Patent Application
20080222093
The present invention relates to a method and system for automatically and adaptively determining execution plans for queries with parameter markers.
Query optimization is central to the efficient operation of a modern relational database system. The query optimizer is typically invoked every time a new query enters the system. The optimizer identifies an efficient execution plan for the query, based on available database statistics and cost functions for the database operators. In commercial systems, great care has been taken to reduce the overhead of query optimization. However, the task of the optimizer is complex, and the join ordering problem alone has complexity that is exponential in the number of tables [13] (see Appendix A for a list of cited references). As a result, the cost of optimization itself may represent a significant fraction of the elapsed time between query submission and answer generation.
If identical queries are submitted, the database system can cache the optimizer's plan the first time, and avoid reoptimization for subsequent query invocations. The query processor merely has to check for syntactic identity of the query with the cached query. This idea can be generalized to queries with parameters. Constants in the query are replaced with “bind variables” to generate a query template, in which the bind variables are parameters. The query processor can then cache a plan for a query template rather than for a query. As a result, frequently-submitted queries that differ only in the constants can avoid the cost of query optimization. Oracles provides such a facility [1], as do DB2® [17] and Microsoft® SQL Server [10].
There is a potential problem with this approach. A single plan is chosen for all instances of a query template. This plan, while optimal in a particular region of the parameter space, may be sub-optimal in another region. Savings achieved by not invoking the query optimizer may be nullified by the choice of a sub-optimal execution plan. In fact, often the difference in cost between the optimizer's plan and the cached plan exceeds the optimization time.
Modern transaction processing systems are often required to handle thousands of transactions per second. Consider, for example, a web-based Online Transaction Processing (OLTP) application, such as an on-line book store described by the TPC-W benchmark [3]. The system executes canned queries that share a small number of pre-defined templates, such as queries generated by the same HTML form, but differ in parameter values. An interactive system is expected to complete query processing and return results to the user in a short amount of time, often less than a second. A single user's queries may exhibit locality in the values of the submitted parameters, in which case a single query execution plan may be good enough. However, this locality is lost when many users interact with the system at any given time. Therefore, to ensure that an optimal plan is chosen for every query invocation, every instance of the query must be optimized anew. Many of these queries involve joins of several database tables and are thus non-trivial to optimize. In this setting, query optimization performed for every query instance adds significant overhead in terms of the overall execution time and CPU utilization.
A number of Parametric Query Optimization solutions have been proposed. The solution proposed by Ioannidis [13] fails to scale in the number of parameters, and does not directly handle continuous attributes. Geometric solutions proposed by Hulgeri and Sudarshan [12] are impractical because of the exponential explosion in the number of parameters and because they do not perform well with a typical real-life workload having multiple categorical attributes or where the underlying data is highly skewed.
Thus, there exists a need to overcome at least one of the preceding deficiencies and limitations of the related art.
The present invention provides a computer-based method of automatically and adaptively determining query execution plans for queries having parameter markers, the method comprising:
generating, by a computing system, a first classifier trained by an initial set of training points;
dynamically updating, by a computing system at a first runtime thereof, at least one of a workload of queries processed by a database of the computing system and database statistics collected by the database for computing a plurality of selectivities;
collecting, by a computing system in an off-line phase thereof, the off-line phase being subsequent to the first runtime, a new set of training points, the collecting responsive to a detection of the dynamically updating;
modifying, by the computing system in the off-line phase, the first classifier into a second classifier, the modifying including utilizing the new set of training points;
receiving, by the computing system at a second runtime thereof, the second runtime being subsequent to the off-line phase, a query for the database, the query including one or more predicates, each predicate including one or more parameter markers bound to one or more actual values, and the one or more predicates associated with one or more selectivities of the plurality of selectivities in a one-to-one correspondence; and
automatically determining a query execution plan by the computing system, the automatically determining including mapping, by the second classifier, the one or more selectivities into the query execution plan, wherein the query execution plan is included in an augmented set of training points, the augmented set including the initial set and the new set.
A system and a computer program product corresponding to the above-summarized method are also described and claimed herein.
Advantageously, the present invention provides machine learning-based algorithms that automatically and adaptively determine query execution plans for queries having parameter markers. Further, these machine learning-based algorithms accurately model the output of a query optimizer, scale gracefully with the number of query parameters, handle nonlinear boundaries in plan space, and achieve high prediction accuracy even when a limited amount of data is available for training.
The task of query optimization in modern relational database systems is important but can be computationally expensive. Parametric query optimization (PQO) has as its goal the prediction of optimal query execution plans based on historical results, without consulting the query optimizer. The machine learning techniques disclosed herein accurately model the output of a query optimizer for queries having parameter markers (a.k.a. parametric queries). The algorithms of the present invention scale gracefully with the number of query parameters, handle non-linear boundaries in plan space, and achieve high prediction accuracy even when a limited amount of data is available for training. Both predicted and actual query execution times are used for learning, and the experimental results disclosed herein demonstrate a total net win of a PQO-based method over a state-of-the-art query optimizer for some workloads. The present invention realizes savings not only in optimization time, but also in query execution time, for an over-all improvement by more than an order of magnitude in some cases.
PQO models the distribution of plans chosen in different regions of the parameter space of a query template [12], or of a set of templates [9]. A PQO system is trained off-line using a number of invocations of the query optimizer on instances of the query template. The result of such training is a function that, given an instance of the query parameters, identifies a plan that is likely to be the optimizer's choice. To be useful, this function must execute significantly faster than the optimizer. The function must also have a compact representation, so that a collection of such functions can be managed in memory.
Hulgeri and Sudarshan [14, 15] explicitly construct a geometric subdivision of the parameter space into convex regions corresponding to individual optimal plans. At runtime, when query parameters are known, an appropriate plan is chosen from the plan space. The technique disclosed herein replaces the explicit geometric constructions of [14, 15] with state-of-the-art machine learning techniques that analyze the training data and generate a set of classifiers that map parameter instances to plans.
