The present invention relates generally to computer processing, and more particularly to techniques for detecting and processing cache hits for queries with aggregates.
Decision support systems are commonly available to assist businesses make proper decisions based on data that may be available from a myriad of different data sources. For example, analytics applications are available to simplify data analysis for specific business or functional areas, such as sales and marketing management, customer relationship management (CRM), financial management, inventory management, category management, and so on. Decision support systems collect and transform data from these various data sources into more meaningful and understandable information to end-users.
A major goal of decision support systems is to provide end-users with results to business questions quickly and accurately. Business questions may be translated to “queries” by a query server, and the queries are then “run” against databases. Depending on the complexity of the queries, a long time may be required by the servers for these databases to return results to the queries. In general, it is desirable to be able to return the results to the end-users as fast as possible.
The processing of queries consumes large amounts of resources at the query and database servers and the computer network that interconnects these servers. The query server normally parses each query received from an end-user into “physical” queries against each database where data for the final results may be stored. The query server would then receive intermediate results from all queried databases and perform any required post-processing to obtain the final results.
Query caching may be used to improve performance by avoiding unnecessary processing. With query caching, the results of prior queries are stored in a local cache, which may be implemented on a storage unit (e.g., a memory unit or a disk drive) that can be more quickly and easily accessed by the query server. If a new query is subsequently received and can use the results stored in the cache, then the database processing for the new query can be avoided. This may then result in a dramatic improvement in the average response time for queries. Besides faster response time, the ability to answer a new query from a local cache conserves resources at (1) the query server, since post processing on the intermediate results received from queried databases may be avoided, (2) the database server, which may then be able to perform other processing, and (3) the network, since intermediate results do not have to be sent over the network to the query server.
The effectiveness of query caching is dependent on the ability to determine whether or not the results for queries that have been saved in the cache (i.e., cached queries) may be used to obtain the results for a new query. In the simple case in which a new query is exactly like one of the cached queries, the results for this cached query may be quickly retrieved and presented to the end-user. However, in many instances, the new query will not be exactly like any of the cached queries.
A major challenge with query caching is then the ability to determine whether or not results for a new query may be obtained from results for a cached query. This will be possible if the new query is a subset of and “subsumed” by the cached query. A cache hit may be declared if the new query can be answered by the cached query, and a cache miss may be declared otherwise.
In one conventional scheme for determining whether or not a new query can be answered by a cached query, the new query is evaluated against each cached query. This scheme provides acceptable performance when the number of cached queries is small (e.g., tens of cached queries), but would be ineffective when the number of cached queries is large (e.g., hundreds or thousands of cached queries).
The determination of whether or not a new query is subsumed by a cached query is also made more challenging by the fact that queries often ask for “aggregated” data, which is data obtained by operating on multiple data entries. For example, a query may ask for the sum of all revenues for each year, which may then be obtained by summing the revenues for all transactions of each year. In this example, “transactions” would be the data entries and “sum” is the aggregation being performed on the data entries. In the process of aggregating data, the value of each individual data entry is lost (i.e., the value of each transaction is not available, in the above example).
Conventionally, for a new query with one or more aggregates to be considered subsumed by a cached query, the aggregates of the new query are required to exactly match the aggregates of the cached query or fall within a very limited range of exceptions. This strict aggregate matching criterion can often result in declaration of a cache miss when in fact the results for the new query may be obtained from the results from the cached query. The ability to detect cache hits may be even more important for queries with aggregates because additional savings may be achieved from not having to perform the aggregation.
As can be seen, techniques that can be used to improve query caching in the presence of aggregates are highly desirable.
Techniques are provided herein to efficiently determine whether or not a new query may be answered by a query stored in a cache. These techniques improve query caching performance by efficiently selecting cached queries for evaluation and increasing the cache hit rate by allowing for inexact matches.
An embodiment of the invention provides a computer program product operative to support query caching. This product includes codes for determining a list of candidate queries stored in a cache that potentially could be used to answer a new query and codes for evaluating each of at least one candidate query in the list to determine whether or not there is a cache hit for the candidate query. The cache hit would indicate that the candidate query could be used to answer the new query. The evaluation is performed using a set of rules that allows for inexact matches of aggregates, if any, included in the new query. A query plan may thereafter be generated for the new query based on a specific candidate query with a cache hit
The list of candidate queries to evaluate may include (1) all queries stored in the cache, (2) cached queries containing exact matches for items in a select list of the new query, or (3) cached queries containing exact and/or inexact matches for the select list items of the new query. Exact matches of select list items may be determined using a select list index, and inexact matches may be determined for aggregates using “aggregate rewrite”, which may be performed using an aggregate index and the select list index, as described in further detail below. Only certain types of aggregates may be rewritten and, in an embodiment, a rewrite may be performed for an aggregate only if an exact match is not found for that aggregate.
