Efficient Parallel Set-Similarity Joins Using MapReduce Rares Vernica, Michael J. Carey, Chen Li Speaker : Razvan Belet Outline • Motivating Scenarios • Background Knowledge • Parallel Set-Similarity Join – Self Join – R-S Join • Evaluation • Conclusions • Strengths & Weaknesses Scenario: Detecting Plagiarism • Before publishing a Journal, editors have to make sure there is no plagiarized paper among the hundreds of papers to be included in the Journal Scenario: Near-duplicate elimination • The archive of a search engine can contain multiple copies of the same page • Reasons: re-crawling, different hosts holding the same redundant copies of a page, etc. Problem Statement Problem Statement: Given two collections of objects/items/records, a similarity metric sim(o1,o2) and a threshold λ , find the pairs of objects/items/records satisfying sim(o1,o2)> λ Solution: • Similarity Join Motivation(2) • Some of the collections are enormous: – Google N-gram database : ~1trillion records – GeneBank : 416GB of data – Facebook : 400 million active users Try to process this data in a parallel, distributed way => MapReduce Outline • Motivating Scenarios • Background Knowledge • Parallel Set-Similarity Join – Self Join – R-S Join • Evaluation • Conclusions Background Knowledge • Set-Similarity Join • Join • Similarity Join • Set-Similarity Join Background Knowledge: Join • Logical operator heavily used in Databases • Whenever it is needed to associate records in 2 tables => use a JOIN • Associates records in the 2 input tables based on a predicate (pred) Consider this information need: for each employee find the department he works in Table Employees LastName DepartmentID Rafferty 31 Jones 33 Steinberg 33 Robinson 34 Smith 34 John NULL Table Departments DepartmentID DepartmentName 31 Sales 33 Engineering 34 Clerical 35 Marketing Background Knowledge: Join • Example :For each employee find the department he works in EMPLOYEES LastName DepID Rafferty 31 Jones 33 Steinberg 33 Robinson 34 Smith 34 John NULL DEPARTMENTS JOINpred Department ID DepartmentNa me 31 Sales 33 Engineering 34 Clerical 35 Marketing pred: EMPLOYEES.DepID= DEPARTMENTS.DerpartmentI D JOIN RESULT LastName DepartmentName Rafferty Sales Jones Engineering Steinberg Engineering … … Background Knowledge: Similarity Join • Special type of join, in which the predicate (pred) is a similarity metric/function: sim(obj1,obj2) T1: T2: a b c … … … … … ... Similarity Joinpred d e c pred: sim(T1.c,T2.c)>threshold … … … … … ... a b c d e … … … … … … … … • Return pair (obj1, ob2) if pred holds: sim(obj1,obj2) > threshold … … … ... … … … … … … … … Background Knowledge: Similarity Join • Examples of sim(obj1,obj2) functions: # of common words sim(paper1,paper2) = # totalwords in the2 papers sim(Si, Tj) |SiTj| , |SiTj| Si, most common words in page i Tj, most common words in page j Similarity Join • sim(obj1,obj2) obj1,obj2 : documents, records in DB tables, user profiles, images, etc. • Particular class of similarity joins: (string/text-) similarity join:obj1, obj2 are strings/texts a b c … … … … … … … … … … … ... Name John W. Smith Marat Safin Rafael P. Nadal … SimilarityJoinpred pred: sim(T1.Name, T2.Name) > 2 d e … … … … Name … Smith, John … Safin, Marat Michailowitsch … Nadal , Rafael Parera ... …. sim(T1.Name,T2.Name)=#common words • Many real-world application => of particular interest Set-Similarity Join(SSJoin) • SSJoin: a powerful primitive for supporting (string-)similarity joins • Input: 2 collections of sets • Goal: Identify all pairs of highly similar sets {word1,word2 ….…. wordn} S1={… } S2={… } …. Sn={… } SSJoinpred pred: sim(Si,Ti)>0.3 |SiTi | sim(Si, Ti) |SiTi | T1={…} T2={…} … Tn={…} {word1,word2 ….…. wordn} Set-Similarity Join • How can a (string-)similarity join be reduced to a SSJoin? SSJoin BasedOn SimilarityJoin • Example: a b c … … … … … … … … Name … {John, W., Smith} … {Marat, Safin} … {Rafael, P., Nadal} ... … SSJoinpred d e Name … … … … {Smith, John} {Safin, Marat, Michailowitsch} {Nadal , Rafael, Parera} …. … … … ... pred: sim(T1.Name, T2.Name) > 0.5 sim(Si, Ti) |SiTi | |SiTi | Set-Similarity Join • Most SSJoin algorithms are signature-based: INPUT: Set collections R and S and threshold λ 1. For each r R, generate signature-set Sign(r) Filtering phase 2. For each s S, generate signature-set Sign(s) 3. Generate all candidate pairs (r, s), rR,sS satisfying Sign(r) ∩ Sign(s) 4. Output any candidate pair (r, s) satisfying Sim(r, s) ≥ λ. Post-filtering phase Set-Similarity Join • Signatures: – Have a filtering effect: SSJoin algorithm compares only candidates not all pairs (in post-filtering phase) – Give the efficiency of the SSJoin algorithm: the smaller the number of candidate pairs, the better – Ensure correctness: Sign(r) ∩ Sign(s) Sim(r, s) ≥ λ; , whenever Set-Similarity Join : Signatures Example • One possible signature scheme: Prefix-filtering • Compute Global Ordering of Tokens: Marat …W. Safin ... Rafael ... Nadal ...P. … Smith …. John • Compute Signature of each input set: take the prefix of length n a b c … … … … … … … … Name … {John, W., Smith} … {Marat, Safin} … {Rafael, P., Nadal} ... … Sign({John, W., Smith})=[W., Smith] Sign({Marat,Safin})=[Marat, Safin] Sign({Rafael, P., Nadal})=[Rafael,Nadal] Set-Similarity Join • Filtering Phase: Before doing the actual SSJoin, cluster/group the candidates a b c Name … … … {John, W., Smith} … … … {Marat, Safin} {Rafael, P., Nadal} … … ... … d e Name … … {Smith, John} … … {Safin,Marat,Michailowitsc} {Nadal , Rafael, Parera} … ... …. … cluster/bucket1 cluster/bucket2 cluster/bucketN • Run the SSjoin on each cluster => less workload Outline • Motivating Scenarios • Background Knowledge • Parallel Set-Similarity Join – Self Join – R-S Join • Evaluation • Conclusions • Strengths & Weaknesses Parallel Set-Similarity Join • Method comprises 3 stages: Compute data statistics for good signatures Stage I: Token Ordering Group candidates based on signature & Compute SSJoin Stage II RID-Pair Generation Generate actual pairs of joined records Stage III: Record Join Explanation of input data • RID = Row ID • a : join column •“A B C” is a string: •Address: “14th Saarbruecker Strasse” •Name: “John W. Smith” Stage I: Data Statistics Compute data statistics for good signatures Stage I: Token Ordering Basic Token Ordering Group candidates based on signature & Compute SSJoin Stage II RID-Pair Generation One Phase Token Ordering Generate actual pairs of joined records Stage III: Record Join Token Ordering • Creates a global ordering of the tokens in the join column, based on their frequency a RID 1 2 Global Ordering: (based on frequency) A B D AA BBDAE E 1 D 2 b c … … … … B 3 A 4 Basic Token Ordering(BTO) • 2 MapReduce cycles: – 1st : computing token frequencies – 2nd: ordering the tokens by their frequencies Basic Token Ordering – 1st MapReduce cycle ,, map: • tokenize the join value of each record • emit each token with no. of occurrences 1 reduce: • for each token, compute total count (frequency) Basic Token Ordering – 2nd MapReduce cycle reduce(use only 1 reducer): map: • emits the value • interchange key with value One Phase Tokens Ordering (OPTO) • alternative to Basic Token Ordering (BTO): – Uses only one MapReduce Cycle (less I/O) – In-memory token sorting, instead of using a reducer OPTO – Details ,, map: • tokenize the join value of each record • emit each token with no. of occurrences 1 Use tear_down method to order the tokens in memory reduce: • for each token, compute count (frequency) Stage II: Group Candidates & Compute SSJoin Individual Tokens Grouping Compute data statistics for good signatures Stage I: Token Ordering Grouped Tokens Grouping Group candidates based on signature & Compute SSJoin Stage II RID-Pair Generation Basic Kernel PPJoin Generate actual pairs of joined records Stage III: Record Join RID-Pair Generation • scans the original input data(records) • outputs the pairs of RIDs corresponding to records satisfying the join predicate(sim) • consists of only one MapReduce cycle Global ordering of tokens obtained in the previous stage RID-Pair Generation: Map Phase • scan input records and for each record: – project it on RID & join attribute – tokenize it – extract prefix according to global ordering of tokens obtained in the Token Ordering stage – route tokens to appropriate reducer Grouping/Routing Strategies • Goal: distribute candidates to the right reducers to minimize reducers’ workload • Like hashing (projected)records to the corresponding candidate-buckets • Each reducer handles one/more candidate-buckets • 2 routing strategies: Using Individual Tokens Using Grouped Tokens Routing: using individual tokens (projected) record • Treats each token as a key token • For each record, generates a (key, value) pair for each of its prefix tokens: Example: • Given the global ordering: Token A B E D G C F Frequency 10 10 22 23 23 40 48 “A B C” => prefix of length 2: A,B => generate/emit 2 (key,value) pairs: • (A, (1,A B C)) • (B, (1,A B C)) Grouping/Routing: using individual tokens • Advantage: – high quality of grouping of candidates( pairs of records that have no chance of being similar, are never routed to the same reducer) • Disadvantage: – high replication of data (same records might be checked for similarity in multiple reducers, i.e. redundant work) Routing: Using Grouped Tokens • Multiple tokens mapped to one synthetic key (different tokens can be mapped to the same key) • For each record, generates a (key, value) pair for each the groups of the prefix tokens: Routing: Using