2. Query Optimization: Relational Queries to Data Mining Most people have Data from which they want information. So, most people need DBMSs whether they know it or not. A major component of any DBMS is the query processor. Queries can range from structure to unstructured: SELECT FROM WHERE Complex queries (nested, EXISTS.. ) FUZZY queries (e.g., BLAST searches,.. OLAP (rollup, drilldown, slice/dice.. Machine LearningData Mining Relational querying Simple Searching and aggregating Supervised - Classification Regression Unsupervised- Clustering Association Rule Mining Although we looked fairly closely at the structured end of this spectrum, much research is yet to be done on that end to solve the problem of delivering standard workload answers with low response times and high throughput (D. DeWitt, ACM SIGMOD’02 plenary symposium). On the Data Mining end, we have barely scratched the surface. (But those scratches have made the difference between becoming the world’s biggest corporation and filing for bankruptcy – Walmart vs. KMart)

3. Some Vertical (column-based) DBMSs Vertical partitioning has been studied within the context of both centralized database system as well as distributed ones. It is a good strategy when small numbers of columns are retrieved by most queries. The decomposition of a relation also permits a number of transactions to execute concurrently. Copeland et al presented an attribute level decomposition storage model (DSM) storing each column of a relational table into a separate binary table. The DSM showed great comparability in performance. Wong et al further took the advantage of encoding attribute values using a small number of bits to reduce the storage space. In this paper, we will decompose attributes of relational tables into bit position level, utilize SPJ query optimization strategy on them, store the query results in one relational table, finally data mine using our very good P-tree methods. Sensage, C-Store, BigTable, SybaseIQ and Vertica are commerical Vertical DBMS based on a columnar architecture. –(1) By vertical partitioning, we only need to read everything we need. This method makes hardware caching work really well and greatly increases the effectiveness of the I/O device. –(2) We encode attribute values into bit vector format, which makes compression easy to do. –(3) SPJ queries can be formulated as Boolean expressions, which facilitates fast implementation on hardware. –(4) Our model is fit not only for query processing but for data mining as well. G.Copeland, S. Khoshafian. A Decomposition Storage Model. Proc. ACM Int. Conf. on Management of Data (SIGMOD’85), pp.268-279, Austin, TX, May 1985. F. Chang, J. Dean, S Ghemawat et al, BigTable, 2006. M. Stonebraker, D. Abadi et al, C-Store, VLDB, 2005. Sensage www.sensage.com Sybase www.sybase.com Vertica www.vertica.com H. K. T. Wong, H.-F. Liu, F. Olken, D. Rotem, and L. Wong. Bit Transposed Files. Proc. Int. Conf. on Very Large Data Bases (VLDB’85), pp.448-457, Stockholm, Sweden, 1985.

4. SPJ Query Optimization Strategies - One-table Selections There are two categories of queries in one-table selections: Equality Queries and Range Queries. Most techniques [WLO+85, OQ97, CI98] used to optimize them employ encoding schemes – equality encoding and range encoding. Chan and Ioannidis [CI99] defined a more general query format called interval query. An interval query on attribute A is a query of the form “x≤A≤y” or “NOT (x≤A≤y)”. It can be an equality query or a range query when x or y satisfies different kinds of conditions. We defined interval P-trees in previous work [DKR+02], which is equivalent to the bit vectors of corresponding intervals. So for each restriction in the form above, we have one corresponding interval P-tree. The ANDing result of all the corresponding interval P-trees represents all the rows satisfy the conjunction of all the restriction in the where clause. [CI98] C.Y. Chan and Y. Ioannidis. Bitmap Index Design and Evaluation. Proc. ACM Intl. Conf. on Management of Data (SIGMOD’98), pp.355-366, Seattle, WA, June 1998. [CI99] C.Y. Chan and Y.E. Ioannidis. An Efficient Bitmap Encoding Scheme for Selection Queries. Proc. ACM Intl. Conf. on Management of Data (SIGMOD’99), pp.216-226, Philadephia, PA, 1999. [DKR + 02] Q. Ding, M. Khan, A. Roy, and W. Perrizo. The P-tree algebra. Proc. ACM Symposium Applied Computing (SAC 2002), pp.426-431, Madrid, Spain, 2002. [OQ97]P. O’Neill and D. Quass. Improved Query Performance with Variant Indexes. Proc. ACM Int. Conf. on Management of Data (SIGMOD’97), pp.38-49, Tucson, AZ, May 1997.