The list of candidate queries to evaluate may thus be determined by (1) for each select list item of the new query with an exact match, obtaining a set of cached queries that include that select list item, (2) for each select list item of the new query without an exact match, obtaining a set of cached queries for that select list item using aggregate rewrite, and (3) forming the list of candidate queries by intersecting the sets of cached queries obtained for all select list items of the new query. Step (2) above may be performed by (1) obtaining a set of aggregates included in the cached queries which could be used to rewrite an aggregate in the new query without an exact match, (2) for each aggregate in this set, obtaining a set of cached queries that include that aggregate, and (3) performing a union of the sets of cached queries obtained for all aggregates in the set. Details for this are described below.
The invention further provides other methods, computer program products, and systems capable of implementing various aspects, embodiments, and features of the invention, as described in further detail below.
The foregoing, together with other aspects of this invention, will become more apparent when referring to the following specification, claims, and accompanying drawings.
Depending on its design, Web/application server 120 may support various software applications (e.g., analytics), with only one application 122 being shown in
Query server 130 receives logical query requests (or queries) from Web/application server 120 , processes each query, and provides the results for the queries back to the Web/application server. Query server 130 performs all of the query processing, including sending SQL requests to the databases. In the embodiment shown in
Databases 150 store data that may be accessed by server 130 as needed. A metadata store 160 stores information about the application environment, data modeling, aggregate navigation, caching, and so on. This information is used by server 130 to perform various processing. A cache 170 provides storage for results for prior queries and is implemented with a storage unit (e.g., a disk drive or a memory unit) that can be accessed more readily and quickly than databases 150.
The processing for a new query may proceed as follows. The query is first provided to SQL generation engine 134, which translates the logical request for this query into an SQL statement. Query processing engine 140 then determines whether or not this query may be answered using a query that is stored in cache 170. If the answer is yes, then the query request is run against the cached query. Otherwise, query processing engine 140 translates the SQL statement into physical queries against one or more databases 150. In either case, computation engine 136 performs post processing of the intermediate results from the databases or the cache to provide the desired results.
A logical request may be issued by an end-user or some other entity to query objects in one or more databases to obtain specific results of interest. The logical request may be formulated using a SELECT statement, which is the basis for querying a structured query language (SQL) database. The SELECT statement for a query, which is also referred to as a query specification, has the following basic syntax:
where
Queries are normally run against one or more databases, each of which may include one or more tables. Each table is uniquely identified by a table name and typically includes a number of columns that contain the data for that table. Each column of a table is also uniquely identified by a column name. However, columns of different tables may have the same name, in which case they may be uniquely identified by the combination of table and column names (e.g., TableX.ColumnY).
A query may ask for data in “raw” form, such as the data in a column of a table. A query may also ask for data to be “aggregated” in some defined manner. For example, the query may ask for the sum of all revenues for each product. The level of aggregation may be specified in the query (e.g., sum of revenues for each product, for each product and in each year, for each year, and so on). Aggregates are described in further detail below.
As shown above, the SELECT statement for a query includes a number of clauses, some of which are optional. The FROM clause includes a list of tables from which the results of the query may be obtained.
The SELECT clause includes a list of one or more items to be provided as the results of the query. Each select list item may be (1) a column in a table identified by the FROM clause (e.g., “revenue”), (2) an aggregate, which is typically a particular operation or function to be performed on a specified column (e.g., “sum(revenue)”), or (3) an expression, which is an operation to be performed on each entry in one or more specified columns (e.g., “revenue+5”). One column is provided in a results table for the query for each item in the select list. Thus, the SELECT clause defines the format of the results table for the query.
The WHERE clause includes one or more conditions or restrictions used to limit the query results. Thus, only data that satisfies the specified conditions is included in the results table. The conditions in the WHERE clause may be viewed as filters for the data obtained by the SELECT clause.
The GROUP BY clause includes one or more columns by which aggregates, if any, in the SELECT clause are computed. The aggregation is performed such that one aggregated value is provided for each unique combination of the elements in the column(s) specified in the GROUP BY clause. For example, if the SELECT clause includes an aggregate “sum(revenue)” and the GROUP BY clause includes “product”, then the sum of revenue is computed for each unique product. The level of aggregation is thus specified by the column(s) in the GROUP BY clause.
The HAVING clause includes one or more conditions or restrictions for aggregates. The conditions in the HAVING clause may be viewed as filters for the data obtained by aggregates in the SELECT clause.
The ORDER BY clause is used to arrange the order of the columns in the output table.
As an example, a SELECT statement for a first query may be written as follows:
The results of this query may be as follows:
This query includes one aggregate: “sum(revenue)”, and the GROUP BY clause includes one column: “year”. Thus, the values in the revenue column are summed for all products for each year.
In one specific SQL instantiation, if the GROUP BY clause is not included in the SELECT statement, then the level of aggregation is determined by the non-aggregate columns in the SELECT clause. For example, if the SELECT statement for the first query did not include the GROUP BY clause, then the level of aggregation would be product and year (i.e., the aggregation is performed for each product of each year).
