Star/join query optimization5960428Abstract Unwieldy star/join queries are performed more efficiently using a filtered fact table. Suitable queries include star/join queries with a large fact table joined with multiple subsidiary dimension tables, where indices exist over fact table join columns. The query is analyzed to prepare a query plan for the dimension table accesses. This plan is supplemented by adding nested loop join operations, where the inner table is a dimension table plan and the outer table is an index scan performed over a fact table index of the join column with the dimension table. The plan is also supplemented by filtering records resulting from the nested loop joins using a sequence of dynamic bit vectors, ultimately yielding a list of probable fact table records. The plan is further supplemented by fetching these records to construct a distilled fact which is used, instead of the large original table, to execute the query in considerably less time. If desired, the supplemented query plan and other competing approaches may studied to provide cost estimates, with the least costly approach being actually implemented. Claims What is claimed is: Description BACKGROUND OF THE INVENTION
TABLE 1
______________________________________
select D1.C1, D2.C2, D3.C3, sum (F.M1), max (F.M2)
from D1, D2, D3, F
where D1.K1 = F.DIM1 and
D1.A1 = "New Jersey" and
D2.K2 = F.DIM2 and
D2.A2 = "45" and
D3.K3 = F.DIM3 and
D3.A3 = "1995"
______________________________________
The typically huge size of the fact table is one reason that star/join queries are so difficult to execute efficiently. With this fact table, even a restrictive join predicate between fact and dimension tables result in unwieldy answer sets. A typical fact table may easily include one billion rows. Thus, with this fact table, even a restrictive join predicate that yields one-thousandth of the fact table's rows would still yield an answer set with one million rows. Engineers and scientists have developed a number of different approaches to address the difficulties of star/join queries. One approach is the "Cartesian product" technique, which is employed by the DB2 database product of International Business Machines Corp. This technique applies the query to the dimension tables by taking their Cartesian product, and then applies this result to the fact table using composite indexes matching the join columns of the fact table. Query execution, though, can still be lengthy if the dimension table predicates are not sufficiently restrictive due to the large size of the Cartesian product of the dimension tables. Another known approach to optimize star/join queries is the "bit map index" technique. This technique creates one bit map index for each fact table column that is joined with a dimension table column. Bit map index entries corresponding to qualifying dimension table join column values are OR-ed together within each dimension, and the results from each dimension are then AND-ed. Rows from the fact table identified by the final bit vector can then be accessed. This approach is described in O'Neil et al., Multi-Table Joins Through Bitmapped Join Indices, SIGMOD Record, Vol. 24, No. 3, September 1995. Although query optimizing techniques such as these constitute a significant advance and enjoy widespread use today, International Business Machines Corp. has continually sought to improve the performance and efficiency of query optimizing techniques. Chiefly, improvement is still sought in the area of query optimization and execution speed. SUMMARY OF THE INVENTION Broadly, the present invention concerns the efficient evaluation and execution of star/join type queries. A query is received and then analyzed to determined whether it is suitable for optimization according to the invention. Suitable queries include star/join queries with a large fact table joined with multiple subsidiary dimension tables. Also, indices over fact table join columns are required. If deemed suitable, the query is further analyzed to prepare a query plan for the dimension table accesses. The query plan is supplemented by adding nested loop join operations, where the outer table is a dimension table access plan and the inner table is an index scan performed over a fact table index of the join column(s) with the dimension table. Record references or record IDs resulting from the nested loop joins are filtered using a sequence of dynamic bit vectors, ultimately yielding a list of probable fact table record references. The probable fact table records can then be accessed to form a distilled fact table. The distilled fact table is then used in place of the original fact table to construct a supplemental execution plan for the entire star/join query. The estimated execution cost of the supplemental execution plan is compared to estimated costs of alternative query execution plans. Alternatively, the supplemental execution plan may be executed straight away, if a companion to other plans is unavailable or undesired. In one embodiment, the invention may be implemented to provide a method to plan and/or execute a star/join type query. In another embodiment, the invention may be implemented to provide an apparatus, such as a digital data processing