1. Field of Invention
The present invention relates generally to the field of query plan optimization in relational databases. More specifically, the present invention is related to optimizing query plans having conjunctive predicates.
2. Discussion of Related Art
Estimating the selectivity of predicates has always been a challenging task for a query optimizer in a relational database management system. A classic problem has been the lack of detailed information about the joint frequency distribution of attribute values in a table of interest. Perhaps ironically, the additional information now available to modern query optimizers has, in a certain sense, made the selectivity-estimation problem even more difficult.
Specifically, consider the problem of estimating the selectivity s1,2, . . . ,n of a conjunctive predicate of the form p1p2 . . . pn, where each pi is either a simple predicate, also known as a Boolean Factor (BF) of the form “column op literal”. Here column is a column name, op is a relational comparison operator such as “=”, “>”, or “LIKE”, and literal is a literal in the domain of the column; some examples of simple predicates are ‘make=“Honda”’ and ‘year>1984’. The phrase “selectivity of a predicate p” is used to indicate the fraction of rows in the table that satisfy p. In older optimizers, statistics are maintained on each individual column, so that the individual selectivities s1, s2, . . . , sn of p1, p2, . . . , pn are available. Such a query optimizer would then impose an independence assumption and estimate the desired selectivity as s1,2, . . . ,n=s1 * s2 * . . . * sn. Such estimates ignore correlations between attribute values, and consequently can be wildly inaccurate, often underestimating the true selectivity by orders of magnitude and leading to a poor choice of query execution plan (QEP).
As an example of using the uniformity assumption, assuming ten distinct values in the MAKE column, the selectivity of the predicate p1: ‘MAKE=“Honda”’ is estimated as s1= 1/10. Similarly, with 100 distinct values in the MODEL column and 10 distinct values in the COLOR column, we obtain s2= 1/100 for p2: MODEL=“Accord” and s3= 1/10 for p3: COLOR=“red”. Advanced commercial optimizers can improve upon these basic estimates by maintaining frequency histograms on the values in individual columns.
As indicated previously, in the absence of other information, current optimizers compute the selectivity of a conjunctive predicate using the independence assumption. For instance, p1,2,3=p1p2p3 is the predicate restricting a query to retrieve all red Honda Accords, and the selectivity of p1,2,3 is computed as s1,2,3=s1 * s2 * s3. In the previous example, the optimizer would estimate the selectivity of red Honda Accords to be 1/10000. As only Honda is the only automobile manufacturer to make Accords, there is a strong correlation between these two columns, and in this case, an actual functional dependency. The actual selectivity of p1,2 must be 1/100. Thus, a more appropriate estimate of the selectivity of p1,2,3 is 1/1000, which is one order of magnitude greater than the estimate using the independence assumption.
Ideally, to overcome the issues associated with the independence assumption, the optimizer should store a multidimensional joint frequency distribution for all of the columns in the database. In practice, the amount of storage required for the full distribution is exponentially large, making this approach infeasible. It has therefore been proposed to store selected multivariate statistics (MVS) that summarize important partial information about the joint distribution. Proposals have ranged from multidimensional histograms on selected columns to other, simpler forms of column-group statistics. Thus, for predicates p1, p2, . . . , pn, the optimizer typically has access to the individual selectivities s1, s2, . . . , sn as well as a limited collection of joint selectivities, such as s1,2, s3,5, and s2,3,4. The independence assumption is then used to “fill in the gaps” in the incomplete information; it is now possible to estimate the unknown selectivity s1,2,3 by s1,2 * s3.
A new and serious problem now arises, however. There may be multiple, non-equivalent ways of estimating the selectivity for a given predicate.
b) shows an alternative QEP that first applies p1p3 and then applies p2. If the optimizer also knows the selectivity s1,3 of p1p3, use of the independence assumption might yield a selectivity estimate sb1,2,3=s1,3 * s2. However, this would result in an inconsistency if, as is likely, sa1,2,3≠sb1,2,3. Potentially, there can be other choices, such as s1 * s2 * s3 or, if s2,3 is known, s1,2*s2,3/s2; the latter estimate amounts to a conditional independence assumption. Any choice of estimate will be arbitrary, since there is no supporting knowledge to justify ignoring a correlation or assuming conditional independence; such a choice would arbitrarily bias the optimizer toward choosing one plan over another. Even worse, if the optimizer does not utilize the same choice of estimate every time that it is required, then different plans will be costed inconsistently, leading to “apples and oranges” comparisons and unreliable plan choices.