5. Vertical Select-Project-Join (SPJ) Queries A Select-Project-Join query has joins, selections and projections. Typically there is a central fact relation (e.g., Enrollments or E below) to which several dimension relations are to be joined (e.g., Student(S), Course(C) below). A bit encoding is shown in reduced font italics for certain attributes, e.g., gen=gender, s=Student#, etc. S|s____|name_|gen| C|c____|name|st|term| E|s____|c____|grade | |0 000|CLAY |M 0| |0 000|BI |ND|F 0| |0 000|1 001|B 10| |1 001|THAIS|M 0| |1 001|DB |ND|S 1| |0 000|0 000|A 11| |2 010|GOOD |F 1| |2 010|DM |NJ|S 1| |3 011|1 001|A 11| |3 011|BAID |F 1| |3 011|DS |ND|F 0| |3 011|3 011|D 00| |4 100|PERRY|M 0| |4 100|SE |NJ|S 1| |1 001|3 011|D 00| |5 101|JOAN |F 1| |5 101|AI |ND|F 0| |1 001|0 000|B 10| |2 010|2 010|B 10| |2 010|3 011|A 11| |4 100|4 100|B 10| |5 101|5 101|B 10| Vertical bit sliced (uncompressed P-trees) attributes stored as: S.s 2 S.s 1 S.s 0 S.gC.c 2 C.c 1 C.c 0 C.tE.s 2 E.s 1 E.s 0 E.c 2 E.c 1 E.c 0 E.g 1 E.g 0 0000000000000110 0010001100000011 1000100100101100 1011101000100010 0101010101101111 0111011001101100 01001010 01000111 10010010 10110110 Vertical (un-bit-sliced) attributes are stored: S.name C.name C.st |CLAY | |BI | |ND| |THAIS| |DB | |ND| |GOOD | |DM | |NJ| |BAID | |DS | |ND| |PERRY| |SE | |NJ| |JOAN | |AI | |ND|

6. O.o 2 0 1 When 1 or more joins are required and there are more than 1 join attributes, e.g., the following SPJ on Student, Course, Offerings, Rooms, Enrollments files (next 5 slides): R:r cap |0 00|30 11| |1 01|20 10| |2 10|30 11| |3 11|10 01| SELECT S.n, C.n FROM S, C, O, R, E WHERE S.s=E.s & C.c=O.c & O.o=E.o & O.r=R.r & S.g=M & C.r=2 & E.g=A & R.c=20; S:s n gen |0 000|A|M| |1 001|T|M| |2 010|S|F| |3 011|B|F| |4 100|C|M| |5 101|J|F| C:c n cred |0 00|B|1 01| |1 01|D|3 11| |2 10|M|3 11| |3 11|S|2 10| E:s o grade |0 000|1 001|2 10| |0 000|0 000|3 11| |3 011|1 001|3 11| |3 011|3 011|0 00| |1 001|3 011|0 00| |1 001|0 000|2 10| |2 010|2 010|2 10| |2 010|7 111|3 11| |4 100|4 100|2 10| |5 101|5 101|2 10| O :o c r |0 000|0 00|0 01| |1 001|0 00|1 01| |2 010|1 01|0 00| |3 011|1 01|1 01| |4 100|2 10|0 00| |5 101|2 10|2 10| |6 110|2 10|3 11| |7 111|3 11|2 10| S.s 2 0 1 0 E.s 2 0 1 C.c 1 0 1 R.r 1 0 1 S.s 1 0 1 S.s 0 0 1 0 1 0 1 S.n A T S B C J S.g M F M F C.c 0 0 1 0 1 C.n B D M S C.r 1 0 1 C.r 0 1 0 R.r 0 0 1 0 1 R.c 1 1 0 R.c 0 1 0 1 O.o 1 0 1 0 1 O.o 0 0 1 0 1 0 1 0 1 O.c 1 0 1 O.c 0 0 1 0 1 O.r 1 0 1 O.r 0 1 0 1 0 1 0 E.s 1 0 1 0 1 0 E.s 0 0 1 0 1 E.o 2 0 1 E.o 1 0 1 0 1 0 E.o 0 1 0 1 0 1 0 1 E.g 1 1 0 1 E.g 0 0 1 0 1 0

7. For selections, S.g=M C.r=2 E.g=A R.c=20 create selection masks (note that C.r=2 is coded in binary as 10 b S.s 2 0 1 0 S.s 1 0 1 S.s 0 0 1 0 1 0 1 S.n A T S B C J S.g M F M F E.s 2 0 1 E.s 1 0 1 0 1 0 E.s 0 0 1 0 1 E.o 2 0 1 0 1 E.o 1 0 1 0 1 0 E.o 0 1 0 1 0 1 0 1 E.g 1 1 0 1 E.g 0 0 1 0 1 0 C.c 1 0 1 C.c 1 0 1 0 1 C.n B D M S C.r 1 0 1 C.r 2 1 0 O.o 2 0 1 O.o 1 0 1 0 1 O.o 0 0 1 0 1 0 1 0 1 O.c 1 0 1 O.c 0 0 1 0 1 O.r 1 0 1 O.r 0 1 0 1 0 1 0 R.r 1 0 1 R.r 0 0 1 0 1 R.c 1 1 0 R.c 0 1 0 1 SELECT S.n, C.n FROM S, C, O, R, E WHERE S.s=E.s & C.c=O.c & O.o=E.o & O.r=R.r & S.g=M & C.r=2 & E.g=A & R.c=20; SM 1 0 1 0 C.r 1 0 1 C.r’ 2 0 1 Cr2 0 1 E.g 1 1 0 1 E.g 0 0 1 0 1 0 EgA 0 1 0 1 0 R.c 1 1 0 R.c’ 0 0 1 0 Rc20 0 1 0 Apply selection masks (Zero out numeric values, blanked out others). S.s 2 0 S.s 1 0 1 0 S.s 0 0 1 0 S.n A T C E.s 2 0 E.s 1 0 1 0 1 0 E.s 0 0 1 0 E.o 2 0 1 0 E.o 1 0 1 0 E.o 0 0 1 0 1 0 C.c 1 0 1 C.c 0 0 1 C.n S O.o 2 0 1 O.o 1 0 1 0 1 001010101001010101 O.c 1 0 1 O.c 0 0 1 0 1 O.r 1 0 1 O.r 0 1 0 1 0 1 0 R.r 1 0 R.r 0 0 1 0