machine, suitable to plan and/or execute a star/join type query. In still another embodiment, the invention may be implemented to provide a signal-bearing medium tangibly embodying a program of machine-readable instructions executable by a digital data processing apparatus to perform method steps to plan and/or execute a star/join type query. The invention affords its users with a number of distinct advantages. Chiefly, the invention performs star/join queries more rapidly than other approaches,. The invention also provides a number of other advantages and benefits, which should be apparent from the following description of the invention. BRIEF DESCRIPTION OF THE DRAWINGS The nature, objects, and advantages of the invention will became more apparent to those skilled in the art after considering the following detailed description in connection with the accompanying drawings, in which like reference numerals designate like parts throughout, wherein: FIG. 1 is a graphical representation of a typical star/join query. FIG. 2 is a block diagram of a digital data processing machine in accordance with the invention. FIG. 3 is a perspective view of an exemplary signal-bearing medium in accordance with the invention. FIG. 4 is a flowchart of an operational sequence for star/join query optimization in accordance with the invention. FIG. 5 is a block diagram of a query optimization plan according to the invention. FIG. 6 is a flowchart illustrating RID filtering according to the invention. FIG. 7 is a block diagram of an illustrative supplemental query plan according to the invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS HARDWARE COMPONENTS & INTERCONNECTIONS One aspect of the invention concerns a digital data processing apparatus, which may be embodied by various hardware components and interconnections. FIG. 2 shows an example of one digital data processing apparatus 200. The apparatus 200 includes a processing unit 202, such as a microprocessor or other processing machine, coupled to a storage unit 206. In the present example, the storage unit 206 includes a fast-access memory 208 and nonvolatile storage 210. The fast-access memory 208 preferably comprises high speed memory circuitry such as random access memory, and may be used to store the programming instructions executed by the processing unit 202. The nonvolatile storage 210 may comprise, for example, DASD (e.g., one or more magnetic data storage disks such as a "hard drive", floppy disk, or RAID array), magnetic tape, electronic read-only memory (e.g., ROM, EPROM, or EEPROM), optical storage (e.g., optical tape, WORM, CD-ROM, DVD), paper "punch" cards, or another suitable signal-bearing medium including transmission media such as digital analog, communication links, or wireless media. The apparatus 200 also includes an input/output 204, such as a line, bus, cable, electromagnetic link, fiber optic link, or other means to exchange data with the processing unit 202. Despite the specific foregoing description, ordinarily skilled artisans (having the benefit of this disclosure) will recognize that the apparatus 200 may be still implemented in a machine of different construction, without departing from the scope of the invention. As a specific example, one of the components 208/210 may be eliminated; furthermore, the storage unit 206 may be provided on-board the processing unit 202, or even provided externally to the apparatus 200. As a specific example, the apparatus 200 may be embodied by a high-power computer such as an IBM RS/6000 machine. OPERATION In addition to the various hardware embodiments described above, a different aspect of the invention concerns a method for optimizing a database query using a digital data processing machine such as the apparatus 200. Signal-Bearing Media Such a method may be implemented, for example, by operating the processing unit 202 to execute a sequence of machine-readable instructions. These instructions may reside in various types of signal-bearing media. In this respect, one aspect of the present invention concerns an article of manufacture or "programmed product", comprising signal-bearing media tangibly embodying a program of machine-readable instructions executable by a digital data processing machine to perform a method to plan and/or execute database queries. This signal-bearing media may comprise, for example, RAI4 or other memory embodied by the fast access memory 208. Alternatively, the instructions may be contained in another signal-bearing media, such as a magnetic data storage diskette 300 (FIG. 3), directly or indirectly accessible by the processing unit 202. Whether contained in the apparatus 200, a separate signal-bearing media, or elsewhere, the instructions may be stored on a variety of machine-readable data storage media, such as DALSD storage (e.g., a conventional "hard drive", floppy disk, or a RAID array), magnetic tape, electronic read-only memory (e.g., ROM, EPROM, or EEPROM), an optical storage device (e.g. WORM, CD-ROM, DVD, optical tape), paper "punch" cards, or other suitable signal-bearing media including transmission media such as digital and analog and communication links and wireless. In an illustrative embodiment of the invention, the machine-readable instructions