Assuming that the QEP in
Cardinality Estimation for Conjunctive Predicates with MVS in DB2
In the absence of a complete set of statistical information, and due to the complexity and the cost of collecting such information, the DB2 UDB optimizer tries to exploit as much statistical data as is available. Shown below are four different scenarios in which it is assumed that the selectivities of simple predicates are known.
Case I: No statistics on conjunctive predicates are known.
For this trivial case, the selectivity of the conjunctive predicate is the product of the selectivities of the individual predicates.
Estimate s1,2,3 given s1, s2, and s3.
s1,2,3=s1 * s2 * s3.
Case II: Statistics on the conjunctive predicates are known and there exists a conjunctive predicate whose BFs match the set of predicates whose combined selectivity is to be determined.
For this simple case, the selectivity of the full conjunctive predicate is the selectivity of the conjunctive predicate whose BFs match the set of predicates in question.
(a) Estimate s1,2,3 given s1,2,3. s1,2,3=s1,2,3. (b) Estimate s1,2,3 given s1,2, s2,3, and s1,2,3. s1,2,3=s1,2,3.
Case III: Statistics on the conjunctive predicates are known and there exist some conjunctive predicates whose intersection is an empty set.
In this case, the selectivity of the conjunctive predicate is the product of the selectivities of all conjunctive predicates and the selectivities of the individual predicates not participating in any conjunctive predicate.
(a) Estimate s1,2,3,4,5 given s1,2 and s3,4 s1,2,3,4,5=s1,2 * s3,4 * s5. (b) Estimate s1,2,3,4,5 given s1,2,5 and s3,4 s1,2,3,4,5=s1,2,5*s3,4.
Although the ad hoc method described above ensures consistency, it ignores valuable knowledge, for example, the correlation between p1 and p3. Moreover, this method complicates the logic of the optimizer, because cumbersome bookkeeping is required to keep track of how an estimate was initially derived and to ensure that it will always be computed in the same way when costing other plans. Even worse, ignoring the known correlation between p1 and p3 also introduces bias towards certain QEPs: if, as is often the case with correlation, s1,3>>s1 * s3, and s1,2>>s1 * s2, and if s1,2 and s1,3 have comparable values, then the optimizer will be biased towards the plan in
One approach to query optimization is to utilize the maximum entropy principle. The maximum entropy principle models all that is known and assumes nothing about the unknown. It is a method for analyzing the available information in order to determine a unique epistemic probability distribution. Information theory defines for a probability distribution p a measure of uncertainty called entropy: H(p)=−Σipi log pi. The ME principle prescribes selection of the unique probability distribution that maximizes the entropy function H(p) consistently with respect to the known information. Entropy maximization without any additional information uses the single constraint that the sum of all probabilities is equal to one. The ME probability distribution then is the uniform distribution. When constraints only involve marginals of a multivariate distribution, the ME solution coincides with the independence assumption.
The present invention first formalizes the problem of consistency issues and bias during QEP evaluation in the presence of partial knowledge about the joint frequency distribution (as embodied in the available MVS). Subsequently, the ME approach to unbiased, efficient, and consistent selectivity estimation is applied, exploiting all available information, for cardinality estimation. The present invention incorporates partial information about multivariate information about multivariate distributions in a constrained optimization by modeling one or more multivariate distributions based on available statistical information, while taking into account all available distributional information and modeling one or more multivariate distributions for which no statistical information is available by maximizing information entropy for each multivariate distribution. Known selectivities are obtained from multidimensional histograms, column group statistics, index statistics, intermediate table statistics, or data learned during query execution.
The present invention also provides for a database query optimizer computing cardinalities of each partial query execution plan, wherein the database query optimizer comprises a selectivity computing unit to compute the selectivity of one or more predicates based on known statistical information in a given probability space, and a missing selectivity computing unit to compute missing selectivities using constrained optimization approach and entropy maximization over the entire probability space. In this embodiment, the computed selecitivities based on statistical information and computed missing selectivities based on constrained optimization approach and entropy maximization are used to provide improved cardinality estimates and reduced query execution times.