The SELECT statement for a second query may be written as follows:
The results of this query may be as follows:
This query is the same as the first query and includes one aggregate: “sum(revenue)”. However, the GROUP BY clause includes two columns: “product” and “year”. Thus, the values in the revenue column are summed for each product of each year.
The SELECT statement for a third query may be written as follows:
This query is similar to the second query, except that only the results for the product/year combinations with the sum of revenue exceeding 800 are provided in the results table.
Different instantiations of SQL may support different features not present in the standard SQL. One specific “extended” SQL instantiation is described below having various capabilities not supported by standard SQL. This extended SQL has the ability to perform “functional” aggregates such as “rank”, “median”, “topn”, “bottomn”, and “running/moving aggregates”. This extended SQL may further have the ability to, for “regular” aggregates such as “sum”, output aggregates with multiple levels of aggregation in a single row of the results table. For example, the extended SQL would allow for the following items in the select list: “country”, “region”, “sum(sales by country, region)”, and “sum(sales by country)”. Standard SQL does not have the ability to support the previous select list. This extended SQL also allows users to place conditions in the WHERE and HAVING clauses interchangeably. The query compiler would then determine the appropriate place for each clause. In much of the following description, the extended SQL is assumed to be used and all conditions are deemed to be placed in the WHERE clause.
Extended SQL also supports another syntax in addition to the one used for standard SQL. For example, the SELECT statement for the first query described above may also be written as:
Referring back to
The above examples show the use of aggregates in queries. One fast way to process a query is to skip the bulk of the processing and use precomputed results. Aggregate tables are examples of such and contain precomputed results for a particular aggregation level. For example, an aggregate table may store sales results for each product by month, when the granularity of detail for the database may be at the day level. To create this aggregate table, a process (often a query) computes the results and then stores them in a table in the database.
As noted above, the processing of queries may require a long time and further consume large amounts of resources. Query caching may be used to store the results of prior queries in a cache so that these results may thereafter be used to obtain results for new queries. If a new query can be answered using the stored results (a cache hit), then the cost of evaluating this query is considerably cheaper than having to pose the query anew against one or more databases (a cache miss). The throughput and performance of the system would be determined by the cache hit rate.
A query caching system typically includes a number of components, two of which are cache hit detection and query plan generation. The cache hit detection component determines whether or not a given new query can be answered by any one of the cached queries. If there is a cache hit, then the query plan generation component generates a query plan for the new query in terms of a selected cached query with the cache hit. The query plan will then be processed to obtain the results for the new query based on the results for the selected cached query.
A major design tradeoff for a query caching system is the cost incurred for processing each new query to detect a cache hit versus the benefits of reduced response time and resources usage whenever a cached query is used to answer the new query. On one extreme, a cache hit detection scheme may require an exact match between a new query and a cached query to be declared a cache hit. An exact match would occur if the elements in each clause of the SELECT statements for the two queries are identical, including the WHERE clause conditions. To determine exact match, the expression tree for the SELECT statement for each cached query may be viewed as a key. An index structure populated with the keys for all cached queries may then be compared with the key for the new query to determine whether or not there is an exact match. Exact match is relatively simple to evaluate but is also the most restrictive, and cache misses may be erroneously declared in instances where cached queries may actually be used to answer the new query.
A cache hit detection scheme that can identify cache hits based on an inexact match between the new query and cached queries requires more processing but may be able to provide improved performance by increasing the cache hit rate. An inexact match cache hit may be determined by “evaluating” the new query against a cached query based on the semantics of queries. For a given query, the semantics specify both the operations to be performed for that query and the order for performing these operations to obtain the results. After evaluating the new query against the cached query, it may be possible to ascertain whether the results for the new query could be obtained from the results for the cached query.
The computational semantics for a query may be expressed as follows:
Whether or not the new query can be answered by a given cached query may be ascertained by performing steps 1 through 7 described above for the new query and the cached query. As illustrated above, the processing to determine inexact match may be complicated.
The cache may be designed to store a large number of queries. In that case, another important design consideration is the selection of queries stored in the cache to evaluate against the new query to detect for cache hits. If the processing to evaluate a cached query for cache hit is simple, as is the case for exact match, then the new query may be evaluated against each cached query, one at a time, until a cache hit is declared. This is often referred to as the “naïve” strategy. However, if the processing to evaluate a cached query for cache hit is more complicated, as may be the case for inexact match, and if the number of queries stored in the cache is large, then the costs of traversing through all cached queries to search for a cache hit may be excessive.
Techniques are provided herein to efficiently determine whether or not a new query may be answered by a cached query. These techniques improve query caching performance by (1) selecting cached queries for evaluation in an efficient manner, and (2) increasing the cache hit rate by allowing for inexact matches. These techniques may be advantageously used in query caching systems that store a large number of queries in the cache.
Initially, a new query is received for processing (step 208). A first list of “candidate” cached queries that potentially may be used to answer the new query is then determined based on items in the select list (step 212). If the cache stores a large number of cached queries, then a large amount of resources may be required to evaluate each cached query to determine whether or not there is a cache hit. Techniques are provided herein to efficiently “compare” the new query against the cached queries to trim down the number of candidate cached queries for evaluation. The first list may be determined as described in further detail below.