may comprise lines of compiled PLX and C++ language code. Overall Sequence of Operation FIG. 4 shows a sequence of method steps 400 to illustrate one example of the method aspect of the present invention. For ease of explanation, but without any limitation intended thereby, the example of FIG. 4 is described in the context of the digital data processing apparatus 200 described above. Generally, the routine 40(0 constructs a plan for executing a star/join query, compares the plan to certain criteria, and executes the plan if it satisfies the criteria. Query Analysis The routine 400 commences in step 402, which receives a query to be optimized. In the present example, the received query is the query 100 (FIG. 1). Step 404 analyzes the query by applying certain predetermined criteria to the query. This is performed to determine whether this query it is a suitable candidate for optimization according to this process. In one embodiment, suitable queries must have the following characteristics: 1. The query involves an extremely large fact table, such as one million or one billion rows; although smaller fact tables may be used, the invention is most efficient with fact tables of possibly unwieldy size. 2. The query involves multiple dimension table joins. 3. An index exists on each column of the fact table being joined with a dimension table. 4. The cardinality of the dimension table is reduced after applying local predicates to the dimension table. Optionally, in determining a suitable query step 404 may ignore parts of a query lacking the desired characteristics. For example, step 404 may ignore any dimension table joins that do not reduce the fact table's size. However, it is contemplated that "snowflake" dimensions, which are query table constructs where the dimension table is joined with other tables separate from the fact table, should be allowable. Furthermore, if the required indices are not provided, the time required to generate them may be estimated and this time included in the time of the overall query plan being constructed in steps 408, 410, 416, 417, and 412. In the present example, step 404 finds that the query 100 is suitable for optimization according to the invention. In this example, the fact table 102 includes a plurality of rows, each row being uniquely identified by a numeric, alphabetic, or alphanumeric code called a row ID (RID). Query Plan: Dimension Table Accesses After step 404 finds that the query is desirable for optimization according to this technique, step 408 prepares query plans for performing each "dimension table access". "Dimension table accesses" include any auxiliary joins between the dimension table and tables other than the fact table (i.e., "snowflakes"). As an example, the dimension table query plans may be implemented as taught by Selinger et al., "Access path Selection in a Relational Database Management System", Proceedings of the ACM SIGMOD Conference, June 1979. The Selinger reference is hereby incorporated by reference in its entirety. Optionally, step 408 may limit its query plan to produce distinct join column values, eliminating any duplicates. In addition, step 408 also preferably suppresses columns of the dimension table other than the join column(s) to the fact table, using an operation such as "select". Supplementing Query Plan: "RID" Boiling After planning the dimension table accesses in step 408, step 410 supplements these plans to identify RIDs of a "distilled" fact table. This process is called staged comparative RID filtering. As explained in greater detail below, the distilled fact table is considerably smaller than the original fact table since it omits records that are irrelevant to the search. This step of supplementing the query plan to eliminate irrelevant RIDs is therefore known as "RID boiling". Accordingly, the original query can be performed substantially faster using the distilled fact table rather than the unwieldy original table. Nested Loop Join FIGS. 5-6 help illustrate the performance of step 410 in greater detail. FIG. 5 diagrams the generation of a supplemental query plan 500 for the query 100 of FIG. 1. Preferably, the plan 500 only includes the dimension table accesses that join to less than all rows of the fact table. Other dimension table operations may be ignored at this stage. In the present example, the query of each dimension table 104-107 reduces the cardinality of the fact table; thus, the plan 500 includes query plans 502-505 corresponding to all four dimension tables. For each dimension table, a corresponding "index scan" operation is performed. The index scans are shown by I-SCANs 508-511. More particularly, an index over the particular fact table column queried against the dimension table is scanned for the value(s) of the join column(s) of the dimension table. Using the exemplary query of FIG. 1 and Table 1, the index scan 508 searches an index over the "DIM1" column in the fact table for entries containing the value(s) of "K1" from dimension table D1. An example of this index is shown in Table 2, below.
TABLE 2
______________________________________
Exemplary Index
Fact Table RID DIM1 value
______________________________________
0001 Arizona
0002 California
0003 Texas
0004 Texas
. .
. .
. .