The present invention approach to incorporate partial information about multivariate distributions in a constrained optimization form comprises: modeling one or more multivariate distributions based on available statistical information in a probability space, and modeling one or more multivariate distributions for which no statistical information is available via a constrained optimization approach and entropy maximization over the entire probability space. The present invention takes into account all available statistical information.
The constrained optimization approach further comprises: defining a selectivity, xb, associated with each atom, iεN pib
In one embodiment, the present invention further partitions, during preprocessing, a predicate space into independent components of nonintersecting knowledge based on both available knowledge and additional information about relevance of selectivities within each component.
In another embodiment, the present invention further identifies, a priori, one or more atoms b for which input knowledge T implies that selectivity xb is zero, and excluding such atoms in computations of said constrained optimization approach.
In yet another embodiment, the present invention further partitions, during preprocessing, a predicate space into independent components of nonintersecting knowledge based on both available knowledge and additional information about relevance of selectivities within each component; identifies, a priori, one or more atoms b for which input knowledge T implies that selectivity xb is zero, and excludes such atoms in computations of said constrained optimization approach.
Computing a solution for all atoms xb, biε{0,1} that maximizes entropy while minimizing the function Σbε{0,1}
In one embodiment the constrained optimization solution is used for cardinality estimation in a database query optimizer.
a)-(c) illustrate a plurality of QEPs and associated selectivity estimates.
While this invention is illustrated and described in a preferred embodiment, the invention may be produced in many different configurations. There is depicted in the drawings, and will herein be described in detail, a preferred embodiment of the invention, with the understanding that the present disclosure is to be considered as an exemplification of the principles of the invention and the associated functional specifications for its construction and is not intended to limit the invention to the embodiment illustrated. Those skilled in the art will envision many other possible variations within the scope of the present invention.
The problem of selectivity estimation for conjunctive predicates, given partial MVS, is formalized using the following terminology. Let P={p1, . . . , pn} be a set of BFs. For any XN={1, . . . , n}, denote by pX the conjunctive predicate iεxpi. Let s be a probability measure over 2N, with the interpretation that sX is the selectivity of the predicate pX. Usually, for |X|=1, the histograms and column statistics from the system catalog determine sX and are all known. For |X|>1, the MVS may be stored in the data-base system catalog either as multidimensional histograms, index statistics, or some other form of column-group statistics or statistics on intermediate tables. In practice, sX is not known for all possible predicate combinations due to the exponential number of combinations of columns that can be used to define MVS. Suppose that sX is known for every X in some collection T⊂2 N. Note that the empty set ◯/ is part of T, as s◯/=1 when applying no predicates. Also note that 2N denotes the powerset of N. Then the selectivity estimation problem is to compute sX for X ε2N\T.
For each predicate p, write p1=p and p0=p. An atom is a term in disjunctive normal form (DNF) over the space of n predicates, i.e., a term of the form iεN pib
C(X)={bε{0,1}n|∀ iεX:bi=1} and
C(◯/)={0,1}n
For the predicates p1 and p1,2 the following are obtained:
C({1})={100, 110, 101, 111}
C({1,2})={110, 111}.
