dbxplorer: a system for keyword b ased search over relational databases

Click here to load reader

Upload: faye

Post on 24-Feb-2016

55 views

Category:

Documents


0 download

DESCRIPTION

DBXplorer: A System for Keyword B ased Search over Relational Databases. Sanjay Agrawal Microsoft Research Surajit Chaudhuri Microsoft Research Gautam Das Microsoft Research. Presented by: DEEP PANCHOLI 1000556121. Introduction. - PowerPoint PPT Presentation

TRANSCRIPT

DBXplorer: A system for keyword based search over relational databases

Sanjay Agrawal Microsoft ResearchSurajit Chaudhuri Microsoft ResearchGautam Das Microsoft ResearchDBXplorer: A System for Keyword Based Search over Relational DatabasesPresented by:DEEP PANCHOLI 1000556121

IntroductionThe two most common types of search are Structured Search and Keyword Based SearchExample of Structured Search http://autos.yahoo.com/A similar example is to search for books in booksellers database e.g. Books->Travel->Maps->Asia->RussiaWe all already know what is keyword based search and one example can be searching for Jim Gray on Microsoft Intranet to obtain matched rows

Introduction (cont)Problems faced with Keyword based search implementationNeed to know schemaNormalized databasesAvailability of indexesBuilt on the concept laid by BANKS paper explained in the last lecture.Symbol tablesCompacting the symbol tablesSearch requirementsOverview of DBXplorerDBXplorer returns all rows either from single table or from multiple tables, using FK-joins, such that each row has all the keywordsPublish1. Identify a database and tables and columns within it that are to be enabled for search2. Create auxiliary tables (Symbol Tables)Search1. Look up the Symbol table2. Searching in possible subsets of tables3. Construct and execute SQL statement and rank the results before displaying to userDifferent Symbol Table DesignsWe will only consider exact match problemTwo important levels of granularitiesColumn level granularity (Pub-Col)Cell level granularity (Pub-Cell)Table=AuthorsFnameLnameJohnMarshall JohnShawnShawnArcherJacquelynMarshallFactors that affect granularitySpace and time requirementsPub-Col is faster and occupies less spaceKeyword search performancePub-Col if there is an index on the columnEase of Symbol table maintenancePub-Col is easier to maintain as it contains updates only if there is addition of a new distinct valuesHence, the Pub-Col alternative is almost always better than Pub-Cell unless if certain columns contain no indexesIf an index is available for column, we should use Pub-Col granularityPub-Col representationStore simply as Keyword-ColIdAlternative is to use Hashvalue-ColId since storing keywords is wasteful as strings can be long and of varying lengthsCompression AlgorithmsFK-Comp: If column c1 is a subset of values in another column c2, we retain only values in c1CP-Comp: It is used when pairs of columns share common keywords but are not tied by FKPub-Col Algorithm

Search ComponentCommon step for all kinds of granularitiesIt makes use of join treesHence, if we join tables that occur in the join tree the resulting relation will contain all potential rows containing all keywords specified in the queryExample of graph treeFinally SQL query is generated and runThe result is then ranked before outputting. The basic approach is to rank them based on the number of joins involved which is similar to Banks approachSearch Algorithm

Case of Token matchesToken matches are matches in which keyword match with a token or a substring of attribute value

Pub-Prefix method efficiently enables token match capabilities by exploiting available B+ tree indexesSymbol table has entry (hash(k),T.C, P)

Case of Token matches (cont)Pub-Prefix method result is comparable to Pub-Cell method when the column width is small (i.e. less than 100 characters)For columns where strings are greater than hundreds of characters, Pub-Cell outperforms Pub-Prefix significantlyImportant issue is to determine the appropriate prefix length stored in symbol table. However, Pub-Prefix method is still being researched uponOther research is going on in field of stemming of query keywordsExperimental ResultsThe experiments were carried out on a 450MHz 256 MB Intel P-3 machine. There were 4 databases used for evaluation:TPC-H data of sizes from 100 to 500 MBUSR is Microsoft employee address DB of 130 MB with 19 tablesML is a 375 MB mailing list DB with 38 tablesKB is a 365 MB DB with 84 tables containing information on articles and help manuals on various shipped productsSystem Architecture for DBXplorer

Experimental Results (cont)In particular the authors show the following:Pub-Col is compact compared to Pub-CellPub-Col scales linearly with data size and is independent of data distributionPub-Prefix is compact compared to Pub-Cell and has a significantly better performance when full text indexes are not presentPub-Col and Pub-Cell symbol table size comparison

Symbol table publishing time comparison

Query performance

Other ObservationsIt was also noticed that search scales with number of query keywords. The query was varied with 2 to 10 keywords and still the average query time was between 1 to 1.3 secondsAlso, it was noticed that FK-Comp and CP-Comp reduce the size of Pub-Col by a factor of 0.45 to 0.90 depending on size of original tableHowever, it was noticed that compression added a negligible overhead on search performanceEffectiveness of Pub-Prefix methodThe Pub-prefix method was tested on workload consisting of 100 random keywords from character column of width 64 bytes in the KB database.It was noticed that the performance of Pub-Prefix increased with increase in Pub-Prefix length and gave the optimum performance at prefix-length of 8This is because as the length increases, beyond a certain limit the optimizer decides to scan the original table compared to index searchConclusionAlthough, we discussed only about a single database query, this technique can be applied to search multiple databases alsoDBXplorer is easy to use with any Database Management systemAs mentioned before, the Pub-Col alternative is the best when columns have indexes on them. A hybrid table can be created so that if there is an index for a column, we use Pub-Col granularity and if there is no index, we use Pub-Cell granularity