Snowflake Inc.

Orestis Kostakis

Qiming Jiang

Justin Langseth

Abdul Q Munir

Jeffrey Rosen

Kristopher Wagner

John Reumann

Jianzhun Du

Prayag Chandran Nirmala

Yijun Xie

Subramanian Muralidhar

Prasanna V Krishnan

Matthew J Glickman

Shakhina Pulatova

Boxin Jiang

Megan Marie Schoendorf

Timur Misirpashaev

Shuodong Dang

A data access event may be recognized, using a browser plug-in, wherein the data access event constitutes a reference to previously obtained data. As a result of recognizing the event, the plug-in may send, to a search engine of a data exchange, a set of extracted terms. The plug-in may receive a set of related data listings related to the set of extracted terms. Upon a selection of a data listing from the set of related data listings, the plug-in may install the data listing to an account.

Browser plug-in for marketplace recommendations

A search engine of a data exchange may receive from a user, a query comprising a set of search terms, and retrieve a set of data listings based on the search terms of the query. A data ranking module of the search engine may analyze each of the set of retrieved data listings to determine, for each of the set of retrieved data listings, a set of listing-specific signals and a set of external signals. Listing-specific signals may correspond to attributes or characteristics of data/content within a data listing, while external signals may correspond to a measure of activity in the data exchange that involves a data listing. Based on the listing-specific signals and the external signals analyzed for each retrieved data listing, the set of retrieved data listings may be ordered and presented to the user.

Search in a data marketplace

A system for improving task scheduling on a cloud data platform is provided. A task to be executed using resources of a computing cluster is received. A task execution plan is generated and information about data to be used for the ask is accessed. Resource requirements for executing the task are predicted by applying machine learning to the task execution plan and the information about the data. Assignment data is generated to execute the task on the resources by applying machine learning information about a current state of the resources and predicted resource requirements.

Task-execution planning using machine learning

A set of affinity metrics may be determined for a set of listings, each listing of the set of listings comprising data to be shared through a data exchange, wherein the set of affinity metrics includes a set of characteristics allowing identification of a listing having one or more characteristics in the set of characteristics. For each pair of listings of the set of listings, an affinity score can be calculated, using the set of affinity metrics, and stored as part of the record in an affinity store. One or more listings of the set of listings using the affinity score between the first listing of the set of listings and the one or more listings of the set of listings can be presented.

Similarity-based listing recommendations in a data exchange

The subject technology identifies a set of functions included in a set of files corresponding to a library. The subject technology, for each function in the set of functions, registers the function as a user defined function (UDF). The subject technology generates a name for the function based at least in part on a predetermined prefix, wherein the predetermined prefix comprises an alphanumeric string. The subject technology generates, using at least a particular set of input parameters utilized by the function and a particular type of parameter of each input parameter of the particular set of input parameters, a particular set of source code. The subject technology stores information corresponding to the function in a metadata database. The subject technology provides access to the function in a different application.

Multiple user defined functions registration

A method includes installing, in a consumer database account, a shared-instance database that includes a shared instance of a provider-account database that resides in a provider database account. The shared-instance database includes a first schema that includes provider-account training data, provider-account scoring data, a training function, and a scoring function. The method also includes invoking the training function from the consumer database account, which results in creation in the consumer database account of a second schema that includes a machine-learning-model instance of a machine learning model, and which also results in training the machine-learning model instance with at least the provider-account training data. Additionally, the method includes generating consumer-account scoring data by inputting, into the trained machine-learning-model instance, consumer-account input data that is stored in the consumer database account. The method also includes storing the consumer-account scoring data in the consumer database account.

Multi-party machine learning using a database cleanroom

Systems, methods, and machine-readable storage devices provide for identifying a user dataset on a distributed database. The system includes generating a similarity score dataset that indicates a similarity between the user dataset and a plurality of datasets of other users of the distributed database. The system generates a plurality of overlap queries that are configured to output overlap datasets between the user dataset and one or more of the plurality of datasets. The system further generates a results dataset by applying one or more of the plurality of overlap queries to a joined dataset comprising data from the user dataset and one of the plurality of datasets of other users on the distributed database.