8. SELECT S.n, C.n FROM S, C, O, R, E WHERE S.s=E.s & C.c=O.c & O.o=E.o & O.r=R.r & S.g=M & C.r=2 & E.g=A & R.c=20; S.s 2 0 S.s 1 0 1 0 S.s 0 0 1 0 S.n A T C E.s 2 0 E.s 1 0 1 0 1 0 E.s 0 0 1 0 E.o 2 0 1 0 E.o 1 0 1 0 E.o 0 0 1 0 1 0 C.c 1 0 1 C.c 0 0 1 C.n S O.o 2 0 1 O.o 1 0 1 0 1 O.o 0 0 1 0 1 0 1 0 1 O.c 1 0 1 O.c 0 0 1 0 1 O.r 1 0 1 O.r 0 1 0 1 0 1 0 R.r 1 0 R.r 0 0 1 0 For the joins, S.s=E.s C.c=O.c O.o=E.o O.r=R.r, one approach is to follow an indexed nested loop like method (note that the P-trees themselves are self indexing). The join O.r=R.r is simply part of a selection on O (R doesn’t contribute output nor participate in any further operations) Use the Rc20-masked R as the inner relation and O as the r-indexed outer relation) to produce a further selection mask for O. Rc20 0 1 0 Get 1 st R.r value, 01 b Mask the corresponding O tuples, P O.r 1 ^P’ O.r 0 O.r 1 0 1 O’.r 0 0 1 0 1 0 1 OM 0 1 0 1 This is the only R.r value (if there were more, one would do the same for each, then OR those masks to get the final O-mask). Next, we apply the O-mask, OM to O O.o 2 0 1 0 1 O.o 1 0 1 O.o 0 0 1 0 1 O.c 1 0 1 0 1 O.c 0 0 1

9. SELECT S.n, C.n FROM S, C, O, R, E WHERE S.s=E.s & C.c=O.c & O.o=E.o & O.r=R.r & S.g=M & C.r=2 & E.g=A & R.c=20; S.s 2 0 S.s 1 0 1 0 S.s 0 0 1 0 S.n A T C E.s 2 0 E.s 1 0 1 0 1 0 E.s 0 0 1 0 E.o 2 0 1 0 E.o 1 0 1 0 E.o 0 0 1 0 1 0 C.c 1 0 1 C.c 0 0 1 C.n S For the final 3 joins C.c=O.c O.o=E.o E.s=S.s the same indexed nested loop like method can be used. O.o 2 0 1 0 1 O.o 1 0 1 O.o 0 0 1 0 1 O.c 1 0 1 0 1 O.c 0 0 1 Get 1 st masked C.c value, 11 b Mask corresponding O tuples: P O.c 1 ^P O.c 0 O.c 1 0 1 0 1 O.c 0 0 1 OM 0 1 Get 1 st masked O.o value, 111 b Mask corresponding E tuples: P E.o 2 ^P E.o 1 ^P E.o 0 E.o 1 0 1 0 E.o 0 0 1 0 1 0 Get 1 st masked E.s value, 010 b Mask corresponding S tuples: P’ S.s 2 ^P S.s 1 ^P’ S.s 0 S’.s 2 1 0 1 0 S.s 1 0 1 0 S’.s 0 1 0 1 0 SM 0 1 0 Get S.n-value(s), C, pair it with C.n-value(s), S, output concatenation, C.n S.n There was just one masked tuple at each stage in this example. In general, one would loop through the masked portion of the extant domain at each level (thus, Indexed Horizontal Nested Loop or IHNL) E.o 2 0 1 0 EM 0 1 0 S C