Compared with earlier geometric approaches, the advantages of using machine learning techniques are: (a) training can be effective with much less training data and (b) the model scales gracefully with the number of parameters. Due to the compactness of the models, classifiers described herein have modest space requirements that are linear in the number of classes, and typically on the order of 10 KB per class. The techniques disclosed herein apply for both qualified and categorical attributes of any datatype. The Experimental Evaluation section presented below demonstrates that the methods disclosed herein accurately predict plans for uniform as well as for skewed data distributions. Further, the experimental results described below demonstrate that the testing functions (i.e., identifying a plan given parameter values) can be performed in less than a millisecond per query, which is typically much cheaper than the cost of query optimization.
The Experimental Evaluation section demonstrates a total net win of the present invention's algorithms compared to either choosing a single plan without reoptimization or reoptimizing each query instance. Both predicted and actual query execution times are used for learning, and achieve an over-all performance improvement by more than an order of magnitude for some workloads.
Moreover, the present invention discloses machine learning-based PQO techniques that are adaptive, so that whenever database statistics or query workload changes, the algorithms of the present invention collect additional training points and use those additional points to modify an old classifier into a new classifier.
The problem addressed by the present invention is defined in this section.
Definition 2.1. A bind variable is a variable that can appear in a predicate within an SQL WHERE clause. A query template with d parameters is an SQL statement containing d bind variables, each occurring exactly once. Bind variables are ordered according to their occurrence, and named b1, . . . , bd respectively. Herein {right arrow over (b)} is used as shorthand for the d-dimensional vector (b1, . . . , bd).
Definition 2.1 does not restrict the data type of parameters (i.e., the parameters may be numeric variables, strings, or even user-defined types). These variables must appear in a WHERE clause. The WHERE clause may belong to the outer query block, or to a nested subquery.
Example 2.1. The following query template has three parameters b1, b2, and b3.
Definition 2.2. Let Q be a query template with d parameters and let {right arrow over (p)} denote a d-dimensional vector of values of the appropriate types for those parameters. A query Q({right arrow over (p)}) is the parameter-free SQL statement derived from Q by replacing each bind variable in {right arrow over (b)} with the corresponding value from {right arrow over (p)}.
Parametric query optimization involves finding good plans for many queries that are derived from the same query template. The regular query optimizer's choice of plan is herein referred to as the optimal plan.
Definition 2.3. One is given query template Q, and a set of historical queries Q({right arrow over (p)}1), . . . , Q({right arrow over (p)}n) derived from Q according to some distribution of values {right arrow over (p)}i. For each query Q({right arrow over (p)}i), suppose that the optimal plan is Pi. The set of queries and their corresponding plans is called the training set, and n is the size of the training set. A training set (a.k.a. training dataset) includes items referred to herein synonymously as training data, training data points, training points and training examples.
A parametric query optimizer (PQO) has an off-line phase and an on-line phase. During the off-line phase, the PQO may read the training set and database statistics to generate some additional information I that is cached by the database. During the on-line phase, the PQO is given a previously unseen query derived from Q using the same parameter distribution that was used in the training set. The PQO is required to choose a valid plan for that query based on the current database statistics and I, but not the training set. The PQO is correct if the chosen plan is the optimal plan for the query. The PQO is permitted to return no plan, which means that it cannot identify the optimal plan with sufficient confidence.
As used herein, database statistics include summary information of a database table such as the number of rows that are in the table, the number of distinct values included in a particular column, the counts for each of those distinct values and the most frequent value of those distinct values, as well as histograms and other summary statistics related to a column or column pairs. Database statistics are used to compute the selectivity of a predicate. Selectivities change as a result of a change in database statistics, which in turn are caused by an update to one or more database tables.
An extended parametric query optimizer (EPQO) operates on an extended training set that contains the set of optimal plans chosen by the optimizer for the queries in the training set, and, for each query, the actual execution time according to each optimal plan. The goal of the EPQO is to choose the plan with the smallest actual execution time.
When the on-line phase returns a plan, the database system typically executes that plan without explicitly calling the regular query optimizer for the query. When no plan is returned, the database system will either optimize the query using the regular query optimizer, or use some standard default plan. A parametric query optimizer can be measured according to several metrics:
These metrics may vary depending on the size of the database, the available statistics and access structures, the query template, the distribution of points in the parameter space, and the size of the training set.
While it is possible that there may be changes in database statistics between the off-line and on-line phases of the PQO, it is hereinafter assumed that the statistics remain valid between phases. This is a fair assumption if off-line training is performed sufficiently regularly (e.g., each time the statistics themselves are recomputed).
In one embodiment, parametric query optimization is performed on the parameter values themselves. In a preferred embodiment, parametric query optimization instead uses as inputs the selectivities of the predicates involving those parameters, for the following reasons:
As used herein, a selectivity is defined as a property of a predicate of a database query. For example, a predicate that is a condition in a WHERE clause is applied to a table. In this example a selectivity indicates the percentage of rows of the table that satisfy the condition. Given binding values of parameter markers within a predicate of a query, selectivities can be computed for that predicate.
In the event that two columns of a table are correlated, the selectivities of two predicates will not be independent. This is a well-known problem in query optimization. One solution to this problem is to keep multidimensional statistics on combinations of columns, and to use these statistics for query optimization [14]. A similar approach applies to parametric query optimization by identifying groups of correlated predicates, and estimating a single combined selectivity for the correlated predicates, which would then be an additional input to the on-line and off-line phases. Hereinafter, the description of parametric query optimization uses single-predicate selectivities only.
Machine Learning is the study of computer algorithms that improve automatically through experience. Recent developments in this field have shown wide applicability of machine learning techniques [2, 7, 20].
A classifier is a computational procedure for deciding which among a number of classes an object belongs to, based on the object's properties (a.k.a. solving a classification problem). A binary classifier has two classes: the positive examples and the negative examples. In a classification problem, objects are represented as labeled feature vectors {({right arrow over (x)}i,yi)} generated from a target true function y=F({right arrow over (x)}), where {right arrow over (x)}i=(x1, x2, . . . , xk) is a list of features, and yiεY is the class label. In the machine learning approach for parametric query optimization in the present invention, the feature vector is a list of selectivity measures corresponding to the binding values for a query template, and labels are plans provided by the query optimizer. The task of inductive learning is to construct a model y=ƒ({right arrow over (x)}) to approximate the true function F.
In the interest of replacing the query optimizer function F with an inductive model ƒ, the present invention discloses modeling techniques that are accurate in their prediction, and efficient in computation and in memory consumption during both model construction and query plan prediction phases.