Additionally, for every T2N, we denote by P(b,T) the set of all known XεT 2N such that px has b as an atom in its DNF representation, i.e., P(b,T)={XεT|∀ iεX: bi=1} ∪ {◯/}. For the atom 011 and T=2{1,2,3} we obtain the set P(b,T)={{2}, {3}, {2,3}, ◯/}. Let xb denote the selectivity of an atom, with xb≧0. Given sX for XεT, we compute sX for X∉T according to the ME principle. To this end, we must solve the following constrained optimization problem:
minimize Σbε{0,1}
given the |T| constraints
ΣbεC(X) xb=sX for all XεT (2)
The constraints are the known selectivities. The solution is a probability distribution with the maximum value of uncertainty, entropy, subject to the constraints. One of the included constraints is s◯/=Σbε{0,1}
Referring now to
s
1
=x
100
+x
110
+x
101
+x
111=0.1 (I)
s
2
=x
010
+x
011
+x
110
+x
111=0.2 (II)
s
3
=x
001
+x
011
+x
101
+x
111=0.25 (III)
s
1,2
=x
110
+x
111=0.05 (IV)
s
1,3
=x
101
+x
111=0.03 (V)
s
◯/=Σbε{0,1}
The task of selectivity estimation is to now compute a solution for all atoms xb, b ε{0,1}3 that maximizes the entropy function −Σbε{0,1}
Shown in
Discussed in this section is an algorithm to compute this result efficiently for an arbitrary number n of simple predicates P and an arbitrary set T of conjunctions over simple predicates, where the probabilities of the conjunctions are known. To solve the constrained optimization problem in its general form, the method of Lagrange multipliers is first applied to obtain a system of optimality equations. Since the entropy function is concave, a solution of this system yields the unique solution of the optimization problem. Specifically, we associate a Lagrange multiplier λX with each known sX. This includes λ◯/, a multiplier associated with the constraint s◯/=1. The Lagrangian is given by
L(x,λ)=Σbε{0,1}
Taking derivatives and setting them equal to zero yields the optimality equations:
1n xb+1=ΣXεP(b,T)λX. (3)
The present invention also provides for a database query optimizer computing cardinalities of each partial query execution plan, wherein the database query optimizer comprises a selectivity computing unit to compute the selectivity of one or more predicates based on known statistical information in a given probability space, and a missing selectivity computing unit to compute missing selectivities using constrained optimization approach and entropy maximization over the entire probability space. In this embodiment, the computed selectivities based on statistical information and computed missing selectivities based on constrained optimization approach and entropy maximization are used to provide improved cardinality estimates and reduced query execution times.
The present invention's approach to incorporate partial information about multivariate distributions in a constrained optimization form comprises: modeling one or more multivariate distributions based on available statistical information in a probability space, and modeling one or more multivariate distributions for which no statistical information is available via a constrained optimization approach and entropy maximization over the entire probability space. The present invention takes into account all available statistical information.
The constrained optimization approach of the present invention further comprises: defining a selectivity, xb, associated with each atom, ΛiεN pib
Computing a solution for all atoms xb, biε{0,1} that maximizes entropy while minimizing the function Σbε{0,1}
Shown in
Using equation (4) to eliminate xb in equation (2), the following is obtained
ΣbεC(Y)ΠXεP(b,T)zX=sYe. (5)
It is usually not possible to analytically compute the Lagrange multipliers Z=(z1, . . . , z|T|) satisfying equation (5). A variant of the iterative scaling algorithm is used to efficiently obtain an approximate ME solution. The algorithm's main objective is to produce a new set Z of multipliers by iteratively refining the old Z0. Each iteration consists of |T| steps, during each of which the algorithm selects one YεT and changes only the multiplier zY associated with the constraint induced by sY, while keeping the other multipliers constant. Therefore, temporarily the current equation (5) is satisfied for Z.
From equation (5) the common Lagrange multiplier zY is factored out. Because this multiplier occurs in all summands on the left side, a solution for zY is found to obtain the iterative scaling equation:
During each iteration step, the algorithm visits all members of Z, updating the value according to (6) for each zY, including Y=◯/. The sequence of iterated solutions converges to some limit Z, and the algorithm terminates when the difference between successive solutions becomes negligible. Z is then used to compute the values for the unknown xb using equation (4), which, in turn, is used to compute sX for X∉T.