Overlap queries on a distributed database

A method includes parsing, by at least one hardware processor, a query to determine query comments and query code associated with the query. A query execution plan is generated based on the query code. Query execution using the query code is performed at a first computing node associated with a query processing pipeline. A detection is made that the query comments are indicative of a software bug in the query code based on analysis of the query comments. The detection is performed at a second computing node associated with a query analysis pipeline. A notification of the software bug and a result of the query execution is output.

Software bugs detection using query analysis

Systems and methods for managing input and output error of a machine learning (ML) model in a database system are presented herein. A set of test queries is executed on a first version of a database system to generate first test data, wherein the first version of the system comprises a ML model to generate an output corresponding to a function of the database system. An error model is trained based on the first test data and second test data generated based on a previous version of the system. The error model determines an error associated with the ML model between the first and previous versions of the system. The first version of the system is deployed with the error model, which corrects an output or an input of the ML model until sufficient data has been produced by the error model to retrain the ML model.

Handling system-characteristics drift in machine learning applications

The subject technology receives a query directed to a set of source tables, each source table organized into a set of micro-partitions. The subject technology determines a set of metadata, the set of metadata comprising table metadata, query metadata, and historical data related to the query. The subject technology predicts, using a machine learning model, an indicator of an amount of computing resources for executing the query based at least in part on the set of metadata. The subject technology generates a query plan for executing the query based at least in part on the predicted indicator of the amount of computing resources. The subject technology executes the query based at least in part on the query plan.

Predictive resource allocation for distributed query execution

A system for generating similarity data for different datasets in a cloud data platform. A first dataset of a plurality of datasets on the cloud data platform is identified, where the first dataset is associated with a first user of the cloud data platform. A semantic type for each feature the first dataset is identified, and each semantic type for the first dataset is compared with existing data of the first user. Semantic types for each feature of each dataset are identified, and each semantic type for the first dataset is compared to each semantic type of each dataset. Overlap requests are generated to output overlap datasets between the first dataset and each of the plurality of datasets. A results dataset is generated by applying the overlap requests to a joined dataset comprising data from the first dataset and data from each of the plurality of datasets.

Overlap results data generation on a cloud data platform

A system for improving task scheduling on a cloud data platform is provided. A task is received, from a user of a cloud data platform, for execution on a dataset of a cloud data platform using a plurality of resources. A task graph is generated, and metadata related to the dataset is accessed for use in execution of the task. A predicted resource profile is generated by applying a first machine learning scheme to the task graph and the metadata of the dataset. Assignment data is generated to execute processes of the task on the plurality of resources. The assignment data generated by applying a second machine learning scheme to current state data of a current computational state of the plurality of resources and the predicted resource profile generated by the first machine learning scheme.

Data-driven task-execution scheduling using machine learning

The subject technology identifies a set of functions in a set of files corresponding to a library. The subject technology, for each function, registers the function as a user defined function (UDF) based on a set of input parameters utilized by the function and a type of parameter of each of the input parameters. The subject technology provides access to each registered function in a different application.

Registration of multiple user defined functions

Disclosed herein are embodiments of systems and methods for analyzing query comments for identifying potential software bugs. In an example, a data platform obtains query comments associated with a query. Based on determining that the query comments include a reference to a software bug of the data platform, the data platform generates a software-bug alert based on the query comments, and transmits the software-bug alert to an endpoint.

Analyzing query comments for identifying potential software bugs

Affinity-based listing recommendations are created and used in a public data exchange. Listings can be evaluated against one another for affinity or similarity such that users working with a particular dataset can be presented with other datasets that share an affinity. Affinity can be determined from both the dataset metadata as well as information from the dataset content. Calculation of affinity scores can be pre-computed and stored, in advance of use, or determined on-the-fly. Presentation of most-similar listings can be deterministic, can contain randomization, can employ time-decay, can be weighted, and can make use of a tiered-sum approach.

Growing community of inventors

Redmond, WA, United States of America

Orestis Kostakis