Error of a machine learning algorithm can be decomposed into bias, variance, and noise, which are discussed in the remainder of this section.
There is no noise in the problem defined in Section 2. It is assumed that the query optimizer function F is deterministic: given the same set of parameter selectivities for a particular query template, the query optimizer will always return the same plan.
To achieve high accuracy, an algorithm is described herein that produces a function ƒ that closely approximates the true function F. Such an algorithm is said to have low bias or systematic error. Several machine learning techniques were excluded because of high bias. Regression techniques predict continuous values and are inappropriate for the present invention's domain in which there are clear discontinuities because the space of plan labels is discrete. The traditional single decision tree algorithm uses linear boundaries and is excluded from this study since the true decision boundary is non-linear in general. A wide variety of clustering algorithms is described in the machine learning literature. Clustering is typically used in unsupervised learning, where class labels are not available. However, PQO is a supervised learning problem—class labels are the query execution plans.
The final consideration is the effect of a small number of training examples on variance. One way to reduce variance is to construct multiple uncorrelated models and combine their predictions via some form of voting [21]. In this work, two algorithms are implemented, each using a different voting method, and compare their accuracy: AdaBoost [20] weighs the training data and constructs multiple classifiers from each weighted sample, while Random Decision Trees [6] utilize a randomization approach during model construction.
Experiments with other machine learning algorithms yielded limited success. Support Vector Machines (SVM) [22] proved to be sensitive to the geometric shape of the plan space, and required kernel selection and parameter tuning for each query. The Naive Bayes classifier works with discrete parameters and is very efficient during training. However, the Naive Bayes classifier also required tuning for each query because it was sensitive to the discretization: the way continuous features were mapped into buckets.
Boosting is a general and provably effective method for improving the accuracy of any learning algorithm. AdaBoost [8] is a widely accepted boosting algorithm that can improve the accuracy of a collection of “weak” learners and produce an arbitrarily accurate “strong” learner. The weak learners are required to be only slightly better than random guessing (i.e., more than 50% accurate in the case of binary classification). AdaBoost was extended in [20] to handle confidence-rated predictions, where weak learners output both the predicted label and a confidence measure as their classification hypothesis.
AdaBoost calls each weak learner repeatedly in a series of rounds t=1, . . . , T. There are various ways to choose T, and one way is described in Section 4.3. The main idea of the AdaBoost algorithm of the present invention is to maintain a distribution of weights over the training set. Initially, the weights of all points are equal. On round t, each weak learner is measured by its error εt. The error is the sum of weights of (a) mis-classified points weighted by the confidence rating c of the prediction, and (b) correctly classified points weighted by 1−c. The weak learner with the lowest error is chosen and is herein referred to as Wt. The weights of Wt's incorrectly classified examples are exponentially increased, and the weights of Wt's correctly classified examples are exponentially decreased. In this way, the weak learners are forced to focus on the difficult examples in the training set. The process is repeated with the new weights. The final strong hypothesis H is the αt-weighted majority vote of W1, . . . , WT, where
AdaBoost has provable bounds on generalization error (i.e., the error on unseen examples that come from the same distribution as the training examples).
AdaBoost is a binary classifier, while PQO is a multi-class problem. The basic AdaBoost algorithm has been extended to incorporate multi-class classification, and is known as AdaBoost.M2 [7]. The AdaBoost.M2 algorithm was implemented but failed to achieve fast convergence. The present invention therefore utilizes an alternative way to adapt AdaBoost to multi-class problems.
The simplest way to adapt a binary classifier to a multi-class problem is by using the “one-vs-all” approach, where a single classifier is built for every class. One-vs-all classification, while simple, is often unable to provide adequate prediction accuracy. There is no classification confidence measure, and if a point is classified positively by more than one classifier, no mechanism exists to break the tie.
The use of error-correcting output codes (ECOC) can improve prediction accuracy of binary one-vs-all classifiers on multi-class problems [5]. An ECOC is a matrix of binary values such as the matrix shown in Table 1. The length of the code is the number of columns in the matrix, and the number of rows corresponds to the number of classes in the learning problem. A single binary classifier (e.g., AdaBoost), is trained for each column in the matrix, with points from classes that have a 1 in the corresponding entry serving as positive examples, and those from classes with a 0 entry as negative examples. During testing, the incoming example is evaluated by every binary classifier, and a bit-string of classification outcomes is obtained. This string is then compared to every row in the matrix, and the Hamming distance (i.e., the number of bits that differ) is calculated. The point is assigned to the class closest in terms of Hamming distance. There is a trade-off between the improvement in prediction accuracy and training time: a greater Hamming distance can be achieved for longer codes, but more binary classifiers need to be trained.
A decision tree is a classifier with a hierarchy of decisions made at each node of the tree. One traverses the tree from root to leaf, choosing the appropriate child based on the decision criterion coded into each node. For example, a node has children for different ranges of the selectivity of the first predicate of a query template.
The Random Decision Tree (RDT) method constructs multiple decision trees “randomly.” The construction selects a feature at random from among those features not yet used in higher levels of the tree. In one embodiment, a feature of the RDT method is a predicate selectivity. A partitioning value for that feature is also selected at random from a distribution. Training data points from the node are then distributed to the node's children. Construction stops when the depth reaches a certain limit, when the number of data points in a node is sufficiently small, or when all points in a node have the same label (i.e., the node is a pure node). The randomized construction is unlike traditional single decision tree algorithms (e.g., C4.5 and ID3 [18]) that use gain functions to choose features and thresholds for tree nodes.
During the on-line phase, each tree is traversed using the actual query selectivities, to arrive at a leaf node L containing a number of plans. A posterior probability is calculated for each plan P. This probability is simply the proportion of the training points in L that are labeled with P. The posterior probabilities are averaged across all trees, and the plan with the highest average is output.
The RDT method reliably estimates probabilities, closely approximates non-linear boundaries, and reduces variance when the number of training examples is small [6].
This section describes the application of AdaBoost and Random Decision Trees to parametric query optimization. Hereinafter, the PQO techniques disclosed herein that apply AdaBoost and RDT are referred to collectively as the algorithms of the present invention. The algorithms of the present invention were implemented on top of an off-the-shelf commercial relational database system, and require no modifications to the query optimizer or any other part of the database system.