If the first list is not empty, as determined in step 214, then the new query is evaluated against each candidate cached query in the first list, one at a time, to determine whether or not the new query can be answered by the cached query (step 216). This evaluation may be performed as described in further detail below. If there is a cache hit for any cached query, as determined in step 218, then a cache hit is declared (step 230) and a query plan is generated for the new query based on the cached query with the cache hit (step 232). Multiple cache hits may be possible for the candidate cached queries in the first list. However, the evaluation in step 216 may stop after the first cache hit has been detected. The process then terminates.
If the first list is empty, as determined in step 214, or if no cache hit is achieved for any of the candidate cached queries in the first list, as determined in step 218, then a second list of candidate cached queries is determined by “rewriting” certain aggregates, if any, included in the new query (step 222). The new query may include one or more aggregates that may not be exactly like the ones in the cached queries. Techniques are provided herein to efficiently determine whether or not each aggregate in the new query may be answered by aggregates in the cached queries even if these aggregates are not exactly alike, as described in further detail below.
If the second list is not empty, as determined in step 224, then the new query is evaluated against each candidate cached query in the second list, one at a time, to determine whether or not the new query can be answered by the cached query (step 226). If there is a cache hit for any of the cached queries, as determined in step 228, then a cache hit is declared (step 230) and a query plan is generated for the new query based on a selected cached query with the cache hit (step 232). Otherwise, if the second list is empty, as determined in step 224, or if no cache hit is achieved for any of the candidate cached queries in the second list, as determined in step 228, then a cache miss is declared (step 240). In this case, the new query is run against the databases (step 242) and the new query and its results are stored in the cache for possible use by future queries (step 244). After steps 232 and 244, the process terminates.
As shown in
As also shown in
Numerous modifications may be made to the process shown in
In an aspect, the first list of candidate cached queries to evaluate for cache hit is determined based on items in the select list. As shown above, the SELECT clause for each query includes a list of one or more items, where each item may be column, an aggregate, or an expression. A cached query may be included in the first candidate list if each item in the select list of the new query has an exact match in the select list of the cached query.
In the embodiment shown in
For each entry in the select list index, field 312 stores a specific select list item and field 314 stores a pointer to each results table whereby the associated query includes this select list item. For example, since “employeeid” is included in the SELECT clause of cached queries 1, 4, and 5, the data item for “employeeid” includes pointers to the results tables for queries 1, 4, and 5. Thus, each unique select list item is associated with a data item that is the set of cached queries containing that particular item.
In one embodiment, whole expressions are stored in the select list index. For this embodiment, an expression such as “productid+5” would be stored in its entirety in the table.
In another embodiment, subexpressions are stored in the select list index. Each expression may be decomposed into subexpressions, and each subexpression may then be stored in the select list index. For the above example, the expression “productid+5” can be decomposed into “productid” and 5, and “productid” may be stored in the table. “Productid” would then be deemed an inexact match for expressions such as “productid+5” and “productid*2” since these expressions may be computed from “productid” by the addition of 5 and the multiplication of 3, respectively.
If an expression includes multiple items stored in the select list index, the set of cached queries for this expression would be the intersection of the sets of cached queries for all items. For example, the set of cached queries for the expression “productid+siteid” would be the intersection of the set of cached queries containing “productid” and the set of cached queries containing “siteid”, since both items are needed to compute “productid+siteid”.
The select list index is updated as necessary to ensure that its contents remain accurate. In particular, whenever an existing query is deleted from the cache, the select list index is updated so that each select list item that references the deleted query is revised or removed from the table. Similarly, whenever a new query is stored in the cache, each select list item included in the new cached query is revised or added to the index.
The select list index in
Initially, the first item in the select list of the new query is identified (step 412). The set of cached queries containing this select list item is then obtained by looking up the select list index (step 414). This can be achieved by looking up the key corresponding to this select list item and retrieving the data item associated with this key. A candidate set is then initialized to include all cached queries contained in the set just obtained from the select list index (step 416).
A determination is then made whether or not the candidate set is empty (step 418). If the answer is yes, then the process proceeds to step 428. This is because any additional select list item can only reduce the candidate set but never enlarge it, and an empty candidate set at any stage would indicate that there are no possible candidate cached queries for the new query.
If the candidate set is not empty, as determined in step 418, then a determination is made whether or not all select list items for the new query have been considered (step 420). If the answer is no, then the next item in the select list of the new query is identified (step 422), and the set of cached queries containing this select list item is obtained by looking up the select list index (step 424). The set of all cached queries that contains all select list items considered thus far may then be determined by intersecting the set just obtained from the select list index with the candidate set (step 426). The new candidate set would include only those cached queries, if any, that are present in both the prior candidate set and the set just obtained. The process then returns to step 418.
If the candidate set is empty, as determined in step 418, or if all select list items for the new query have been considered, as determined in step 420, then the cached queries included in the candidate set are provided as the first list of candidate cached queries to be evaluated (step 428). This first list may be a null list if no cached query contains all of the select list items included in the new query. The process then terminates.