9761 New Jersey
______________________________________
The result of this index scan is a listing of RIDs identifying fact table rows satisfying the index scan. For the example of Table 1 and FIG. 1, the index scan results list all RIDs corresponding to fact table rows whose DIM1 column contains "New Jersey". Each index scan operation may also be referred to a "nested loop join". The nested loop joins are shown by 514-517. Each nested loop join effectively "joins" an "outer table" (the dimension table query) with an "inner table" (the fact table index). The results of each nested loop join is a listing of fact table RIDs, which may itself be viewed as a table. Staged Comparative RID Filtering The RIDs resulting from the nested loop joins 514-517 are fed to a series of "dynamic bit vectors", which complete the RID boiling process. Generally, the RIDs from a first nested loop join operation are used to "build" a first set of one or more RID bit vectors. Then, the RIDs from each subsequent nested loop join operation are used to generate bit vectors that are filtered by the previous bit vector(s). This process continues until reaching the final nested loop join 517, whose resultant RIDs are evaluated to determine whether they are represented in the previous bit vector(s). This process is shown more thoroughly in FIGS. 6-7. FIG. 6 shows a sequence 600 for RID filtering according to the invention. After the routine 600 begins in step 602, step 604 builds one or more bit vectors from the RIDs of one of the nested loop joins 514-517. As discussed below, step 604 preferably starts with the nested loop join 514-517 having the least number of resultant RIDs. In the illustrated example, this is the nested loop join 514. As shown in FIGS. 5-7, step 604 utilizes each resultant RID of the nested loop join 514 to select a bit in each of one or more bit vectors. Each bit vector comprises a single column of multiple binary values, all initially set to zero. The length of the bit vectors is set in proportion to the number of expected RIDs from the corresponding nested loop join operation. In the example of FIG. 7, three bit vectors 704-706 are shown, each bit vector corresponding exclusively to a particular hash function, each of the hash functions necessarily being independent of each other. All bit vectors 704-706 have the same length, which is selected in advance to accommodate the expected number of necessary RID values. The exemplary RID 702 is hashed with first, second, and third hash functions 708-710, which select bits from the first, second, and third bit vectors 704-706, respectively. More particularly, each hashing of the RID 702 identifies a bit in the corresponding bit vector that should be "set" to binary one. For instance, applying the hash function 708 to the RID 702 determines that a bit 714 of the bit vector 705 should be set to binary one. Similarly, applying the hash function 709 to the RID 702 determines that a bit 712 of the bit vector 704 should be set. Likewise, applying the hash function 710 to the RID 702 determines that a bit 716 of the bit vector 704 should be set. The bit vectors resulting from the first nested loop join 514 are represented by 520. If hashing of a RID from a subsequent nested loop join finds a binary "one" already set in each bit vector, the fact table row corresponding to the RID might be relevant to the overall query. RIDs that hash to a zero bit in one or more bit vectors 704-706, however, definitely represent rows of the fact table not pertinent to the overall query. RID boiling therefore represents a process of exclusion. Preferably, each hash function is a member of a set of "universal" hash functions, meaning that their output values are evenly distributed and also completely independent of the output values of the other hash functions. Thus, the results of each hash function 708-710 is preferably independent of the others. One example of a suitable hash function is discussed in greater detail in J. L. Carter et al., "Universal Classes of Hash Functions", Journal of Computer and System Sciences, Vol. 18, No. 2 (April 1979), pp. 143-154, which is hereby incorporated by reference in its entirety. As an example, hashing of a RID may set a bit whose position is given according to Equation 1, below. bit position =hash(RID) % (bit vector size) ›1! where: hash(RID): represents a numerical term resulting from application of the hash function; bit vector size: is the number of bits in the bit vector; and %: represents the mathematical modulo operation, i.e., x % y=remainder (x/y). Referring to FIGS. 5-7, after the first set of bit vectors 520 is built in step 604, step 606 proceeds to the next dimension table plan. This is the plan 503 in the present example. Preferably, step 604 starts with the dimension table nested loop join having the smallest number of matching fact table rows, and step 606 sequentially advances to dimension tables with larger and larger join cardinalities. Thus, the example of FIG. 5 is preferred when the nested loop join 514 produces the fewest RIDs, with the operations 515, 516, and 517 producing successively greater numbers of RIDs. Step 608 then asks whether this is the final dimension table to be processed. In the present example, where the dimension table plans are processed from left to right, the final dimension query plan to be processed is the plan 505. As alternative, step 608 may consider the number of bits set in the just-completed bit vectors and declare the next dimension table to be the "final" table if the number