In step 206, a new user query 106 is issued and received by query plan cache 110. The selectivities of each predicate with parameter markers are given as input to the classifier built in step 204. Using the selectivities as input, the classifier outputs a predicted query execution plan. In step 208, database system 102 (see
Steps 202 and 204 are performed in a first off-line phase (i.e., not at runtime). Step 206 is performed at runtime. Steps 208 and 210 provide the adaptive capabilities of the present invention in a second off-line phase. The second off-line phase is shorter than the first off-line phase because the required training in the second off-line phase is incremental and trains only the new training data points 112 (see
This section presents an overview of the processes of building, using and adapting an AdaBoost-based classifier for the PQO technique of the present invention. Unless otherwise specified, the steps in each of these processes are performed by query plan learner 104 (see
4.1.1 Building a Multi-Class Classifier using Boosting
In step 306, the number of classes in the problem is set to k, where k=(number of classes with greater than the predetermined threshold coverage)+1. Step 308 generates an ECOC table having a length of 2*k. The classifier building process of
4.1.2 Training a Binary Classifier with AdaBoost with Confidence-Rated Predictions
The determination of the number of rounds for the training phase includes computing the overall error on the training data every X rounds, where X is predetermined (e.g., X=20 rounds). If the computed error is below a predetermined rate (e.g., 5%), then the training phase stops. If the error is above the predetermined rate, then the training continues for another X rounds. If the error is the same as it was X rounds ago, then training stops. If the error is higher than it was X rounds ago, then the algorithm rolls back to the state X rounds ago, and training stops.
In step 322, the model is output. If T is the number of training rounds in the training phase of steps 318-321, then the model consists of T weak learners and T weights, one weight for each weak learner. Each of the T weak learners is chosen at step 319 in one of the rounds of training. Each of the T weights is assigned at step 320 in one of the training rounds. The binary classifier training process of
In step 340, the new training data points 112 (see
In the workload change adaptation process of
In the workload change adaptation process of
If necessary based on predefined criteria, data points from the new class are assigned higher weights in step 366 to ensure that the binary classification algorithms concentrate on the new data points. In step 368, no changes to the ECOC table are made and all binary classifiers are re-trained for a predetermined number of rounds (i.e., are not fully trained) to incorporate the new training data. The adaptation process of
When deciding whether class A or class B is more likely on a single dimension d, weak learners for A_d and B_d are queried, and both return a probability. The class with the highest probability is chosen and the winning probability is returned as the weight of the hypothesis.
This section presents an overview of the processes of building, using and adapting a RDT-based classifier for the PQO technique of the present invention.
In step 404, at each internal node of the current RDT being constructed, query plan learner 104 (see
In step 420, each posterior probability from the multiple RDTs is output and the outputted posterior probabilities are averaged across the multiple RDTs for each query execution plan. In step 422, a loss function is used to choose an optimal average posterior probability and the query execution plan associated therewith is selected as the prediction of the output of the query optimizer of database system 102 (see
In one embodiment, an ECOC length between 2*c and 3*c is used, where c is the number of classes.
Choosing the weak learner appropriate for the domain was the main challenge of the AdaBoost implementation. The choice of a weak learner was guided by the observation that the selectivity of a single parameter can be used to discriminate between an optimal and a sub-optimal plan for a query in a particular selectivity region.
Consider a query 500 in
For queries with d parameters, the optimizer chooses a query execution plan based on individual selectivities and/or on products of any subset of the d selectivities. Products of selectivities naturally correspond to estimates of the relative size of intermediate or final results during plan execution. Explicit enumeration of all possible products (i.e., of all possible subsets of parameters) is exponential. The weak learners in the AdaBoost implementation are designed to avoid the exponential explosion and to consider the selectivity of one parameter at a time. For Plan 1 in
The design of the weak learners in the AdaBoost implementation is based on the above simple observation. Each weak learner is a discrete (i.e., bucketized) vector of weighted probabilities. The probabilities represent the likelihood that a particular plan is chosen by the optimizer when the selectivity falls within each bucket. The weights are adjusted over time by the AdaBoost meta-learner. A weak learner of this kind is defined for each parameter, and for each plan. Such weak learners are unary—they always claim that the point is a member of the class. The weak learners encode the strength of the claim in their confidence measure, which is proportional to the appropriate element of the weighted probability vector. The probability distribution is calculated using the conditional probability presented below:
As is apparent from formula (1), the AdaBoost implementation needs to consider only how many points that fall within the selectivity range of interest also map to the particular plan label. The initial distributions for Plans 1, 2 and 3 given selectivities of b1 are listed in graphs 700 in
The algorithm used in the implementation of AdaBoost has two parameters that influence prediction accuracy: the number of training rounds T for each binary classifier and the number of buckets in the probability distributions B. Each of these parameters is discussed below.
AdaBoost adjusts the weights of correctly and incorrectly classified points exponentially, and provides for exponential convergence. Each binary classifier is trained in increments of T rounds, and prediction accuracy is measured with respect to the training set after each T rounds. In one embodiment, the predefined increment of T rounds is 20 rounds. Training continues until one of the following conditions is met: (a) the total number of rounds is equal to a predefined round limit (e.g., 100 rounds), (b) accuracy on the training set reaches a predefined accuracy threshold (e.g., 95%) or (c) accuracy on the training set does not improve compared to T rounds ago.
The AdaBoost implementation uses equi-width histograms with B equal to a predefined number of buckets (e.g., B=20 buckets), each encompassing a predefined percentage (e.g., 5%) of the selectivity range. The setting of B=20 buckets, each encompassing 5% of the selectivity range works well for all query templates in the experiments described below in Section 5.
To adapt RDT for query plan prediction, one important improvement is made based on knowledge about the behavior of the optimizer. While predicates are still chosen at random, the decision threshold is no longer chosen at random. Instead, for a randomly chosen predicate, a threshold with the highest information gain is computed. In this way, it is more likely to generate pure nodes, which leads to smaller trees. The adaptation of RDT in the present invention is more efficient than Breiman's Random Forest (RF) [4]. The RF algorithm uses computationally intensive bootstrap sampling (i.e., random sampling with replacement) from the training set, while RDT uses the original training set. Additionally, RF evaluates information gain for a set of features, while RDT considers a single feature at a time. Finally, RF uses voting to classify a point, and RDT uses averaged probabilities, which may benefit prediction accuracy, particularly for multi-class problems [6].