The process shown in
In the above description, the first list of candidate cached queries is determined based on exact match of select list items. Items for other clauses in the SQL statement may also be stored in an index, in conjunction with or as substitute for the select list items, and this is within the scope of the invention. For example, if a sufficient number of tables or data sources are accessed, then these may also be stored in the index. For example, one entry may be provided in the index for each unique item in the FROM clause. The items in the FROM clause of the new query may then be processed in similar manner as that described above for the select list items.
Queries often include aggregates. Each aggregate typically denotes a particular operation (e.g., sum) to be performed on a particular data column with a particular level of aggregation, all of which are specified in the SELECT statement for the query.
The set of all aggregates supported by a particular SQL instantiation may be classified into a number of different categories based on the properties of the resultant data generated by the aggregate operation or function. These categories include “regular” aggregates, “distributive” aggregates, and “functional” aggregates. Distributive aggregates include operations such as SUM, COUNT, MAX, and MIN. Regular aggregates are a superset of distributive aggregates (e.g., average is a regular aggregate that is not a distributive aggregate). Functional aggregates include operations such as RANK, PERCENTILE, TOPN, BOTTOMN, and MAVG. Of these categories, distributive aggregates are of special interest because the results of a distributive aggregate operation may be subjected to the same operation with a coarser level of aggregation and still produce the correct results.
For example, a cached query may contain the aggregate “sum(revenue by productid, year)”, which sums revenue for each product and in each year. A new query may then ask for the aggregate “sum(revenue by year)”, which sums revenue for each year. In this case, the results for the new query can be obtained from the results for the cached query by performing an additional aggregation over all products for each year. Similarly, if another query asks for the aggregate “sum(revenue by productid)”, which sums revenue for each product, then the results for this query can be obtained from the results for the cached query by performing an additional aggregation over all years for each product. The aggregate “sum(revenue by year)” and “sum(revenue by productid)” in the two new queries may be rewritten in terms of the aggregate “sum(revenue by productid, year)” in the cached query.
The process of aggregating over a coarser level of aggregation is also referred to as “aggregate roll-up”. Aggregate roll-up can be validly applied for certain aggregates, such as distributive aggregates, to produce the correct results. Aggregate roll-up would produce incorrect results if applied to some other aggregates. For example, the aggregate “avg(revenue by productid)” could not be obtained from the aggregate “avg(revenue by productid, year)” unless additional information regarding revenue is available.
Since distributive aggregates can be rolled up, they are good candidates for being rewritten if they appear in the new query. In one embodiment, using a naïve strategy, each distributive aggregate in the new query is checked against all aggregates in all cached queries for a possible rewrite. This embodiment can provide satisfactory performance when the number of cached queries is not too large. In another embodiment, the search for aggregates in the cached queries for a possible rewrite of a distributive aggregate in the new query is made using an aggregate index.
In the embodiment shown in
The aggregate index is updated as necessary to ensure that its contents remain accurate. In particular, whenever an existing query is deleted from the cache or a new query is stored to the cache, the aggregate index is updated so that aggregates, if any, included in the deleted or added query are deleted from or added to the index.
In the first step of the process, the set of cached queries for each “matched” select list item in the new query is obtained using the select list index (step 602). A matched select list item is one that is found in the select list index, and an unmatched select list item is one that is not found in the index.
The set of aggregates is then obtained for each unmatched select list item in the new query using the aggregate index (step 604). The aggregate set for each unmatched select list item is further operated on to obtain the set of cached queries for the select list item. In particular, for each aggregate set, the set of cached queries for each aggregate in the set is first obtained using the select list index, and a union is then performed on the sets of cached queries for all aggregates in the set to obtain the set of cached queries for the unmatched select list item (step 606).
A set of cached queries is thus obtained for each matched select list item in step 602 and a set of cached queries is obtained for each unmatched select list item in steps 604 and 606. An intersection is then performed on all of these sets of cached queries to provide the second list of candidate cached queries for the new query (step 608).
Aggregate rewrite can increase the search space of candidate cached queries. In an embodiment, to prevent a possible explosion in the size of the search space, a rewrite is attempted for an aggregate (e.g., “aggregate(qtysold by sited)”) only if an exact match for the aggregate cannot be found in the select list index. Moreover, the increase in the search space may be minimized by choosing a specific order in which to process the items in the select list of the new query. Since only aggregates are candidates for rewrites and since rewrites can increase search space, they may be processed last.
For process 222b, a candidate set is first initialized to include all queries stored in the cache (step 610). The first item in the select list of the new query is then identified (step 612). A determination is then made whether or not this select list item is found in the select list index (i.e., whether or not this is a match or unmatched select list item) (step 614). If the select list item is found in the select list index, then the set of cached queries containing this select list item is obtained by looking up the select list index (step 616).