10. SELECT S.n, C.n FROM S, C, O, R, E WHERE S.s=E.s & C.c=O.c & O.o=E.o & O.r=R.r & S.g=M & C.r=2 & E.g=A & R.c=20; S.s 1 0 S.s 2 0 1 0 S.s 3 0 1 0 S.n A T C E.s 1 0 E.s 2 0 1 0 1 0 E.s 3 0 1 0 E.o 1 0 1 0 E.o 2 0 1 0 E.o 3 0 1 0 1 0 C.c 1 0 1 C.c 1 0 1 C.n S Having done the query tree sequentially (selections first, then joins and projections) it appears that the entire query tree could be done in one combined step by looping through the masked C tuples, for each C.n value, determine if there is an S.n value that should be paired with it by logical operations output those S.n, C.n pair(s), if any, else go to the next masked C.n value. Does this lead to a one-pass vertical query optimizer?!?!?! Can the indexed nested loop like algorithm be modified to loop horizontally? (across bit positions, rather than down tuples?) O.o 1 0 1 0 1 O.o 2 0 1 O.o 3 0 1 0 1 O.c 1 0 1 0 1 O.c 2 0 1

11. DISTINCT Keyword, GROUP BY Clause, ORDER BY Clause, HAVING Clause and Aggregate Operations Duplicate elimination after a projection (SQL DISTINCT keyword) is one of the most expensive operations in query optimisation. In general, it is as expensive as the join operation. However, in our approach, it can automatically be done while forming the output tuples (since that is done in an order). While forming all output records for a particular value of the ORDER BY attribute, duplicates can be easily eliminated without the need for an expensive algorithm. The ORDER BY and GROUP BY clauses are very commonly used in queries and can require a sorting of the output relation. However, in our approach, if the central relation is chosen to be the one with the sort attribute and the surrogation is according to the attribute order (typically the case – always the case for numeric attributes), then the final output records can be put together and aggregated in the requested order without a separate sort step at no additional cost. Aggregation operators such as COUNT, SUM, AVG, MAX, and MIN can be implemented without additional cost during the output formation step and any HAVING decision can be made as output records are being composed, as well (See Yue Cui’s Master’s thesis in NDSU library for vertical aggregation computations using P-trees.) If the Count aggregate is requested by itself, we note that P-trees automatically provide the full counts for any predicate with just one multiway AND operation.

12. The following example illustrates these points. SELECTDISTINCT C.c, R.capacity FROM S,C,E,O,R WHERE S.s=E.s AND C.c=O.c AND O.o=E.o AND O.r=R.r AND C.cred>1 AND (E.grade='B' OR E.grade='A') AND R.capacity>10 ORDER BY C.c; S___________ C___________ E_________________ O_______________ R_____________ |s |n|gen| |c |n|cred| |s |o |grade| |o |c |r | |r |capacity| |0 000|A|M 0| |0 00|B|1 01| |0 000|1 001|2 10| |0 000|0 00|0 01| |0 00|30 11| |1 001|T|M 0| |1 01|D|3 11| |0 000|0 000|3 11| |1 001|0 00|1 01| |1 01|20 10| |2 010|S|F 1| |2 10|M|3 11| |3 011|1 001|3 11| |2 010|1 01|0 00| |2 10|30 11| |3 011|B|F 1| |3 11|S|2 10| |3 011|3 011|0 00| |3 011|1 01|1 01| |3 11|10 01| |4 100|C|M 0| |1 001|3 011|0 00| |4 100|2 10|0 00| |5 101|J|F 1| |1 001|0 000|2 10| |5 101|2 10|2 10| Sn |2 010|2 010|2 10| |6 110|2 10|3 11| A |2 010|3 011|3 11| |7 111|3 11|2 10| T |4 100|4 100|2 10| S |5 101|5 101|2 10| Ss1 Ss2 Ss3 Sgen B 0011 0000 0101 0001 C Egrade1 Egrade2 Cn 00 11 01 11 J 1101 0100 Cc1 Cc2 Ccred1 Ccred2 B 1011 1001 00 01 01 11 D Es1 Es2 Es3 Eo1 Eo2 Eo3 11 00 11 01 11 10 M 0000 0000 0011 0000 0010 1010 S 0000 1111 1100 0000 0111 1101 Rr1 Rr2 Rcap1 Rcap2 11 00 01 11 00 01 00 01 11 10 11 01 10 11 Oo1 Oo2 Oo3 Oc1 Oc2 Or1 Or2 0011 0000 0101 0011 0000 0001 1100 0011 1111 0101 0011 1101 0011 0110 Apply selection masks: mE =Egrade1 mR =Rcap1 mC =Ccred1 1101 11 01 1011 10 11 11

