Databricks Inc.

Cloudera, Inc.

Alexander Behm

Shoumik Palkar

Sriram Krishnamurthy

Mostafa Mokhtar

Ankur Dave

Shi Xin

Herman Rudolf Petrus Catharina Van Hövell Tot Westerflier

Prashanth Menon

Utkarsh Agarwal

David Cashman

Ryan Deng

The present application discloses a method, system, and computer system for processing data. The method includes obtaining a query plan for processing input data in response to a query, obtaining the input data, selecting a batch of the input data, creating a metadata structure for the batch, allocating one or more contiguous parts of a memory for processing the batch, processing the batch in accordance with the metadata structure to generate resulting data, and storing each array of the resulting data for the batch in one of the one or more contiguous parts of the memory.

Nested array batch processing

Disclosed herein is a method, system, or non-transitory computer readable medium for evaluating a query on a columnar dataset comprising one or more dictionaries associated with columns in the dataset. The method includes receiving a request to perform a query comprising at least a operator and a request to return information about a value of interest in a columnar dataset stored on cloud storage. At least one column in the columnar dataset is based on a dictionary. The dictionary maps one or more values for a column to one or more respective identifiers. The method determines whether to perform dictionary filtering for the query by calculating a metric based on one or more factors. Responsive to the metric being below a threshold, which may be predetermined, the method performs the dictionary filtering.

Dictionary filtering and evaluation in columnar databases

Disclosed herein is a method, system, or non-transitory computer readable medium for evaluating a query on a columnar dataset comprising one or more dictionaries associated with columns in the dataset. The method includes receiving a request to perform a query comprising at least an operator for a columnar dataset on cloud storage. At least one column in the dataset is based on a dictionary, and the dictionary maps one or more values for a column to one or more respective identifiers. The method evaluates the operator on one or more values of the dictionary to generate an updated dictionary comprising updated values. The method may decode the updated dictionary into an updated column comprising updated data values.

Evaluating expressions over dictionary data

The present application discloses a method, system, and computer system for parsing files. The method includes receiving an indication that a first file is to be processed, determining to begin processing the first file using a first processing engine based at least in part on one or more predefined heuristics, indicating to process the first file using a first processing engine, determining whether a particular error in processing the first file using the first processing engine has been detected, in response to determining that the particular error has been detected, indicate to stop processing the first file using the first processing engine and indicate to continue processing using a second processing engine, and storing in memory information obtained based on processing the first file by one or more of the first processing engine and the second processing engine.

Scan parsing

A system includes a plurality of computing units. A first computing unit of the plurality of computing units comprises: a communication interface configured to receive an indication to roll up data in a data table; and a processor coupled to the communication interface and configured to: build a preaggregation hash table based at least in part on a set of columns and the data table by aggregating input rows of the data table; for each preaggregated hash table entry of the preaggregated hash table: provide the preaggregated hash table entry to a second computing unit of the plurality of computing units based at least in part on a distribution hash value; receive a set of received entries from computing units of the plurality of computing units; and build an aggregation hash table based at least in part on the set of received entries by aggregating the set of received entries.

Hash based rollup with passthrough

Disclosed herein is a method for determining whether to apply a lazy materialization technique to a query run. A data processing service receives a request to perform a query identifying a filter column and a non-filter column in a columnar database. The data processing service accesses a first task of contiguous rows in the filter column from a cloud-based object storage. The data processing service applies a filter defined by the query to the first task. The data processing service generates filter results for the first task that may include a percentage of the first task discarded and a run-time. The data processing service determines, based on the filter results for the first task, a likelihood value that indicates a likelihood of gaining a performance benefit by applying the lazy materialization technique to a second task of the query.

Adaptive approach to lazy materialization in database scans using pushed filters

The problem of distinct value estimation has many applications, but is particularly important in the field of database technology where such information is utilized by query planners to generate and optimize query plans. Introduced is a novel technique for estimating the number of distinct values in a given dataset without scanning all of the values in the dataset. In an example embodiment, the introduced technique includes gathering multiple intermediate probabilistic estimates based on varying samples of the dataset, 2) plotting the multiple intermediate probabilistic estimates against indications of sample size, 3) fitting a function to the plotted data points, and 4) determining an overall distinct value estimate by extrapolating the objective function to an estimated or known total number of values in the dataset.

Distinct value estimation for query planning

A system comprises an interface, a processor, and a memory. The interface is configured to receive a query. The processor is configured to: determine a set of nodes for the query; determine whether a node of the set of nodes comprises a first engine node type or a second engine node type, wherein determining whether the node of the set of nodes comprises the first engine node type or the second engine node type is based at least in part on determining whether the node is able to be executed in a second engine; and generate a plan based at least in part on the set of nodes. The memory is coupled to the processor and is configured to provide the processor with instructions.

Integrated native vectorized engine for computation

A system for spilling comprises an interface and a processor. The interface is configured to receive an indication to perform a GROUP BY operation, wherein the indication comprises an input table and a grouping column. The processor is configured to: for each input table entry of the input table, determine a key, wherein the key is based at least in part on the input table entry and the grouping column; add the key to a grouping hash table, wherein adding the key to the grouping hash table comprises last-in, first-out (LIFO) spilling when necessary; create an output table based at least in part on the grouping hash table; and provide the output table.

Growing community of inventors

Lafayette, CA, United States of America

Alexander Behm