In the RDT implementation of the present invention, the minimum number of training examples per leaf node is chosen to be 2, which is the default in traditional decision tree methods and a predefined number of trees are constructed. In one embodiment, 10 trees are constructed due to a reported result that there is no significant improvement in accuracy when more than 10 trees are constructed [6]. The depth of each tree is limited according to predefined criteria. In one embodiment, the tree depth is limited to 5 times the number of features, which allows a partitioning of the range of selectivity measures into up to 6 ranges. Since the number of training examples is at most a few hundred, and no empty nodes are generated, each tree is expected to be reasonable in size.
The accuracy of a classifier depends largely on the availability of sufficient training data. In many domains the number of training examples per class is highly non-uniform: some classes are represented by many more examples than others. During the experimental evaluation of AdaBoost it was noted that even on large training sets, the algorithm achieved higher prediction accuracy for the more common classes than for the less common. Not only were the points from the less common classes classified incorrectly, but a disproportionately large number of points were erroneously attributed to the less common classes.
It is often better to make no prediction for a test point than to classify that point incorrectly. For PQO, a misclassified point may incur a misprediction penalty that by far exceeds the optimization overhead. The algorithm of the present invention requires a reliable measure of prediction confidence that directs the algorithm to give up on a point and generate an uncertain classification (i.e., a classification of “no plan”).
The Hamming distance as a measure of prediction confidence was attempted to be used. The algorithm was trained as before, but test points that fell outside the Hamming distance threshold of the closest class were classified as uncertain. It was observed that previously correctly-classified points were now being classified as uncertain at approximately the same rate as the misclassified points, irrespective of the Hamming distance threshold. As a result, the overall prediction accuracy did not increase. It was concluded that the mechanism for deciding the confidence of a prediction needed to be incorporated during the training phase to ensure the proper generation of training classes.
During the training phase, all points that represent uncommon classes (i.e., classes of size smaller than a predefined threshold S) are placed into a single unreliable class. Hereinafter, S is also referred to as the class size threshold. The classifier is then trained as before, except that there are now fewer training classes, with all uncommon plans now mapped to a single class. During the test phase, all points that are classified as belonging to the unreliable class are now given an uncertain classification. In one embodiment, to determine a class size threshold, cross-validation is used on several datasets with a training set of size 500. The term cross-validation is used to describe the choice of some aspect of the machine learning model empirically, using available training data. By gradually increasing the threshold, it was found that plans represented by fewer than 20 query instances were too small to classify reliably. For a training set of size 500, this corresponds to not making a prediction for plans that take up less than 4% of the plan space. It was found that the class size threshold of 20 points worked well for all query templates and training set sizes, and it is concluded that this parameter does not need to be learned with each new query template.
The training time of the algorithm is linear in the number of classes, and is reduced by grouping all uncommon plans together, as there are now fewer binary classifiers to train.
This technique reduces the misprediction rate by 5-10% for most queries, at the cost of some “no prediction” outcomes. For a mispredicted plan, the penalty depends on how suboptimal the chosen plan is. For a “no prediction” outcome, the optimizer is called and then an optimal plan is used. The penalty for the “no prediction” outcome is therefore the cost of running the optimizer. The only way to compare a two-outcome classifier with a three-outcome classifier is to compute the total expected time with appropriate empirical measurements of the appropriate penalties. When this comparison was done for AdaBoost, it was found that the three-outcome classifier performed as well or better than the two-outcome classifier for almost all queries. Therefore, only experimental results for the three-outcome version of AdaBoost are reported in Section 5.
A three-outcome version of RDTs was not implemented primarily because the two-outcome version performed very well, as is shown in Section 5.
Given a query template, and domains and distribution (e.g., uniform, Gaussian, or Zipf) of the query parameters, a training set of the appropriate size is automatically generated for the query. For each point in the parameter space, the regular query optimizer is called and returns the query execution plan. The query is not executed. Alternatively, the real-time query workload is recorded, along with the optimizer's predictions. This dataset is then used to train the classification algorithm. Training is computationally expensive and happens off-line. This is in line with how most query optimizers collect statistics.
For AdaBoost, during this phase each binary classifier builds the model of the training dataset. A binary classifier executes between 20 and 100 training rounds T, and considers each available weak learner on every round. The number of binary classifiers is linear in the number of classes c, and the number of weak learners is exactly the number of query parameters d. Each weak learner computes probability distributions over the dataset, and is linear in the size of the dataset n. The time complexity of the training phase is therefore O(T*d*c*n). The output of training is the set of T discrete probability distributions (each containing B=20 entries) and T weights for each class, along with the optimal execution plan for that class. The space complexity of the model is O(T*B*c).
For Random Decision Trees, k random trees are independently constructed during the training phase. It has been reported in [6] that k=10 returns satisfactory accuracy. The time complexity of training is O(k*d*log d*n*log n). The output of the training phase has space complexity of O(d*log*n).
When a new query matches an already-seen template for which the classifier has been trained, the instance will be classified on-line by the algorithm. If the classification outcome points to one of the learned classes, the plan for that class will be returned to the query execution module, and the query optimizer will not be invoked. If, however, an uncertain outcome is generated, the optimizer will be called upon to generate an execution plan. Optionally, query parameters and plan information can be stored for future re-training.
The performance of the classifiers of the present invention was evaluated on the DB2 Universal Database version 8.2, a commercial RDBMS with a cost-based optimizer. All experiments were executed on a Pentium 4 3.0 GHz CPU, with 512 MB of RAM, running the Linux operating system. The optimizer was executing at the highest optimization level (i.e., level 9). The performance of AdaBoost and RDT were evaluated with respect to two database schemas. The first schema conforms to the TPC-W benchmark [3], with uniformly distributed data. Two database sizes were used for the TPC-W experiments: 20 MB and 100 MB. Varying the size of the database resulted in a different number and types of optimal plans, and the shape of the plan space varied significantly. The second schema is the DMV database [15]: a synthetic database with skewed data distributions and correlations. A 60 MB database was used for the second schema. In the results discussed below, the scale of an experiment refers to the database size in megabytes.