13. results in, Es1 Es2 Es3 Eo1 Eo2 Eo3 Rr1 Rr2 Cc1 Cc2 00 0 00 0 00 1 00 0 00 0 10 0 00 01 0 1 0 00 1 11 1 00 0 00 0 11 1 01 1 0 11 01 11 00 01 11 00 01 Semijoin (toward center), E  O(on o=0,1,2,3,4,5), R  O(on r=0,1,2), C  O(on c=1,2,3), reduces Oo1 Oo2 Oo3 Oc1 Oc2 Or1 Or2 0011 0000 0101 0011 0000 0001 1100 0011 1111 0101 0011 1101 0011 0110 to Oo1 Oo2 Oo3 Oc1 Oc2 Or1 Or2 11 00 01 11 00 01 00 00 11 01 00 11 00 01 Thus, the participants are c=1,2; r=0,1,2; o=2,3,4,5. Semijoining back again produces the following. Cc1 Cc2 Rr1 Rr2 Es1 Es2 Es3 Eo1 Eo2 Eo3 0 1 00 01 00 11 00 00 11 01 1 0 1 0 11 00 01 11 00 01 Thus, s partic are s=2,4,5. Ss1 Ss2 Ss3 11 00 01 0 1 0 Output tuples are determined from participating O.c P-trees. RC(P O.c (2)) = RC(Oc 1 ^Oc 2 ’)=2, since Oc1 ^ Oc2’ 11 11 = 11 00 00 00 Since the 1-bits are in positions 4 and 5, the two O-tuples have O.o surrogate values 4 and 5. The r-values at positions 4 and 5 of O.r are 0 and 2. Thus, we retrieve the R.capacity values at offsets 0 and 2. However, both of these R.capacity values are 30. Thus, this duplication is discovered without sorting or additional processing. The only output is (2,30). Similarly, RCntP O.c (1) = RCntOc 1 ’^Oc 2 =2, Oc1’ ^ Oc2 00 00 = 00 11 11 11 Finally note, if ORDER BY clause is over an attribute which is not in the relation O (e.g., over student number, s) then we center the query tree (or wheel) on a fact file that contains the ORDER BY attribute (e.g., on E in this case). If the ORDER BY attribute is not in any fact file (in a dimension file only) then the final query tree can be re-arranged to center on the dimension file containing that attribute. Since output ordering and duplicate elimination are traditionally very expensive sub-operations of SPJ query processing, the fact that our BDM model and the P-tree data structure provide a fast and efficient way to accomplish these operations is a very favorable aspect of the approach.

14. Combining Data Mining and Query Processing Many data mining request involve pre-selection, pre-join, and pre-projection on a database to isolate the specific data subset to which the data mining algorithm is to be applied. For example, in the above database, one might be interested in all Association Rules of a given support threshold and confidence threshold but only on the result relations of the complex SPJ query shown. The brute force way to do this is to first join all relations into one universal relation and then to mine that gigantic relation. This is not a feasible solution in most cases due to the size of the resulting universal relation. Furthermore, often some selection on that universal relation is desirable prior to the mining step. Our approach accommodates combinations of querying and data mining without necessitation the creation of a massive universal relation as an intermediate step. Essentially, the full vertical partitioning and P-trees provide a selection and join path which can be combined with the data mining algorithm to produce the desired solution without extensive processing and massive space requirements. The collection of P-trees and BSQ files constitute a lossless, compressed version of the universal relation. Therefore the above techniques, when combined with the required data mining algorithm can produce the combination result very efficiently and directly.

15. O.o 2 0 1 R:r cap |0 00|30 11| |1 01|20 10| |2 10|20 10| |3 11|10 01| S:s n gen |0 000|A|M| |1 001|T|M| |2 010|S|F| |3 011|B|F| |4 100|C|M| |5 101|J|F| C:c n cred |0 00|B|1 01| |1 01|D|3 11| |2 10|M|3 11| |3 11|S|2 10| E:s o grade |0 000|1 001|2 10| |0 000|0 000|3 11| |3 011|1 001|3 11| |3 011|3 011|0 00| |1 001|3 011|0 00| |1 001|0 000|2 10| |2 010|2 010|2 10| |2 010|7 111|3 11| |4 100|4 100|2 10| |5 101|5 101|2 10| O :o c r |0 000|0 00|0 01| |1 001|0 00|1 01| |2 010|1 01|0 00| |3 011|1 01|1 01| |4 100|2 10|0 00| |5 101|2 10|2 10| |6 110|2 10|3 11| |7 111|3 11|2 10| S.s 2 0 1 0 E.s 2 0 1 C.c 1 0 1 R.r 1 0 1 S.s 1 0 1 S.s 0 0 1 0 1 0 1 S.n A T S B C J S.g 0 1 0 1 C.c 0 0 1 0 1 C.n B D M S C.r 1 0 1 C.r 0 1 0 R.r 0 0 1 0 1 R.c 1 1 0 R.c 0 1 0 1 O.o 1 0 1 0 1 O.o 0 0 1 0 1 0 1 0 1 O.c 1 0 1 O.c 0 0 1 0 1 O.r 1 0 1 O.r 0 1 0 1 0 1 0 E.s 1 0 1 0 1 0 E.s 0 0 1 0 1 E.o 2 0 1 E.o 1 0 1 0 1 0 E.o 0 1 0 1 0 1 0 1 E.g 1 1 0 1 E.g 0 0 1 0 1 0 Horizontal Indexed Nested Loop Join??? SELECT * FROM S,E WHERE S.s=E.s 1 st if 0<rc(S.s 2 ) thenif rc(S.s 2 )<|S| thenif 0<rc(S.s 1 )^ thenif rc(S.s 1 )<|S| thenif 0<rc(S.s 0 ) thenif rc(S.s 0 )<|S| then… So depth-first traversal down the bitslice tree for S.s, skipping all values that are not present, and for each S.s value that is present, one and gives that value P-tree in E (index into E) so optimal retrieval can be done. If the Ptrees are organized according to physical boundaries as below, then is there a P-tree based Hybrid Hash join that allows us to avoid excessive rereads of extents? It seems clear that compressing bit vectors into P-trees based, not on 1/2 d boundaries, but on page and extent boundaries is important. Use the Dr. Md Masum Serazi approach, but with the following levels (possibly collapsing levels 0 and 1 together) The level-0 fanout is the bfr of the page blocks. The level-1 fanout is the extent size (# of blocks per extent). The level-2 fanout is the (maximum) number of extents per file. The level-3 fanout is the number of files in the DB The real advantage of this approach may to apply it to join algorithms where the location of join Attribute values is known ( see V. Goli’s thesis) since we know the location of all values Through ANDs.