Otherwise, if the select list item is not found in the select list index, then a determination is made whether or not the item is an aggregate (step 622). If the answer is no, then the select list item is deemed to not be included in any of the cached queries, and the process proceeds to step 644 where a null second list is provided for the new query. Otherwise, if the select list item is an aggregate, then the item being aggregated upon is determined (step 624). The set of aggregates in the cached queries that contain or aggregated over this item is then obtained by looking up the aggregate index (step 626).
A temporary set is next initialized to null (step 628). The first aggregate in the aggregate set is then identified for consideration (step 630) and this aggregate is evaluated to see whether or not it can be used for a rewrite of the select list aggregate in the new query currently being processed (step 632). A rewrite would be possible if the select list aggregate merely roll ups the results of the aggregate under consideration. As part of the rewrite evaluation, a check is performed to ensure that all columns identified in the select list aggregate are included in the select list of the cached query. For example, if “aggregate(qtysold by siteid)” is to be rewritten using “aggregate(qtysold by siteid, productid)”, then the item “productid” needs to be present in the select list of the cached query.
If a rewrite is not possible, as determined in step 634, then the process proceeds to step 640. Otherwise, if a rewrite is possible, then the set of cached queries containing the aggregate under consideration is obtained by looking up the select list index (step 636), and a union is performed between this set of cached queries and the temporary set. The union is performed because any of the aggregates that can be used for a rewrite may thereafter be selected for the select list aggregate.
A determination is then made whether or not all aggregates in the aggregate set have been considered (step 640). If the answer is no, then the next aggregate in the set is identified (step 642), and the process returns to step 632 to evaluate this aggregate. Otherwise, if all aggregates in the set have been considered, then the process proceeds to step 652.
A set of cached queries is obtained for each match select list item in step 616 and a set of cached queries (which is the temporary set) is obtained for each unmatched select list item in steps 622 through 642. An intersection is then performed between the set of cached queries obtained for the select list item currently being processed and the candidate set (step 652).
If the candidate set is not empty (as determined in step 654) and if all select list items for the new query have not been considered (as determined in step 656), then the next item in the select list of the new query is identified (step 658), and the process returns to step 614 to process this select list item. Otherwise, if the candidate set is empty or if all select list items for the new query have been considered, then the cached queries included in the candidate set are provided as the second list of candidate cached queries for the new query (step 660). This second list may be a null list if no cached query contains all of the select list items (or rewrite candidates) included in the new query. The process then terminates.
In an embodiment, aggregate rewrites may also be performed on subexpressions. For example, consider the expression “sum(revenue by country)*2”. This expression may be decomposed into “sum(revenue by country)” and 2. The subexpression “sum(revenue by country)” may then be written in terms of the aggregate “sum(revenue by country, year)”.
The operation of processes shown in
Using the example aggregate index shown in
The select list index is then accessed to obtain (1) the set of cached queries containing the first aggregate in the set “sum(revenue by siteid, productid)”, which may be, for example, {1, 2, 7}, and (2) the set of cached queries containing the second aggregate in the set “sum(revenue by siteid, month)”, which may be, for example, {3, 9}. The union of these two sets would be {1, 2, 3, 7, 9}, which is the set of cached queries that support the second select list item “sum(revenue by siteid)” in the new query.
The set obtained for the first select list item {2, 3} is then intersected with the set obtained for the second select list item {1, 2, 3, 7, 9} to provide the second list of candidate cached queries {2, 3} that supports all select list items of the new query.
The aggregate index shown in
The candidate selection process described above ensures compatibility between the select list of the new query and the select list of each candidate cached query. However, each query may specify one or more filters via the WHERE and HAVING clauses. These filters limit the results provided for the query based on the conditions or restrictions specified by the filters. Thus, each candidate cached query would need to be evaluated to ensure that the filters, if any, specified by the cached query did not filter out any data needed by the new query. This evaluation is then to verify that the filters present in the new and cached queries are compatible.
For standard SQL, the WHERE clauses cannot contain aggregates. In some other SQL instantiations, such as the extended SQL described above, aggregates may be allowed in the WHERE and HAVING clauses, in which case the evaluation would need to account for any aggregates that may appear in these clauses. The server can be designed to automatically determine where each predicate should be placed. In the following analysis, an assumption is made that all filters are present in the WHERE clause prior to being appropriately placed.
Filters are often defined using “WHERE clause conjuncts”, which can be described as follows. Consider an SQL query:
This query asks for the lastname and firstname of all male employees with age greater than 50. In this query, “gender=“Male”” and “age>50” are WHERE clause conjuncts. They are referred to as conjuncts because both conditions have to be satisfied for an employee's name to satisfy this WHERE clause condition. This is because the two conditions are joined with the AND operator. If these conditions had been joined with the OR operator, then they would be referred to as “WHERE clause disjuncts”. Any logical expression can be normalized so that it consists of a sequence of conjuncts. The following description assumes that logical expressions are normalized before they are processed.
Initially, a determination is made whether or not the WHERE clause for the new query exactly matches that of the target query (step 710). If the answer is yes, which is the simplest case, then there is a cache hit (step 714).
