois++ Index Service", Bunyip, CNIDR, EIT, February 1996. [RFC1914] Faltstrom, P., Schoultz, R., C. Weider, "How to Interact with a Whois++ Mesh", Bunyip, KTHNOC, February 1996. [CIP-MIME] Allen, J., M. Mealling, "MIME Object Definitions for the Common Indexing Protocol (CIP)", IETF FIND WG, June 1997. [CIP-TRANSPORT] Allen, J., P. Leach, "CIP Transport Protocols", WebTV, Microsoft, June 1997. Appendix A: Glossary application domain: A problem domain to which CIP is applied which has indexing requirements which are not subsumed by any existing problem domain. Separate application domains require separate index object specifications, and potentially separate CIP meshes. See index object specification. centroid: An index object type used with Whois++. In CIP versions before version 3, the index was not extensible, and could only take the form of a centroid. A centroid is a list of (template name, attribute name, token) tuples with duplicate headers removed. dataset: A collection of data (real or virtual) over which an index is created. When a CIP server aggregates two or more indices, the resultant index represents the index from a "virtual dataset", spanning the previous two datasets. Dataset Identifier: An identifier chosen from any part of the ISO/CCITT OID space which uniquely identifies a given dataset among all datasets indexed by CIP. DSI: See Dataset Identifier. DSI-description: A human readable string optionally carried along with DSI's to make them more user-friendly. See dataset Identifier. index object: The embodiment of the indices passed by CIP. An index object consists of some control attributes and an opaque payload. index object specification: A document describing an index object type for use with the CIP system described in this document. See index object and payload. index pushing: The act of presenting, unsolicited, an index to a peer CIP server. MIME: see Multipurpose Internet Mail Extensions Multipurpose Internet Mail Extensions: A set of rules for encoding Internet Mail messages that gives them richer structure. CIP uses MIME rules to simplify object encoding issues. MIME is specified in RFC-1521 and RFC-1522. payload: The application domain specific indexing information stored inside an index object. The format of the payload is specified externally to this document, and depends on the type of the containing index object. polled server: A CIP server which receives a request to generate and pass an index to a peer server. polling server: A CIP server which generates a request to a peer server for its index. referral chain: The set of referrals generated by the process of routing a query. See query routing. query routing: Based on reference to indexing information, redirecting and replicating queries through a distributed database system towards the servers holding the actual results.
Datatracker
Report a datatracker bug
draft-ietf-find-cip-arch-00
Internet-Draft
| Document | Document type |
This is an older version of an Internet-Draft that was ultimately published as RFC 2651.
Expired & archived
|
|
|---|---|---|---|
| Select version | |||
| Compare versions | |||
| Author | |||
| RFC stream |
|
||
| Other formats | |||
| Additional resources | Mailing list discussion |