16. E.s 2 1 0 1 0 1 0 1 E.s 1 0 1 0 1 0 1 0 E.s 0 0 1 0 1 0 1 0 S.s 2 0 1 S.s 1 0 1 0 1 S.s 0 0 1 0 1 0 1 0 1 S.a 1 0 1 0 C.c 1 0 1 C.c 0 0 1 0 1 C.n 1 0 1 E.g 0 1 0 1 0 1 0 E.c 1 0 1 0 1 0 1 0 1 0 E.c 0 1 0 1 0 1 0 1 0 1 0 1 0 Graph G=(N,E) is (T,I)-bipartite iff N=T  !I and  e={e 1,e 2 }  E, if e 1  T [I] then e 2  I [T]. WOLOG write e={e T,e I } (E is directed from T to I e=(e T,e I ) ) E={ {e k,T,e k,I } | k=1..|E|} or the edge relationship can be expressed as tIset, E T = { (t,Iset(t) | t  T and Iset(t)={i|{t,i}  E} iTset, E I = { (i,Iset(i) | where Iset(i)={t | {t,i}  E} tImap, E Tb ={ (t,b 1,...,b |I| ) | where b k =1 iff e k,T =t} iTmap, E Ib ={ (i,b 1,...,b |T| ) | where b k =1 iff e k,I =t} Given a star schema with fact, E and dimensions, S, C. E is a ER-relationship between entities, S and C and is therefore a bipartite graph, G=(N,E) where N is the disjoint union of S and C. Given a join S.s with E.s, JoinIndex (JI) is a relationship between S and E, giving a bipartite graph, G=(S  !E,JI). The sEmap of this relationship is the association matrix of Qiang Ding's thesis.

17. Desirable Features of a Distributed DBMS: LOCATION TRANSPARENCY is achieved if a user can access needed data without having to know which site has that data. -simplifies logic of programs -allows data movement as usage patterns change A data object (typically a file) is fragmented if it is divided into multiple pieces for storage and/or placement purposes at different sites. e.g., accounts files: Fargo customer accounts can be stored in Fargo, Grand Forks customer accounts can be stored in Grand Forks...) FRAGMENTATION TRANSPARENCY is achieved if users can access needed data without having to know whether it is fragmented. a data object (typically a record or file) is REPLICATED if it has ≥ 1 physical copy -distributed replication advantages include availability -disadvantages include increased update overhead. REPLICATION TRANSPARENCY is achieved if users can access needed data without knowing whether or not it is replicated. Additional desirable DDBMS features include: LOCAL AUTONOMY is achieved if the system is distributed consistent with the logical and physical distribution of the enterprise. It allows local control over local data, It allows local accountability and less dependency on remote Data Processing Support for INCREMENTAL GROWTH AVAILABILITY and RELIABILITY. QUERY PROCESSING in Distributed DBMSs (DDBMSs)

18. Distributed systems can more easily allow for graceful (and unlimited) growth simply by adding additional sites. The DDBMS software should allow for adding sites easily. Reliability can be provided by replicating data. The DDBMS should allow for replication to enhance reliability and availability in the presence of failures of sites or links. DISTRIBUTED QUERIES Query Optimization Methods can be STATIC: strategy of transmissions and local processing activities is fully determined before execution begins (at compile time). DYNAMIC: Each step is decided after seeing results of previous steps. Response time usually is dominated by transmission costs (i.e., local processing times are negligible by comparison - assumed 0?). One model is to take RESPONSE time to be linear in number of bytes, X, sent: R(X) = AX + B B is the fixed (setup?) cost of the transmission and AX is the variable cost (depending on message size only, not distance). What assumptions does this make? (next slide) QUERY PROCESSING in Distributed DBMSs (DDBMSs)

19. Bandwidth = The number of fbits per second that can be sent. delay Time to send a message from point A to point B Propagation = Distance / SpeedOfLight: The time between when the last bit enters and last bit leaves the link. Transmit = Size / Bandwidth: The time between when 1 st bit enters and last bit enters the link. Components of delay = Propagation + Transmit + Queue (=delays in send and demultiplexing queues) Propagation versus Transmit delay If you’re sending 1 byte, propagation delay dominates. If you’re sending 500 MB, transmit delay dominates