Otherwise, if the WHERE clauses do not exactly match, then a determination is made whether or not all non-matching WHERE clause conjuncts reference aggregates (step 720). A WHERE clause conjunct is said to reference an aggregate if the aggregate is used in the WHERE clause conjunct (e.g., “sum(revenue by productid)>1000”). Step 720 effectively determines if there is a HAVING clause constraint. If the answer is yes, then a determination is made whether or not the WHERE clause contains rewrite AND the matching filter has aggregate (step 722). There is a cache miss if the answer is yes (step 724) and a cache hit if the answer is no (step 726).
If the answer to step 720 is no, then a determination is made whether or not all non-matching WHERE clause conjuncts reference non-aggregates (step 730). Step 730 effectively determines if there is a WHERE clause constraint. If the answer is yes, then a determination is made whether the select list or the WHERE clause contains any aggregate (step 732). There is a cache miss if the answer is yes (step 734) and a cache hit if the answer is no (step 736).
If the answer to step 730 is no, then a determination is made whether the select list or the WHERE clause contains any aggregate (step 742). There is a cache miss if the answer is yes (step 744) and a cache hit if the answer is no (step 746). The process then terminates.
The algorithm shown in
For clarity, specific examples are provided in Appendix A to illustrate the criteria described above in
Initially, a determination is made whether or not the WHERE clause for the new query exactly matches that of the target query (step 810). If the answer is yes, then a determination is made whether or not the WHERE clause contains a functional aggregate (step 820). If the answer is also yes, then the following three criteria apply. First, there is a cache hit if all inexact matches in the select list reference non-aggregates (steps 822 and 824). Second, there is a cache miss if one or more inexact matches in the select list reference a functional aggregate where this aggregate is “above” the hit (i.e., the functional aggregate is computed from the cached rows) (steps 826 and 828). Third, there is a cache hit if the select list does not contain a rewrite (steps 832 and 834) and a cache miss otherwise.
If the answer to step 820 is no, then a determination is made whether or not the WHERE clause contains no aggregates (step 840). There is a cache hit if the answer is yes (step 842). If the answer to step 840 is no, then a determination is made whether or not the WHERE clause contains regular aggregates (step 850). If the answer is yes, then there is a cache hit if the select list does not contain an aggregate rewrite (steps 852 and 854) and a cache miss otherwise (step 856).
In
In
If the answer to step 880 is no, then a determination is made whether or not the non-matching WHERE clause conjuncts contain a mix of aggregates and non-aggregates (step 890). If the answer is yes, then the same two criteria for step 880 also apply, as shown in
The algorithm shown in
For clarity, specific examples are provided in Appendix A to illustrate the criteria described above in
Given a new query (Q) and a set of cached queries (Q1, Q2 . . . QN), a determination needs to be made as to which specific cached query to use to generate the results for the new query. One or more cached queries may be hits and would be possibilities. Different amounts of processing may be required to be performed on the results for different cached queries to obtain the results for the new query. In one embodiment, the first candidate cached query that hits is selected for use to generate the results for the new query. In a second embodiment, some or all cached queries that hit may be further evaluated to determine the best one to use. The first embodiment may be used if the benefit from using the best cached query is outweighed by the cost of determining the best one.
Once a specific cached query has been selected for use to generate the results for the new query, the query plan generation phase is entered (step 232 in
In the query plan generation, if an expression in the select list of the new query contains an aggregate and there is an exact match for this expression in the select list index, then the expression is treated as a column index and not as an aggregate by the query compiler. For example, the new query may contain the expression “rank(qtysold)+5” in the select list. If there is an exact match for this expression in the select list index, then the expression is rendered in the new plan as a column index of “rank(qtysold)+5” in the target query's select list (which may be 1, 2, 3, and or on).
A major issue in query plan generation is to ensure that the correct order of filter execution is preserved. In an embodiment, filter execution order is preserved by inserting additional levels of nesting and moving filters where appropriate. In one implementation, rewrite rules for an expanded summary filter examine predicates in the detail filter and move them to the summary filter if appropriate. A summary filter is one normally included in the HAVING clause in standard SQL, and a detail filter is one normally included in the WHERE clause in standard SQL. The rewrite rules also replace summary filter lists by converting them to detail filters in a newly added layer of nesting. In general, the query plan produced by the plan generation should be equivalent to a query plan produced without query caching.
A number of examples are provided in Appendix A to illustrate the evaluation performed in
Memory subsystem 912 may include a RAM 932 and a ROM 934 used to store codes and data that implement various aspects of the invention. For example, memory subsystem 912 may be used for metadata store 160 and possibly cache 170 in
Input device interface 916 provides interface with various input devices such as a keyboard 952, a pointing device 954 (e.g., a mouse, a trackball, a touch pad, a graphics tablet, a scanner, or a touch screen), and other input device(s) 956. Output device interface 918 provides an interface with various output devices such as a display 962 (e.g., a CRT or an LCD) and other output device(s) 964. Network interface 920 provides an interface for system 900 to communicate with other computers coupled to network 112.