The choice of relatively small databases is consistent with the target applications that demand very fast query response. When query execution time is small, optimization time becomes a significant component of the overall time. The space overhead of the algorithms of the present invention was no more than 30 KB for all experiments.
In this section, the performance of the present invention's algorithms is evaluated with respect to 6 query templates. The templates are chosen so that optimization time constitutes a significant portion of the overall time. These are the queries that can potentially benefit from Parametric Query Optimization. Although many more templates were considered, the ones chosen to be highlighted herein are the cases where the algorithms of the present invention make the most difference. For many other queries, such as queries that are trivial to optimize or take a long time to execute, the opportunities for improvement were small. To show the applicability of the algorithms of the present invention to a variety of query templates, templates are chosen with a varying number and types of parameters, and with a varying number of joins.
The template TPCW-1 is template 500 in
For each of the above templates in FIGS. 5 and 8-12, up to 1000 random parameter combinations are generated. The distribution of each parameter matches the distribution of the column to which the parameter is compared. In some cases it is impossible to generate 1000 distinct parameter combinations because of the limited number of distinct values in the database. In such cases the largest possible dataset is generated. The training set of size 200 is chosen uniformly at random from the dataset; the remainder of the dataset is used for testing. The measurements used in this section represent averages over 5 random splits of the dataset into training and testing sets.
A successful machine learning technique achieves high prediction accuracy with a limited number of training examples. The effects of training set size on accuracy were studied and it was observed that the classification accuracy for RDTs improves only 0-3% when the size of the training set is increased beyond 200 points. For RDTs, all plans were considered during training and testing. In one embodiment, the implementation of AdaBoost never chooses a plan if that plan is represented by fewer than 20 points in the training set. Increasing the size of the training set reduces the proportion of unclassified points, but does not significantly change the ratio of correct to incorrect classifications among the points for which classification was attempted. Increasing the size of the training set would enhance the performance of AdaBoost. However, the results presented herein are based on this limited training set because (a) it is less time-consuming to train on a small training set, and (b) for larger training sets, only a marginal improvement in classification accuracy for RDTs was observed. For a fair comparison, the training set size is kept constant, and 200 training points are used for all experiments in this section.
The accuracy and overall performance improvement achieved by the algorithms of the present invention for the first 4 query templates (i.e., the templates in FIGS. 5 and 8-10) is presented in Sections 5.3 and 5.4. The final two templates, DMV-2 and DMV-3 (i.e., the templates in
5.3 Learning from Optimizer Predictions
In the first part of the experiments, classifiers were trained on the plan space induced by the query optimizer, assuming that the query execution plan chosen by the optimizer is in fact the optimal plan for any given query instance. Table 1300 in
The third column (i.e., the column labeled Plans) of table 1300 shows the number of distinct query plans that were generated using 200 training queries. The fourth column (i.e., the column labeled Common Plans) of table 1300 shows how many of those distinct query plans met the AdaBoost threshold that requires a plan to be chosen at least 20 times out of 200. The last column (i.e., the column labeled Coverage) of table 1300 shows the fraction of the training points that employ plans that are above threshold. Table 1300 shows, for example, that for query DMV-1, 87% of the 200 training points optimized to one of the 3 plans that are above threshold.
Accuracy results are summarized in table 1400 in
The classifiers choose a suboptimal plan some of the time. Nevertheless, it is possible that this suboptimal plan has a performance comparable to the optimal plan. The remainder of this section investigates how much worse a suboptimal plan can be, and to what extent the misprediction penalty can be offset by the reduction in optimization overhead. The performance of the classifiers are compared with two alternatives: (a) the cost of the plan generated by the optimizer plus the optimization overhead for every query (OPT) and (b) total execution time of every query according to every plan in the optimal plan space, weighted by the frequency of each plan, plus the cost of a single query optimization (AVG). The second alternative represents the situation where a single plan is chosen for all instances of a query template. Typically, the first plan that is encountered by the system is chosen, and the probability of any particular plan to be encountered first is approximately proportional to the frequency of that plan in the dataset.
It would be difficult to ask the commercial system being used to estimate the cost of an arbitrary plan on different parameters. Thus, the actual execution time of each query is measured. It was also discovered that, given a single query template, optimization time may vary depending on the plan that is ultimately chosen by the optimizer. This is because the choice of a plan affects how many plans are pruned at intermediate stages of plan generation. Therefore, it is not assumed that optimization time is constant for a given query template, and the optimization time is measured per plan.
Table 1500 in
In table 1500, total times for both RDT and AdaBoost include the execution times of all points in the test set according to the chosen plan plus the overhead of the classification algorithms, which is on the order of 10 microseconds per point for both algorithms. AdaBoost also includes the optimization time for the points for which an uncertain prediction was generated, and the cost of the optimal plan is used in this case.
As is apparent from
It appears more difficult to outperform AVG for some queries. Despite high prediction accuracy on TPCW-3, even RDT performs 8% worse than AVG. This query has the highest optimization overhead: optimization takes a total of 29.5 sec, and execution—a total of 0.3 sec for 800 queries in the test set. Furthermore, there is no significant difference between the plans—executing all queries according to any single plan differs only marginally from optimal execution.
There is another important reason that limits the performance improvements realized by the algorithms of the present invention. When training on the plan space induced by the optimizer, it is expected that the plans chosen by the optimizer are indeed optimal. This characterization of the plans chosen by the optimizer, however, is not always the case. Section 5.4 explores a method for correcting the optimizer.
5.4 Learning from Observation
The task of a relational optimizer is to choose a reasonably good plan among an exponential number of query execution plans. The optimizer does not explore the entire plan space, and brings no guarantee of global plan optimality. Additionally, the optimizer is guided by statistics that, even if accurate, represent only a summary of the data distributions and selectivities. For frequently executed short-running queries, it may be possible to build in “corrections” to the optimizer based on past execution time data. This is the Extended Parametric Query Optimization (EPQO) version of the problem as described above in Definition 3.
In the remainder of this section, the query execution plans returned by the optimizer are referred to as optimal plans, and plans that were observed to perform best are referred to as best plans.