20. A STATIC, QUERY PROCESSING ALGORITHM usually takes as input: database statistics such as relation sizes attribute sizes projected sizes of attributes produces as output: a strategy for answering the query (a pattern of what transmissions to make, when, where and what local processing to do, when and where) Usually involves 4 phases: LOCAL PROCESSING phase: do all processing that can be done initially at each site that doesn't require data interchange between sites. (e.g., local selections, joins and projections) The result this phase is that there will be one participating relation at each participating site. REDUCTION phase: selected "semijoins" to be done to reduce the size of participating relations by eliminating tuples that are not needed in answering the query. TRANSPORT phase: send one relation from each participating site (the result of the reduction phase) to the querying site. COMPLETION phase: finishing up processing using those relations to get final answer (e.g., final projects, selects, joins) QUERY PROCESSING in Distributed DBMSs (DDBMSs) 2

21. What is the SEMIJOIN of R 1 (A,B) to R 2 (A,C) on A? (written: R 1 :A→ R 2 ). 1. projection R onto A (again, the result is written as R 1 [A]) 2. R 1 [A]:A→ R 2 (Which selects those tuples of R 2 that will participate in the join). The result of R 1 [A]:A→ R 2 is the sub-relation of R 2 of only those R 2 -tuples which will participate in the full join of R 1 JOIN A R 2 on A (eliminates non-participants at the cost of generating (and sending, if R 2 is located at a different site than R 1 ) the R 1 -join attribute values). A semijoin can be viewed as a special SELECTION operator also, since it selects out those tuples of R 2 that have a matching A-value in R 1. Thus the semijoin is perfect for reducing the size of relations before they are sent to the querying site. But note that semijoins don't always end up reducing the size of a relation. QUERY PROCESSING in Distributed DBMSs (DDBMSs)

22. For example, STUDENT-FILE S#|SNAME |LCODE 25|CLAY |NJ5101 32|THAISZ|NJ5102 38|GOOD |FL6321 17|BAID |NY2091 57|BROWN |NY2092 ENROLL FILE S#|C#|GRADE 32|8 | 89 32|7 | 91 32|6 | 62 38|6 | 98 STATIC QUERY PROCESSING Example in DDBMSs ENROLL S# →STUDENT 1. project ENROLL onto the S# attribute: S# 32 38 2. Join the two relations on S# S# S# |SNAME |LCODE |32| 25|CLAY |NJ5101 |38| join 32|THAISZ|NJ5102 38|GOOD |FL6321 17|BAID |NY2091 57|BROWN |NY2092 resulting in: S# |SNAME |LCODE 32|THAISZ|NJ5102 38|GOOD |FL6321

23. STATIC QUERY PROCESSING Example in DDBMSs semijoins don't always end up reducing the size of a relation. Consider STUDENT S#→ ENROLL Project STUDENT onto the S# attribute and join it with ENROLL: S# S#|C#|GRADE 25 join 32|8 | 89 32 32|7 | 91 38 32|6 | 62 17 38|6 | 98 57 resulting in the entirety of STUDENT again (no tuples eliminated)! So let's make it a rule: - never semijoin the primary key to a foreign key, because it will always result in no reduction.

24. Distributed Semijoin of R 1 at site1 to R 2 at site2 along A At site1: R 1 At site 2: R 2 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 1 A 2 a A A B C C E A F d 1 a C D D E A A B B e 2 b A B C D B A B A g 3 c D D B B A C A C e E B A A C C D D Assume response time for transmission of X bytes between any 2 sites is R(X) = X + 10 time units. 1. projection R 1 [A] 2. transmission of R 1 [A] to the site2. 3. R 1 [A] A-join R 2 (select R 2 -tuples that participate in join) Consider the following distributed query: Assume SELECT R 1.A 2, R 2.A 2 FROM R 1,R 2 WHERE R 1.A 1 = R 2.A 1 arrives at site3.

25. a A A B C C E A F a C D D E A A B B b A B C D B A B A c D D B B A C A C e E B A A C C D D Distributed Semijoin of R 1 at site1 to R 2 at site2 along A STRATEGY 1 Strategy-1: (No reduction phase). 1. Send R 1 to site3: 45 bytes sent. Cost is R(45)=45+10 = 55 2. Send R 2 to site3: 6 bytes sent. Cost of R(6)= 6+10 = 16 3. Final join (cost = 0) site2: R 2 A 1 A 2 a A A B C C E A F a C D D E A A B B b A B C D B A B A c D D B B A C A C e E B A A C C D D At site1: R 1 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 site3 d 1 e 2 g 3 result: e EBAACCDD2 Response time= 71 Strategy 1': If 1. and 2. are done in parallel, the response time= 55

