Computer systems generally include one or more processors interfaced to a temporary data-storage device such as a memory device and one or more persistent data-storage devices such as disk drives. Each disk drive generally has an associated disk controller. Data is transferred between the disk drives and the disk controllers. Data is also transferred between the disk controller(s) and the memory device over a communications bus or similar.
Data organization in a computer system such as that above is important in relational database systems that deal with complex queries against large volumes of data. Relational database systems allow data to be stored in tables that are organized as both a set of columns and a set of rows. Standard commands are used to define the columns and rows of tables and data is subsequently entered in accordance with the defined structure.
The defined table structure is locally maintained but may not correspond to the physical organization of the data. In a parallel shared nothing relational database data can be stored across multiple data-storage facilities, each data-storage facility in turn including one or more disk drives.
Hash partitioning is a partitioning scheme in which a predefined hash function and map is used to assign rows in a table to respective processing modules and data-storage facilities. The hashing function generates a hash bucket number and the hash numbers are mapped to data-storage facilities.
Data skew during join operations is a troublesome problem for a massively parallel (MPP) database engine. Even though individual tables have effective primary index choices, there are queries that join the two tables that can suffer from skewed processing because there might be tens of millions of rows in one table all pointing to the same key in the second table.
Current approaches to planning joins either ignore the effects of skewing entirely or else require duplicating the “smaller” spool file on all nodes. With large spool files on large database configurations that may be a very costly approach.
The techniques described below show how a database query optimizer can dynamically detect at query execution time that excessive data skew is probably going to occur on a foreign key join and how to insert additional steps into the query plan that will mitigate the effects of excessive skewing. These techniques do not require that the user submitting his or her query provide any hints or special syntax in their SQL and there is no additional overhead on the database administrator to do more than to regularly collect demographic statistics—currently a common practice in the industry.
Described below is a method for determining when a database system query optimizer should employ join skew avoidance steps. The method includes dynamically calculating the worst-case anticipated frequency distribution for a particular relation along a particular set of join column(s) at query execution time. The calculated frequency distribution value is compared to a skew threshold, the skew threshold representing the number of rows on the same distinct value that would lead to avoidable processing inefficiencies. It is then determined that the database system query optimizer should employ join skew avoidance steps if the calculated frequency distribution value exceeds the skew threshold.
Also described below is a method for dynamically determining at query execution time the data values from a set of relational join column(s) that will exceed a skew threshold. The method includes creating a skewed value temporary table, the table associating data values for common join columns between the first and second relation with partitioning factors representing distinct numbers, the partitioning factors used to replicate a proper subset of the second relation rows during the subsequent join operation. The temporary table is populated with one or more join column values from the first relation that have a row frequency greater than the skew threshold and the value-specific partitioning factor.
The data warehouse 100 includes one or more processing modules 1051 . . . M that manage the storage and retrieval of data in data-storage facilities 1101 . . . N. Each of the processing modules 1051 . . . M manages a portion of a database that is stored in a corresponding one of the data-storage facilities 1101 . . . N. Each of the data-storage facilities 1101 . . . N includes one or more disk drives.
The system stores data in one or more tables in the data-storage facilities 1101 . . . N. The rows 1151 . . . Z of the tables are stored across multiple data-storage facilities 1101 . . . N to ensure that the system workload is distributed evenly across the processing modules 1051 . . . N. A parsing engine 120 organizes the storage of data and the distribution of table rows 1151 . . . Z among the processing modules 1051 . . . N. The parsing engine 120 also coordinates the retrieval of data from the data-storage facilities 1101 . . . N over network 125 in response to queries received from a user at a mainframe 130 or a client computer 135 connected to a network 140. The database system 100 usually receives queries and commands to build tables in a standard format, such as SQL.
Product table 240 has product_id 245 as a primary key. Product table 240 has further fields for example product_name 250 as well as other fields.
There is a foreign key relationship between big table 205 and product table 240. This means that big table 205 and product table 240 share one or more common field(s) but the shared field(s) is the primary key of only one of the tables. The product_id field 215 in big table 205 is not a primary key whereas product_id field 245 is a primary key for product table 240. The respective primary keys could well be effective primary index choices for their respective tables. However queries that join the two tables can suffer from skew processing because there might be tens of millions of rows in the big table 205 all pointing to the same product_id key 245 in product table 240.
Users are made aware of the impact of skew when queries execute slowly and ultimately return an error of “out of spool” conditions. General system performance is impacted if the only response by database administrators is to grant such users a larger spool allocation because the processing modules that run hot due to skewed joins end up slowing overall query throughputs. That is, some processing modules 1051 . . . M work harder whereas other processing modules sit idle when they could be helping to complete the join operation.
Currently, the main technique for avoiding the processing impact of a skewed join is to duplicate the smaller of the two tables (for example, product table 240) across all of the processing modules 1051 . . . M so that the rows from the larger table (for example, big table 205) can remain local to the nodes where they were stored prior to the join step. That approach can be problematic, however, if product table 240 is only small in a relative sense—it may contain millions of rows, too. In that case, the processing overhead of duplicating product table 240 across all of the processing modules 1051 . . . M might be significant enough that users and administrators still see “out of spool” conditions and slow performance. What is significant to recognize, however, is that out of possibly millions of rows in product table 240 there may be only a few hundred product_id key values that result in severe skewing. That is, if the database query optimizer can detect in advance that only certain rows from product table 240 need to be replicated across multiple processing modules 105, query throughputs can be maximised at the minimum overall cost to system performance.