To check the performance of the optimizer, actual query execution times are reviewed for every query according to every plan that was optimal in some region of the selectivity space. It is observed that this method introduces some noise—executing a particular query according to the same plan several times may yield slightly different execution times, especially if queries are fast-running. Additionally, for a particular query the best plan may perform only marginally better than the optimal plan. If all points are re-labeled with best labels, the resulting plan space is highly irregular and very difficult to learn. A simple method is used for reducing the noise. A point is re-labeled with its best label only if the best plan outperforms the optimal plan by more than 30%. The new labeling is referred to as HYBRID.
Having re-labeled the space, the total execution times of all queries in the training set according to the optimal labeling are compared with the execution times according to the hybrid labeling. These results are summarized in table 1600 in
Prediction accuracy of AdaBoost and RDT on the test sets is summarized in table 1700 in
Table 1800 in
Based on these results, machine learning techniques can benefit from learning on the HYBRID space. Generating HYBRID involves optimizing and executing each query in the training set according to every plan that is optimal in some region of the plan space. This is much more expensive than generating OPT, which does not involve any query executions, but is not prohibitively expensive for the target workloads of fast-running queries.
The HYBRID space can be used to improve the performance of the algorithms of the present invention. It can also be used to determine whether to use one of the algorithms of the present invention, or to simply optimize once and execute all queries according to a single plan, and perhaps suggest a good candidate plan. This approach works best for queries for which there is no significant difference between individual plans, such as TPCW-3 on the 20 MB database.
5.5 Queries with a Sparse Plan Space
This section focuses on query templates DMV-2 and DMV-3. For queries that utilize only, or mostly, equality predicates, many different parameter values will map to the same set of selectivities. In the extreme case, for a query that contains only equality predicates on uniformly distributed attributes, all points will map to a single set of selectivity measures, which in turn maps to a single optimal plan. It was observed that 165 unique points were present in the selectivity space of DMV-2, mapping to 9 plans. For DMV-3, 24 unique points represented 6 distinct plans.
For such queries, a much simpler method to construct a query plan cache can be used. With a sufficiently large training set, not only all classes, but also all points, will be represented in the training set. A hash table that stores the mapping between the selectivities and the optimal plans will suffice; no learning is required, since the algorithm will never have to deal with any previously unseen selectivities.
However, it is difficult to tell if the training set is exhaustive by looking at that set alone. This final set of experiments demonstrates that the algorithms of the present invention can be used for queries such as DMV-2 and DMV-3. In both cases, RDT and AdaBoost are trained on all available points, and return the prediction accuracy on the training sets.
RDT achieves very high prediction accuracy on both sets: 163 out of 165 points are classified correctly for DMV-2, and 22 out of 24 points are correct for DMV-3. AdaBoost does well on DMV-3, with only 1 misprediction out of 24, but fails to learn the space for DMV-2—the algorithm only attempts to make a prediction 7% of the time, and is 75% accurate when it does predict. However, this failure to learn is apparent at the end of the training phase. The accuracy on the training set indicates that the classifier did not learn the space sufficiently well and therefore should not be used.
As has been demonstrated thus far, machine learning techniques can be used to achieve an overall performance improvement for some queries. However, these techniques are only practical if it is also possible to automatically identify cases where using the algorithms of the present invention is not beneficial, and even impairs performance.
For AdaBoost, the accuracy on the training set is a good indication of accuracy on unseen data, provided that it comes from the same distribution as the training set. If the classifier fails to achieve the desired accuracy on the training set, or returns the “no plan” prediction too often (as for DMV-2), the system declares that the classifier cannot be used, and possibly collects more training data.
Random Decision Trees always achieve very high accuracy on the training dataset. However, this accuracy carries no guarantee that the algorithm will perform well on unseen data. A simple validation technique can be adapted to verify whether an RDT classifier is ready to be used for a particular query template. Similar to the experiments disclosed herein, the training dataset can be split into two subsets. RDT can be trained on the first subset, and validated against the second.
As an additional safety measure that ensures that the classifiers are up-to-date and accurate, the optimizer is invoked when the system is idle to validate a number of recent predictions made by the algorithms of the present invention. If it is determined that the classifier no longer achieves the desired accuracy, it can be re-trained on this more recent data set.
The algorithms of the present invention targets workloads of a large number of fast-executing parametric queries, with the objective of optimizing the aggregate behavior of the system. Thus, the implementation described herein targets average-case performance. An alternative, and possibly conflicting, objective may be to target worst-case performance. Please see [15] for a discussion of risk vs. opportunity considerations. The machine learning methods disclosed herein can be configured to optimize for different objectives in a simple yet principled way: by adjusting the loss function used during training. To optimize for the average case, the present invention penalizes all mis-predictions during the training phase equally. If the objective of the present invention was to bound worst-case performance, the loss function would incorporate the actual mis-prediction penalty, possibly in an exponential way.
The experimental evaluation herein reports average-case results. However, worst-case results were also measured. The worst-case behavior is strictly better than OPT for most queries. For some worst-case queries, the cost of the algorithms of the present invention is a significant multiple of the cost of OPT. These results, however, cannot be compared directly to the maximum degradation results reported in AniPQO [11]. The measurements for the present invention are based on actual execution time, while AniPQO is based on optimizer estimates that are more likely to be stable at plan boundaries, where most mis-predictions take place. A full comparison of the present invention to AniPQO is included in the Section 7. Additionally, the worst-case performance for the present invention may be biased by the fact that all queries used herein have very short execution times.
Ioannidis and Kang [14] describe the applicability of several randomized algorithms to Parametric Query Optimization, applied to run-time parameters that take on discrete values, such as the number of buffer pages allocated to a query and types of available indexes. The algorithms of the present invention work for both discrete and continuous parameters.
Reddy and Haritsa [19] study properties of plan spaces on a suite of commercial query optimizers. Reddy and Haritsa challenge the foundations underlying traditional geometric PQO approaches by observing that plan optimality regions may not be convex and may not be limited to a single contiguous region. These characteristics of plan optimality regions were observed for some of the query templates described herein (e.g., DMV-1), and the algorithms of the present invention were able to learn such plan spaces successfully.