26. degdeg e E B A A C C D D a A A B C C E A F a C D D E A A B B b A B C D B A B A c D D B B A C A C e E B A A C C D D Distributed Semijoin of R 1 at site1 to R 2 at site2 along A; STRATEGY 2 1. Send R 2 [A] to site1; site2: R 2 A 1 A 2 At site1: R 1 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 site3 result: e EBAACCDD2 Response time = 48 Strategy 2': If 1. and 3. are (can be?) done in parallel, the response time = 32 2. Send R 2 [A]  R 1 to site3. 9 bytes sent. Cost=R(9)= 19 3. Send R 2 to site3; 6 bytes sent. Cost=R(3) = 16 4. JOIN R 2 [A]  R 1 and R 2 on A 1 at site3. Cost = 0 do R 2 [A]  R 1 d 1 e 2 g 3 3 bytes sent. Cost=R(3)= 13 d 1 e 2 g 3

27. degdeg e E B A A C C D D a A A B C C E A F a C D D E A A B B b A B C D B A B A c D D B B A C A C e E B A A C C D D Distributed Semijoin of R 1 at site1 to R 2 at site2 along A; STRATEGY 3 1. Send R 1 [A] to site2; site2: R 2 A 1 A 2 At site1: R 1 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 site3 result: e EBAACCDD2 Response time = 122 Strategy 3': If 1. and 3. are (can be?) done in parallel, the response time = 67 2. Send R 1 [A]  R 2 to site3. 2 bytes sent. Cost=R(2)= 12 3. Send R 1 to site3; 45 bytes sent. Cost=R(45) = 55 4. JOIN R 1 [A]  R 2 and R 1 on A 1 at site3. Cost = 0 do R 1 [A]  R 2 d 1 e 2 g 3 45 bytes sent. Cost=R(45)= 55 d 1 e 2 g 3

28. Distributed Query Processor DQP) must pick the strategy! For static algorithms, the hardest job of the Distributed Query Processor (which is at site3 where the query came in and must be processed) is to pick among these 6 alternatives (if other transmission and local processing cost are used, there would be a vastly different set of alternative strategies). The DQP at site3 must pick a strategy without seeing the data at aites 1 and 2. E.g., if the DQP decides that a "one semijoin strategy is best, should it be 2, versus 3 (or 2' versus 3' if the network accomodates parallel transmissions from a given send site). Note the vast difference is cost (2 costs 48 and 3 costs 122, though both are 1 semijoin strategies! 3 costs more than the no semijoin strategy which is 1 at a cost of 55). The DQP has a need for estimates of the two semijoin result sizes, since the actual results are not known in advance at site 3. That estimation method is important, but difficult, since the situation can be very different than the above.

29. d 1 e 2 g 3 q 4 q 5 v 7 d 1 e 2 g 3 q 4 q 5 v 7 d A A B C C E A F d C D D E A A B B e A B C D B A B A g D D B B A C A C e E B A A C C D D degqqvdegqqv d A A B C C E A F d C D D E A A B B e A B C D B A B A g D D B B A C A C e E B A A C C D D STRATEGY 2 with different R 1 and R 2 data 1. Send R 2 [A] to site1; site2: R 2 A 1 A 2 At site1: R 1 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 site3 result: dEBAACCDD1 Response time = 93 dCDDEAABB1 eABCDBABA2 gDDBBACAC3 eEBAACCDD2 Strategy 1 has same cost=71 so Strategy 1 is better! How should semijoin results be estimated? 2. Send R 2 [A]  R 1 to site3. 45 bytes sent. Cost=R(45)= 55 3. Send R 2 to site3; 12 bytes sent. Cost=R(12) = 22 4. JOIN R 2 [A]  R 1 and R 2 on A 1 at site3. Cost = 0 do R 2 [A]  R 1 6 bytes sent. Cost=R(3)= 16

30. Selectivity Theory for estimating semijoin results. The work of Hevner and Yao assumes data values are uniformly distributed and attribute-distributions are independent of each other. Results estimated as follows: (assuming A 1 has domain {a,b,...z}). The Selectivity of attributed R 1 is the ratio of the number of values present (size of the extant domain) over the number of values possible (size of full domain). Therefore the selectivity of R 1.A is 3/26. Using selectivity theory, we estimate the size of semijoin, R 1 A-semijoin R2 as: (Original size of R 2 )*(selectivity of incoming, R 1.A): 45 * 3/26 = 5.2 Selectivity theory estimates 5.2 bytes of R 1 survive semijoin. This is close for the first example database state and the algorithm proposed by Hevner & Yao (ALGORITHM-GENERAL) would correctly select method 1. However, it is way off in the second database state but ALGORITHM-GENERAL would still select strategy-1 (not best for this DB state).

31. UPDATE PROPAGATION IN DISTRIBUTED DATABASES UPDATE PROPAGATION: To update any replicated data item, the DDBMS must propagate the new value consistently to all copies. IMMEDIATE method: update all copies (the update fails if even 1 copy is unavailable) PRIMARY method: designate 1 copy as primary for each item. Update is deemed complete (COMMITTED) when primary copy is updated. Primary copy site is responsible for broadcasting the update to the other sites. Broadcast can be done in parallel while the transaction is contining, however that runs counter to local autonomy theme