Many other devices or subsystems (not shown) may also be coupled to system 900. In addition, it is not necessary for all of the devices shown in
Headings are included herein for reference and to aid in locating certain sections. These headings are not intended to limit the scope of the concepts described therein under, and these concepts may have applicability in other sections throughout the entire specification.
The foregoing description of the specific embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without the use of the inventive faculty. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein, and as defined by the following claims.
WHERE rank(employeeid) in (2,5).
WHERE rank(employeeid) in (2, 5)
WHERE rank(qtysold)<10.
WHERE rank(qtysold)<10 AND aggregate(qtysold BY siteid)>1000
The aggregate in the WHERE clause is a rewrite.
FROM sales
WHERE rank(qtysold)<10.
FROM sales
WHERE rank(qtysold)<10 AND aggregate(qtysold BY siteid)>1000
WHERE rank(employeeid) in (2, 5)
WHERE rank(qtysold)<5.
WHERE rank(qtysold)<5 and rank(revenue)<4.
WHERE rank(revenue)<5.
WHERE qtysold<1000.0 and rank(revenue)<5
WHERE employeeid in (2, 3, 5)
FROM sales
WHERE rank(qtysold)<10
FROM sales
WHERE rank(qtysold)<10 AND productid<5
WHERE rank(qtysold)<10
WHERE rank(qtysold)<10 AND productid<5
WHERE lastname =‘King’
WHERE lastname =‘Smith’
WHERE rank(qtysold)<10
FROM employee
WHERE rank(qtysold)<10
WHERE rank(qtysold)<10
WHERE rank(qtysold)<10
WHERE rank(qtysold)<5
WHERE rank(qtysold)<5
WHERE rank(revenue)<5
WHERE rank(revenue)<5
WHERE rank(qtysold)<10
FROM sales
WHERE rank(qtysold)<10
WHERE rank(qtysold)<10
WHERE rank(qtysold)<10
WHERE employeeid<5
WHERE employeeid<5
FROM sales
WHERE aggregate(qtysold BY siteid)>1000
WHERE aggregate(qtysold BY siteid)>1000
WHERE aggregate(qtysold BY siteid)>1000
WHERE aggregate(qtysold BY siteid)>1000
WHERE rank(revenue)<5
WHERE rank(revenue)<3
WHERE rank(revenue)<5
WHERE rank(revenue)<3
WHERE rank(qtysold)<10
FROM sales
WHERE rank(qtysold)<10 AND aggregate(qtysold BY siteid)>1000
WHERE productid<10
FROM sales
WHERE productid<10 AND rank(qtysold)<20
WHERE rank(revenue)<5
WHERE rank(revenue) in (1,3)
WHERE employeeid<5 and rank(revenue)<6
WHERE employeeid in (2,4) and rank(revenue)<6
WHERE employeeid<5 and rank(revenue)<6
WHERE employeeid in (1,2,4) and rank(revenue)<6
WHERE rank(qtysold)<100
FROM sales
WHERE rank(qtysold)<100 AND aggregate(qtysold BY siteid)>1000
WHERE employeeid<5 and rank(revenue)<6
WHERE employeeid in (1,2,4) and rank(revenue)<6
The present patent application is a divisional of U.S. patent application Ser. No. 10/186,344, filed on Jun. 27, 2002 now U.S. Pat. No. 7,499,910, entitled “Detecting and Processing Cache Hits for Queries with Aggregates” and is incorporated by reference herein in its entirety and for all purposes.
Number | Name | Date | Kind |
---|---|---|---|
5727196 | Strauss et al. | Mar 1998 | A |
5781896 | Dalal | Jul 1998 | A |
6073129 | Levine et al. | Jun 2000 | A |
6202060 | Tran | Mar 2001 | B1 |
6347314 | Chidlovskii | Feb 2002 | B1 |
6438594 | Bowman-Amuah | Aug 2002 | B1 |
6446062 | Levine | Sep 2002 | B1 |
6466931 | Attaluri et al. | Oct 2002 | B1 |
6601062 | Deshpande et al. | Jul 2003 | B1 |
6618719 | Andrei | Sep 2003 | B1 |
6629094 | Colby et al. | Sep 2003 | B1 |
6708166 | Dysart et al. | Mar 2004 | B1 |
6732091 | Middelfart | May 2004 | B1 |
6950823 | Amiri et al. | Sep 2005 | B2 |
7188100 | De Bellis et al. | Mar 2007 | B2 |
20010049685 | Carey et al. | Dec 2001 | A1 |
20020099691 | Lore et al. | Jul 2002 | A1 |
20020129032 | Bakalash et al. | Sep 2002 | A1 |
20030200194 | Arnold et al. | Oct 2003 | A1 |
20040230572 | Omoigui | Nov 2004 | A1 |
Number | Date | Country | |
---|---|---|---|
20080077557 A1 | Mar 2008 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 10186344 | Jun 2002 | US |
Child | 11931947 | US |