Hulgeri and Sudarshan [12] propose AniPQO: a geometric solution to PQO that samples the parameter space and attempts to approximate boundaries between regions of plan optimality by constructing an explicit decomposition of the space. AniPQO extends the optimizer to enable plan-cost evaluation probes that return the estimated cost of a given plan at a given point in the parameter space. Commercial optimizers do not readily provide such functionality. Hulgeri and Sudarshan argue that such probes are cheaper than regular optimizer and extend a Volcano-based optimizer with which they are working to provide for such probes. In working with a commercial optimizer, it was found that plan-cost evaluation probes would be nearly as expensive as regular optimizer calls. Considering that the probes also produce cost estimates, and not actual costs, it was decided that this type of evaluation would not be used and instead the algorithms of the present invention were evaluated with respect to actual execution time. For this reason, a direct side-by-side comparison between AniPQO and the algorithms of the present invention cannot be presented.
AniPQO theoretically works with any number of query parameters, discrete and continuous, but is impractical for more than 4 parameters because the size of the training set (i.e., the number of optimizer calls) grows exponentially with the number of parameters. Either of the algorithms of the present invention leverages machine learning techniques to successfully learn the plan space with far fewer training points, and hence scales beyond 4 parameters. In fact, each of the experiments described herein use no more than 200 training points, compared to over 1000 points for many queries in [12]. This advantage comes from variance-reducing techniques inherent to RDTs and AdaBoost.
While geometric in nature, RDTs and AdaBoost are both less sensitive to the shape of the plan space compared to explicit geometric techniques like AniPQO. AniPQO was demonstrated to work on a class of SPJ queries with a “smooth” plan space (i.e., optimality regions are convex and there are no discontinuities in the plan space). The algorithms of the present invention work for a larger class of queries: SPJ and aggregate, with continuous and discrete attributes, on uniform and skewed data sets.
Because the cost of testing is small, the algorithms of the present invention can be used to optimize correlated subqueries. Queries containing correlated subqueries often cannot be rewritten into a single query block. In such cases, commercial optimizers optimize the outer query and the inner subquery as separate query blocks. From the point of view of the inner subquery, the outer references are parameters. The inner query may also have bind variables of global scope. The inner subquery will be called multiple times with various values for the parameters. Known commercial systems currently optimize a query once. While some systems abandon a query execution and reoptimize when the current execution looks bad [15], no conventional system routinely reoptimizes a correlated subquery depending on the values of the outer reference.
Machine learning methods can be used to accurately model predictions of a relational query optimizer. Based on these models, one can derive plans much more cheaply than the optimizer can. Further, the plans generated this way perform better than using a single pre-optimized plan for a query template.
An RDT-based method outperformed a method based on AdaBoost, although AdaBoost still outperformed the optimizer for most examples. It is common in machine learning that different classes of problems have different “best” learning algorithms. This concept is referred to as inductive bias in the machine learning literature [16].
Previous PQO methods considered only optimizer cost estimates and reasoned about the misprediction penalty in terms of those estimates. In the techniques disclosed herein, query execution time is measured and the actual misprediction penalty is derived. The present invention is the first to demonstrate that a PQO approach can result in savings beyond query optimization time, and achieve a significant overall net win over the methods currently available in relational database systems.
Since the optimizer provides just a model of the execution cost, sometimes the actual best plan is not the one chosen by the optimizer. A technique described above shows how to “correct” the optimizer's plan selection within the machine learning model, using actual performance results for a query template. A performance improvement of more than an order of magnitude can be achieved for some queries.
Local memory elements of memory 1904 are employed during actual execution of the program code of query execution plan determination system 1914. Cache memory elements of memory 1904 provide temporary storage of at least some program code in order to reduce the number of times code must be retrieved from bulk storage during execution. Further, memory 1904 may include other systems not shown in
Memory 1904 may comprise any known type of data storage and/or transmission media, including bulk storage, magnetic media, optical media, random access memory (RAM), read-only memory (ROM), a data cache, a data object, etc. Storage unit 1912 is, for example, a magnetic disk drive or an optical disk drive that stores data. Moreover, similar to CPU 1902, memory 1904 may reside at a single physical location, comprising one or more types of data storage, or be distributed across a plurality of physical systems in various forms. Further, memory 1904 can include data distributed across, for example, a LAN, WAN or storage area network (SAN) (not shown).
I/O interface 1906 comprises any system for exchanging information to or from an external source. I/O devices 1910 comprise any known type of external device, including a display monitor, keyboard, mouse, printer, speakers, handheld device, printer, facsimile, etc. Bus 1908 provides a communication link between each of the components in computing unit 1900, and may comprise any type of transmission link, including electrical, optical, wireless, etc.
I/O interface 1906 also allows computing unit 1900 to store and retrieve information (e.g., program instructions or data) from an auxiliary storage device (e.g., storage unit 1912). The auxiliary storage device may be a non-volatile storage device (e.g., a CD-ROM drive which receives a CD-ROM disk). Computing unit 1900 can store and retrieve information from other auxiliary storage devices (not shown), which can include a direct access storage device (DASD) (e.g., hard disk or floppy diskette), a magneto-optical disk drive, a tape drive, or a wireless communication device.
The invention can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment containing both hardware and software elements. In a preferred embodiment, the invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc.
Furthermore, the invention can take the form of a computer program product accessible from a computer-usable or computer-readable medium providing program code of query execution plan determination system 1914 for use by or in connection with a computing unit 1900 or any instruction execution system to provide and facilitate the capabilities of the present invention. For the purposes of this description, a computer-usable or computer-readable medium can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
The medium can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium. Examples of a computer-readable medium include a semiconductor or solid state memory, magnetic tape, a removable computer diskette, RAM 1904, ROM, a rigid magnetic disk and an optical disk. Current examples of optical disks include compact disk-read-only memory (CD-ROM), compact disk-read/write (CD-R/W) and DVD.
The flow diagrams depicted herein are provided by way of example. There may be variations to these diagrams or the steps (or operations) described herein without departing from the spirit of the invention. For instance, in certain cases, the steps may be performed in differing order, or steps may be added, deleted or modified. All of these variations are considered a part of the present invention as recited in the appended claims.
While embodiments of the present invention have been described herein for purposes of illustration, many modifications and changes will become apparent to those skilled in the art. Accordingly, the appended claims are intended to encompass all such modifications and changes as fall within the true spirit and scope of this invention.
This application is a continuation application claiming priority to Ser. No. 11/673,091, filed Feb. 9, 2007.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 11673091 | Feb 2007 | US |
| Child | 12125221 | US |