Grouped Tokens Example: • Given the global ordering: Token A B E D G C F Frequency 10 10 22 23 23 40 48 “A B C” => prefix of length 2: A,B Suppose A,B belong to group X and C belongs to group Y => generate/emit 2 (key,value) pairs: • (X, (1,A B C)) • (Y, (1,A B C)) Grouping/Routing: Using Grouped Tokens • The groups of tokens (X,Y) are formed assigning tokens to groups in a Round-Robin manner Token A B E D G C F Frequency 10 10 22 23 23 40 48 AD F BG Group1 Group2 EC Group3 • Groups will be balanced w.r.t the sum of frequencies of token belonging to one specific group Grouping/Routing: Using Grouped Tokens • Advantage: – Replication of data is not so pervasive • Disadvantage: – Quality of grouping is not so high (records having no chance of being similar are sent to the same reducer which checks their similarity) RID-Pair Generation: Reduce Phase • This is the core of the entire method • Each reducer processes one/more buckets • In each bucket, the reducer looks for pairs of join attribute values satisfying the join predicate If the similarity of the 2 candidates >= threshold => output their ids and also their similarity Bucket of candidates RID-Pair Generation: Reduce Phase • Computing similarity of the candidates in a bucket comes in 2 flavors: • Basic Kernel : uses 2 nested loops to verify each pair of candidates in the bucket • Indexed Kernel : uses a PPJoin+ index RID-Pair Generation: Basic Kernel • Straightforward method for finding candidates satisfying the join predicate • Quadratic complexity : O(#candidates2) reduce: foreach candidate in bucket for each cand in bucket\{candidate} if sim(candidate,cand)>= threshold emit((candidateRID, candRID), sim) RID-Pair Generation:PPJoin+ • Uses a special index data structure • Not so straightforward to implement • Much more efficient reduce: probe PPJoinIndex with join attr value of current_candidate => a list RIDs satisfying the join predicate add the current_candidate to the PPJoinIndex Stage III: Generate pairs of joined records Compute data statistics for good signatures Stage I Group candidates based on signature & Compute SSJoin Generate actual pairs of joined records Stage II Basic Record Join Stage III One Phase Record Join Record Join • Until now we have only pairs of RIDs, but we need actual records • Use the RID pairs generated in the previous stage to join the actual records • Main idea: – bring in the rest of the each record (everything excepting the RID which we already have) • 2 approaches: – Basic Record Join (BRJ) – One-Phase Record Join (OPRJ) Record Join: Basic Record Join • Uses 2 MapReduce cycles – 1st cycle: fills in the record information for each half of each pair – 2nd cycle: brings together the previously filled in records Record Join: One Phase Record Join • Uses only one MapReduce cycle R-S Join • Challenge: We now have 2 different record sources => 2 different input streams • Map Reduce can work on only 1 input stream • 2nd and 3rd stage affected • Solution: extend (key, value) pairs so that it includes a relation tag for each record Outline • Motivating Scenarios • Background Knowledge • Parallel Set-Similarity Join – Self Join – R-S Join • Evaluation • Conclusions • Strengths & Weaknesses Evaluation • Cluster: 10-node IBM x3650, running Hadoop • Data sets: • DBLP: 1.2M publications • CITESEERX: 1.3M publication • Consider only the header of each paper(i.e author, title, date of publication, etc.) • Data size synthetically increased (by various factors) • Measure: • Absolute running time • Speedup • Scaleup Self-Join running time • Best algorithm: BTO-PKOPRJ • Most expensive stage: the RID-pair generation Self-Join Speedup • Fixed data size, vary the cluster size • Best time: BTO-PKOPRJ Self-Join Scaleup • Increase data size and cluster size together by the same factor • Best time: BTO-PKOPRJ R-S Join Performance • Mostly, the same behavior R-S Join Performance Outline • Motivating Scenarios • Background Knowledge • Parallel Set-Similarity Join – Self Join – R-S Join • Evaluation • Conclusions • Strengths & Weaknesses Conclusions • Efficient way of computing Set-Similarity Join • Useful in many data cleaning scenarios • SSJoin and MapReduce: one solution for huge datasets • Very efficient when based on prefix-filtering and PPJoin+ • Scales-up up nicely Strengths & Weaknesses • Strengths: – – – – More efficient than single-node/local SSJoin Failure safer than single-node SSJoin Uses powerful filtering methods (routing strategies) Uses PPJoinIndex (data structure optimized for SSJoin) • Weaknesses: – This implementation is applicable only to string-based input data – Supposes the dictionary and RID-pairs list fit in main memory – Repeated tokenization – Evaluation based on synthetically increased data Questions Thank you!