In this embodiment, the constraint optimization problem is solved in its general form making use of the Newton-Raphson algorithm to numerically find a root of a complicated function or curve. Newton's method relies on the observation that this tangent line often intersects with the x-axis at a closer approximation to the function's root. Hence when trying to seek any f(x)=0 one often can find a better approximation to the root by:
where xi is initialized with an initial guess. Applying this formula continuously on the obtained xi+1, will result in a convergence of the x-value, and in a final guess of the function's root. When solving the constraint optimization problem the following set of equations are utilized:
1n xb+1−Σ{λX|p(b)}−λΩ=0 (8)
ΣbεC(X)
These equations represent optimality conditions for the Maximum Entropy optimization problem. Equation (8) depends on the two variables xb and λX; the vector yX for the set of λX s, including . Equation (9) depends on xb. Furthermore, entire vector of functions in equation (8) is denoted by F=(f0, . . . ,fN−1) and the vector of functions in (9) by G=(g0, . . . ,g|T|−1). Roots are then approximated by Newton's method satisfying:
F
i(xb, yX)=0 (10)
G
j(xb)=0 (11)
The function can be expressed as polynomials in the neighborhood of the vectors xb and yX, hence can be expanded as the corresponding Taylor series:
The only remaining part necessary for consideration is the nonlinear unknown part of equations (12) expressed by: O(xb2)+O(yx2) and the nonlinear unknown part of (13) expressed as O(xb2). Since an approximation is sought, these nonlinear parts are neglected and the remaining system of linear equations is solved. Matrices in the Taylor series are Jacobian matrices. Let these Jacobian matrices be denoted as follows:
This gives rise to the following matrix notation A*h=d:
Subsequently, the system of linear equations in (14) is iteratively solved for the vector h in this Newton Raphson algorithm embodiment of the present invention. In each iterative step, the solution vector h is used, to update the current xb and yX:
x
b+1
=x
b
+Δx
b (15)
y
X+1
=y
X
+Δy
X (16)
Thus, the next approximation of the roots of (10) and (11) are obtained. In contrast to the iterative scaling algorithm, the xb and yX values are not changed one at a time, but all at once. The calculations of (14), (15), (16) are continuously performed, until the vectors Δxb and ΔyX of the solution vector h are approximately zero. At this point, the constraints of equations (8) and (9) are best approximated for all xb and yX, and hence satisfy the Maximum Entropy Principle. The final values of xb then can be used to calculate the unknown values sX for X∉T.
An important pre-processing step for the Newton Raphson approach is to properly define the starting values for the set of xb. To guarantee a convergence of the solution vector h, these initial values should not randomly be assigned. A sophisticated approach to solve this problem is given by linear programming. By minimizing an objective function subject to the constraints given by all known T, expressed as lower and upper boundaries, and the constraints 0≦xb≦1 expressed as upper and lower boundaries either, a simplex algorithm is utilized to determine the optimal starting point for the algorithm expressed in xb. Thus, not only is the convergence behavior of the algorithm improved, but also the steps necessary for iterations are minimized as well. This is because the values are already given a consistent distribution and well-defined direction.
Integration into DB2 Optimizer
The existing DB2 UDB optimizer precomputes selectivities from single-column statistics and MVS prior to plan enumeration, costing, and selection. When costing a plan that involves a conjunctive predicate p, the optimizer estimates p's selectivity by using a precomputed selectivity if available, or else combining several precomputed selectivities using the previously mentioned ad hoc method outlined. The optimizer takes precautions to avoid inconsistencies by keeping track of how a missing estimate for a predicate p was computed when it is first requested by a subplan during join enumeration. Subsequent requests for p's selectivity estimate will use this recorded information to ensure that the estimate is always derived in the same way. While this approach avoids inconsistencies, it does not utilize all available information as outlined previously and also causes an arbitrary bias towards particular plans, depending on how the selectivity was derived initially; this initial derivation depends in turn on the order in which selectivity estimates are requested during optimization.
The ME model of the present invention is seamlessly integrated into the DB2 UDB optimizer. The pre-computation phase is extended to not only compute selectivities based on statistical information, but also to pre-compute all missing selectivities using the iterative scaling algorithm. This pre-computation eliminates the need to use the heuristics for costing. It also eliminates the need to keep track of how selectivities were combined in order to avoid inconsistencies during estimation, because the ME estimates are inherently consistent with the statistics and generally applicable without causing any bias. The DB2 optimizer uses these pre-computed selectivity estimates during dynamic programming to compute the cardinalities of each partial QEP it considers. Extensions to the cardinality model enable the optimizer to use all available statistics in a consistent way, for all plans in the plan space. This improved knowledge results in better query plans and improved query execution times. Thus, the query optimizer's logic is simplified, as consistency checks and record-keeping are no longer necessary during cardinality estimation.
Even though the iterative scaling algorithm has a complexity of O(|T|·2|P|), we often do not have to execute scaling for the entire predicate set P. In many situations, we can compute the maximum entropy solution by partitioning P into several independent subsets and by executing the scaling algorithm on each subset. This approach reduces the complexity substantially. In the following, we discuss several partitioning strategies that make the iterative scaling algorithm feasible for even extremely complex queries.
