摘要:
Inverse distribution operations are performed on a large distributed parallel database comprising a plurality of distributed data segments to determine a data value at a predetermined percentile of a sorted dataset formed on one segment. Data elements from across the segments may be first grouped, either by partitioning keys or by hashing, the groups are sorted into a predetermined order, and data values corresponding to the desired percentile are picked up at a row location of the corresponding data element of each group. For a global dataset that is spread across the database segments, a local sort of data elements is performed on each segment, and the data elements from the local sorts are streamed in overall sorted order to one segment to form the sorted dataset.
摘要:
The invention identifies and caches query plans in a shared-nothing distributed data store that are unlikely to become invalid because they do not reference objects that are likely to be changed or deleted. Plans that are likely to become invalid and are not cached are those plans that reference data that is partitioned across segment/query execution nodes of the data store, plans that are complex, and plans that reference objects that are not “built-in” (primitive) objects. The effect is that most plans which are generated on a query dispatch (master) node are not cached, whereas most plans generated on an execution (segment) node are cached.
摘要:
A script is run on a database to transform the metadata and produce an upgraded database. A new database corresponding to the upgraded database is initialized, and the metadata in the new database catalog is verified by comparing it to the upgraded database metadata. A fast verification is performed on a partial upgrade by dumping the catalogs of master nodes and comparing the results, and a thorough verification is performed on a full upgrade by querying and comparing both master node catalogs and segment node catalogs.
摘要:
A catalog of one or more metadata objects associated with metadata is stored in a persistent manner. The data objects are stored at least in part on a user node included in a data management system. A request associated with accessing a metadata object included in the persistent catalog is received. In response to receiving the request, access to the requested metadata object is provided.
摘要:
A computer readable storage medium includes executable instructions to define a map-reduce document that coordinates processing of data in a distributed database. The map-reduce document complies with a map-reduce specification that integrates map-reduce functions with queries in a query language. The operations specified by the map-reduce document are executed in the distributed database.
摘要:
Testing a database is disclosed. A test description defining a test of a database is received. The test description specifies a relative timing between an issuance of a first command via a first session of the database when the test is conducted and an issuance of a second command via a second session of the database when the test is conducted. The database is configured to conduct the test defined in the test description.
摘要:
Formatting data is disclosed. An indication of a specified data to be formatted between a format of a database and a format external to database is received. A formatter of the database is used to format the specified data between the format of the database and the format external to the database. The formatter has been integrated with the database using a formatter code defined external to the database.
摘要:
A method for aggregating sparse data in a multidimensional array by using a composite join hierarchy created by segmenting the data so that each segment of the hierarchy processed is smaller and more likely to fit in memory. The method employs a recursive sub-cubing mechanism wherein an n-dimensional cube is broken into a number of (n−1)-dimensional cubes and each of those cubes are solved as (n−2)-dimensional cubes etc. Within each division, the processing is segmented by hierarchy level so a dimension with three hierarchy levels (for example, month-quarter-year) would form three separate subcubes with one less dimension. This algorithm produces one ‘worklist’ for every combination of hierarchy levels in the cube. Each of these worklists is represented as a bitmap of the cells contained within it and may be used as a basis of generating more aggregate worklists. To minimize the need for input-output data transfers, all the derived worklists of a single worklist are generated at the same time. This is accomplished without keeping more than n-worklists active at any given time to reduce the number of input-output data transfers needed without requiring substantially larger memory space.
摘要:
A process for upgrading a mirrored shared-nothing database system comprises a sequence of short well-defined idempotent steps, and at least one non-idempotent step involving transforming a master catalog. The upgrade process is managed and controlled by a state machine that has a persistent memory running on the master node. In the event of a failure or crash during an idempotent step, the process stops the database in the current state and repeats the step. If a failure or crash occurs during a non-idempotent step, the upgrade process is rolled back to the beginning and repeated.
摘要:
In a multidimensional database, an aggregation operation is performed in an optimal manner by storing the values included in the/aggregation operation on the same disk page. A sparsity manager determines aggregate values that are computed from other data values during the aggregation operation. Each aggregate value is associated with one or more data values that are used during the aggregation operation to compute the aggregate value. The sparsity manager stores the associated data values in proximity to each other, such as on the same disk page, so that multiple disk page fetches may not be required for the same set of data values during the aggregation operation. The data values used in the aggregation operation can therefore be fetched once from a disk page, and thereafter are found in memory, such as on a cache page corresponding to the disk page. In this manner, multiple fetches for the same disk page during the aggregation operation are avoided.