The techniques described below attempt to distribute the workload associated necessary to process a skewed join across more processing modules 105 without having to resort to duplicating the smaller spool. The techniques described below permit the database system query optimizer to determine whether it is more effective to employ join the new skew avoidance steps versus the more traditional techniques that it already considers.
It is first determined whether the normal query plan would hash distribute the big table as indicated at 305. If the answer is no then a conventional “duplicate small spool” technique is used 310.
On the other hand if a normal query plan would hash distribute a big table then the expected worst case skew count is computed from collected statistics 315. If the expected skew is greater 320 than the threshold a then the skew avoidance steps are inserted 325 into the query plan. On the other hand if the expected skew is not greater than the threshold value α then 330 a conventional “hash distribute” technique is used.
As shown in
P is joined to the small table and hash distributed on (K, Null) as indicated at 350 when no match. Small table rows are replicated KP times when there is a match on skewed value.
The join is finished on the big table to small table 355 from redistributed spool files of steps 345 and 350 above.
Collected statistics on big table 205 identify the loner or mode values associated with the product_id column having the greatest frequency within the table. It typically is the case that not all of the data values that can cause skewing for joins will be reported in the demographic statistics collected by the database administrator.
The technique described below uses collected statistics to assess the probability that excessive skewing will occur and, if so, add extra steps to the query plan to mitigate the processing impact.
In one example the statistics indicate that big table 205 contains 8 billion rows. The worst case skewing on a single product_id key value is 50 million rows but the user query is not trying to pull all rows out of big table 205—other filter conditions on the query are expected to qualify only 10 million rows from big table 205 that need to be joined to product table 240. Therefore, the anticipated skew value is 10 million×50 million divided by 8 billion=62,500 rows. This anticipated skew value exceeds the target row count or skew threshold. The conclusion is that excessive skewing is expected so a modified join operation is recommended. If the anticipated skew value was less than the skew threshold then a conventional join operation is performed.
In a modified join operation a spool file of the big table 205 is kept local to an individual processing module 105.
An aggregate scan of the spool file generated from big table 205 is performed to determine the most skewed values and associated partition factor. The result is stored in a duplicated table referred to as spool P that is duplicated on all processing modules 105. The following SQL statements simulate this situation. The statements identify those keys that have frequency greater than or equal to α and computes a partitioning factor KP. For purposes of our example, K refers to the product_id column in big table 205. That is, K is the foreign key column(s) in big table 205 that will be used to join over to product table 204.
The temporary duplicated table spool P is merge joined or hash joined to big table 205 on a right outer join via K so that each row that was qualified out of big table 205 is paired up with the anticipated skew factor, KP, for that key value. The result of the first temporary distributed table spool X is hash redistributed on the composite key of (K,β) where β is effectively a randomized integer value between zero and KP−1. In short, the net effect of this aggregate scan, join, and re-distribution is to distribute rows from big table 205 that are NOT expected to be skewed according to the hash of (K, NULL) and to distribute rows that ARE expected to be skewed across multiple processing modules 105 in a randomized fashion depending upon what β works out to be.
The temporary duplicated table spool P is also merge joined or hash joined to product table 204 on a right outer join via K so that we can redistribute those rows appropriately, too. The difference, however, is that rows from product table 204 that match one of the skewed keys in P will be replicated KP times to whichever processing modules 105 own the hash of (K, 0), (K, 1), . . . , (K, KP−1). The assumption, of course, is that there are only a limited number of rows in spool P so this replication is much less overhead than duplicating the entire product table 204 to all processing modules 105.
In some cases, the partition factor for a key may exceed the total number of processing modules in the configuration. In this case KP is capped at the maximum number of processing modules 105. The assumption about a in the preconditions for falling into this logic ensures that spool P is always tiny enough that the duplication overhead is reasonable. The most significant added cost to this newer plan versus conventional approaches is the aggregate scan.
It does not matter if the user query request uses inner or outer joins between big table 205 and product table 204—the techniques are effectively the same. Also, generalizing the example above to involve composite key joins on K1, K2 . . . Ky instead of a single column K works the same.
From a user stand point there is nothing special that has to be altered in SQL statements. The user simply needs to presume that the database administrator collects statistics and keeps them current enough to reflect the actual data demographics on the big table.
The text above describes one or more specific embodiments of a broader invention. The invention also is carried out in a variety of alternative embodiments and thus is not limited to those described here. Those other embodiments are also within the scope of the following claims.
Number | Name | Date | Kind |
---|---|---|---|
5121494 | Dias et al. | Jun 1992 | A |
5873074 | Kashyap et al. | Feb 1999 | A |
20080059412 | Tarin | Mar 2008 | A1 |
20090024568 | Al-Omari et al. | Jan 2009 | A1 |
Number | Date | Country | |
---|---|---|---|
20090248617 A1 | Oct 2009 | US |