P can be partitioned into non-overlapping subsets and simplified, depending on the available knowledge T, reducing the dimensionality of the problem space and thus the complexity of the scaling algorithm. Suppose that (a) for U⊂N, sU is known or can be computed (possibly with closed formulas) in a way consistent with the ME principle using only a subset of the information in T, and (b) the selectivities sX with X⊂U are not important. Then it is possible to replace the predicates denoted by U in P with a single predicate, reducing |P|, |T|, and the overall complexity of the problem. Below we provide several examples of when this is possible.
If T does not contain any information about a subset U of N, except for the single selectivities (i.e., |X|=1 for all XεU), then the selectivity sU can be computed as sU=ΠiεU si. All simple predicates pi, iεU, can then be removed from P, and all pieces of knowledge in T that intersect U can be removed, creating a new set of knowledge T*. Adding a new single predicate U with known selectivity thus reduces the complexity of iterative scaling to O(|T*|·2|P|−|U|+1). Similarly, for UεT and |U|>1, all simple predicates pi, iεU, can be removed from P and replaced by a new single predicate pU with known selectivity, again reducing the complexity of iterative scaling to O(|T|·2|P|−|U|+1). Also, for the special case in which S contains knowledge as described in Section 4.1, Case 2.2, the selectivity of U can be computed as described there, and the complexity can be reduced to O(|T*|·2|P|−2).
The general approach to reducing complexity uses a preprocessing step in which the predicate space P is partitioned into independent components of nonintersecting knowledge based on both T and additional information about the relevance of the selectivities within each component. The selectivity of each component is computed either by using closed formulas or running iterative scaling; the latter process is fast because each component has relatively low complexity.
It is sometimes possible to identify a priori certain atoms b for which the input knowledge T implies that the selectivity xb is zero. Iterative scaling can then exclude these atoms from the iteration step (8) in the algorithm. There are two ways to deduce from T that xb=0 for some bε{0,1}n.
These simple pre-processing steps significantly reduce the number of iterations needed for convergence. For several real-world queries and real-world data, the running time was improved by orders of magnitude, producing ME distributions for more than twenty predicates on a single table with sub-second execution times.
Additionally, the present invention provides for an article of manufacture comprising computer readable program code contained within implementing one or more modules to utilize all available conjunctive predicate information to predict selectivities during query plan optimization. Furthermore, the present invention includes a computer program code-based product, which is a storage medium having program code stored therein which can be used to instruct a computer to perform any of the methods associated with the present invention. The computer storage medium includes any of, but is not limited to, the following: CD-ROM, DVD, magnetic tape, optical disc, hard drive, floppy disk, ferroelectric memory, flash memory, ferromagnetic memory, optical storage, charge coupled devices, magnetic or optical cards, smart cards, EEPROM, EPROM, RAM, ROM, DRAM, SRAM, SDRAM, or any other appropriate static or dynamic memory or data storage devices.
Implemented in computer program code based products are software modules for: modeling one or more multivariate distributions based on available statistical information in a probability space, modeling one or more multivariate distributions for which no statistical information is available via a constrained optimization approach and entropy maximization over said probability space, and wherein said method takes into account all available statistical information.
A system and method has been shown in the above embodiments for the effective implementation of selectivity estimation for conjunctive predicates in the presence of partial knowledge about multivariate data distributions. While various preferred embodiments have been shown and described, it will be understood that there is no intent to limit the invention by such disclosure, but rather, it is intended to cover all modifications falling within the spirit and scope of the invention, as defined in the appended claims. For example, the present invention should not be limited by type of algorithm used to solve the constrained optimization problem, software/program, computing environment, or specific computing hardware.
The above enhancements are implemented in various computing environments. For example, the present invention may be implemented on a conventional IBM PC or equivalent, or relational database. All programming and data related thereto are stored in computer memory, static or dynamic, and may be retrieved by the user in any of: conventional computer storage or display (i.e., CRT) formats. The programming of the present invention may be implemented by one of skill in the art of database programming.
This application is a continuation of U.S. application Ser. No. 11/190,947, filed Jul. 28, 2005, pending.
Number | Date | Country | |
---|---|---|---|
Parent | 11190947 | Jul 2005 | US |
Child | 12042065 | US |