of bits is below a predetermined threshold. If step 608 finds that the current plan is not the final, step 610 performs a "probe and build" operation. Specifically, step 608 builds a set of new bit vectors (not shown) by processing resultant RIDs of the current dimension table according to FIG. 7. Each RID is hashed using the same hash functions 708-710. The bit positions corresponding to the hash function output in the old bit vectors are then checked. This constitutes a "probe" operation. Then, if all the checked bit positions are set to one for the current RID, the corresponding bit positions in the new bit vectors are set to one; this constitutes a "build" operation. Later, when step 608 arrives at the last dimension table, step 612 performs a probe operation for each RID from this last dimension table join. The RIDs that pass the probe operation represent rows of a fact table that has been distilled by removing rows irrelevant to the present query. In the illustrated example, the last dimension table plan is 505, which is associated with the nested loop join 517. The probe operation of step 612 separately evaluates each RID resulting from the nested loop join 517. Each RID is hashed by the same three hash functions used to process the other dimension tables. If the bits indicated by the hash functions are all set in the corresponding bit vectors from the previous dimension table plan, the fact table row corresponding to the RID still might be relevant to the overall query. Using the illustrated example, each RID from the nested loop join 517 is hashed three times to identify bits from the bit vector 522; if these bits are already set, this RID is potentially relevant to the overall query. RIDs that yield a zero in one or more bit vectors, however, definitely represent rows of the fact table not pertinent to the overall query. After probing all RIDs of the nested loop join 517, the surviving RIDs are output in step 614. Alternatively, if desired, RIDs may be probed and output individually. After step 614, the routine 600 ends in step 616. A similar problem to RID filtering, "set intersection filtering", is discussed in the following reference, the entirety of which is hereby incorporated by reference: B. H. Bloom, "Space/Time Trade-Offs in Hash Coding with Allowable Errors", Comm. ACM 13, 7 (July 1970), pp. 422-426. Fetching Fact Table Rows After step 410 prepares the supplemental query plan, step 416 further supplements the query plan by planning a fetch of the fact table rows identified by the "boiled" RIDs of step 410. Preferably, these rows are assembled in memory, such as the memory 208, to actually construct a distilled fact table. Alternatively, the identified row,; may be fetched into a non-table structure, or even identified in-place without fetching. As described below, the supplemented query plan is subsequently executed upon the identified rows, whether located in the distilled fact table, another data structure, or specified locations of the original fact table. As an optional enhancement, step 416 may also sort the fact table RIDs before fetching fact table rows to improve execution of the query. For instance. the rows may be sorted numerically according to their RIDs. This sorting would therefore group nearby rows, thereby streamlining the subsequent fetching of fact table rows. Completing Query Plan After step 416, step 417 completes the original star/join query plan by joining the distilled fact table with all the dimension tables. If desired, known query optimization techniques can be used to determine the best query execution plan for the star/join query given the distilled fact table. Such techniques include, for example, those of Selinger et al., referenced above. Evaluating Supplemented Query Plan After step 417 completes the supplemented query plan, step 412 evaluates the execution of this plan upon the distilled fact table. The supplemented plan is evaluated against certain predetermined criteria, which are selected depending upon the particular implementation needs. Preferably, the execution time of the supplemented plan is evaluated by comparing it to the execution time of one or more alternative plans. As an alternative, the supplemented plan may be evaluated by comparing it to a desired execution time. Regardless of the evaluation used, if the supplemented query plan is not satisfactory according to the criteria of step 412, then step 414 advances to step 406, which executes the query using a different approach. After step 406, the routine ends in step 420. Executing Planned Query Using Distilled Fact Table If step 414 finds the supplemental query plan satisfactory, step 418 executes the planned supplemented query upon the table rows collected or otherwise identified in step 416. This step includes any dimension table accesses that were ignored in creating the supplemental plan 500, e.g., those dimension table accesses failing to reduce the cardinality of the fact table. This query executes with considerable speed because it is performed on the distilled table rather than the substantially larger original table. After executing the query in step 418, the routine 400 ends in step 420. OTHER EMBODIMENTS While there have been shown what are presently considered to be preferred embodiments of the invention, it will be apparent to those skilled in the art that various changes and modifications can be made herein without departing from the scope of the invention as defined by the appended claims.
|
Same subclass Same class Consider this |
||||||||||