Snowflake Inc.

Florian Andreas Funke

Thierry Cruanes

Benoit Dageville

Jiaqi Yan

Peter Povinec

Guangyan Hu

Bowei Chen

Allison Waingold Lee

Marcin Zukowski

Selcuk Aya

Xinzhu Cai

Megha Thakkar

A method implementing a fault-tolerant data warehouse including allocating a plurality of processing units to a data warehouse, the processing units located in different availability zones, an availability zone comprising one or more data centers. The method further includes, as a result of monitoring a number of queries running at an input degree of parallelism on the plurality of processing units of the data warehouse, determining that the number of queries is serviceable by one fewer processing units. The method further includes routing a query from a first processing unit to a second processing unit within the data warehouse, the query having a common session identifier with another query previously provided to the second processing unit, the second processing unit determined to be caching a data segment associated with a cloud storage resource, usable by the query, and removing the first processing unit from the data warehouse.

Multi-cluster warehouse

Joining data using a disjunctive operator is described. An example computer-implemented method can include generating, with a processing device, a query plan for a query, the query comprising a join operator expression for a disjunctive predicate, wherein the join operator expression includes a conjunctive predicate and a disjunctive operator. The method may further include generating a bloom filter for the disjunctive operator. Additionally, the method may include generating a result set as a result of evaluating the join operator expression using the disjunctive operator and bloom filter for the disjunctive predicate.

System and method for disjunctive joins

Joining data using a disjunctive operator using a lookup table is described. An example computer-implemented method can include receiving a query with a set of conjunctive predicates and a set of disjunctive predicates. The method may also include generating a lookup table for each predicate in the sets of conjunctive predicates and disjunctive predicates. The method, for each row in a probe-side table, may also further include looking up a value associated with that row in each of the lookup tables and adding the row to a results set when there is a match. Additionally, the method may also include returning the results set.

System and method for disjunctive joins using a lookup table

Provided herein are systems and methods for handling build-side skew. For example, a method includes computing a plurality of hash values for a join operation. The join operation uses a corresponding plurality of row sets. The plurality of hash values are sampled to detect a frequent hash value. A build-side row set is partitioned using the frequent hash value to generate a partitioned build-side row set. The build-side row set is selected from the plurality of row sets. The partitioned build-side row set is distributed to a plurality of hash-join-build (HJB) instances executing at a corresponding plurality of servers.

Build-side skew handling for hash-partitioning hash joins

A method implementing a fault-tolerant data warehouse using availability zones includes allocating a plurality of processing units to a data warehouse, the processing units located in different availability zones, an availability zone comprising one or more data centers. The method further includes routing a query to a processing unit within the data warehouse, the query having a common session identifier with a query previously provided to the processing unit, the processing unit determined to be caching a data segment associated with a cloud storage resource independent of the plurality of processing units. The method further includes, as a result of monitoring a number of queries running at an input degree of parallelism, determining that the processing capacity of the processing units has reached a threshold; and changing a total number of processing units using the input degree of parallelism and the number of queries.

Data in a micro-partition of a table is stored in a compressed form. In response to a database query on the table comprising a filter, the portion of the data on which the filter operates is decompressed, without decompressing other portions of the data. Using the filter on the decompressed portion of the data, the portions of the data that are responsive to the filter are determined and decompressed. The responsive data is returned in response to the database query. When a query is run on a table that is compressed using dictionary compression, the uncompressed data may be returned along with the dictionary look-up values. The recipient of the data may use the dictionary look-up values for memoization, reducing the amount of computation required to process the returned data.

Efficient database query evaluation

The subject technology receives a query plan, the query plan comprising a set of query operations, the set of query operations including at least one aggregation and at least one join operation. The subject technology analyzes the query plan to identify an aggregation that is redundant. The subject technology removes the aggregation based at least in part on the analyzing. The subject technology determines at least one aggregation property corresponding to at least one query operation of the query plan. The subject technology inserts at least one adaptive aggregation operator in the query plan based at least in part on the at least one aggregation property, the at least one aggregation property comprising a set of aggregation properties. The subject technology provides a modified query plan based at least in part on the inserted at least one adaptive aggregation operator in the query plan.

Placement of adaptive aggregation operators and properties in a query plan

A method includes determining that an amount of available space in a first memory used by a set of relational queries is insufficient for a query, wherein the query comprises a join operation. A first partition of a set of partitions is identified, wherein the first partition possesses a smallest available probe memory size of the set of partitions and a build memory size greater than or equal to a threshold memory size, wherein the threshold memory size is a percentage of a maximum build memory size, and the largest partition of the set of partitions has the maximum build memory size. The first partition is copied from the first memory to a second memory.

Systems and methods for spilling data for hash joins

A method for a multi-cluster warehouse includes allocating processing units as part of a data warehouse. The processing units access data within one or more databases in cloud storage resources. The method also includes providing one or more queries to each processing unit within the data warehouse. In response to the queries, each processing unit performs database operations on a particular portion of a database table. The method also includes monitoring a workload of the processing units to determine that a processing capacity of the processing units has reached a threshold processing capacity. The method also includes dynamically adding additional processing units to and removing processing units from the data warehouse as needed based on a workload of the processing units.

A method for a multi-cluster warehouse includes allocating a plurality of compute clusters as part of a virtual warehouse. The compute clusters are used to access and perform queries against one or more databases in one or more cloud storage resources. The method includes providing queries for the virtual warehouse to each of the plurality of compute clusters. Each of the plurality of compute clusters of the virtual warehouse receives a plurality of queries so that the computing load is spread across the different clusters. The method also includes dynamically adding compute clusters to and removing compute clusters from the virtual warehouse as needed based on a workload of the plurality of compute clusters.

The subject technology receives a query plan, the query plan comprising a set of query operations, the set of query operations including at least one aggregation. The subject technology analyzes the at least one aggregation to generate a modified query plan, the modified query plan including at least a top aggregation operator, an intermediate aggregation operator, and a bottom aggregation operator. The subject technology performs, with respect to the intermediate aggregation operator, at least one operation comprising: the subject technology receives an input intermediate data type; the subject technology performs an internalize operation on the input intermediate data type to generate an internal state; the subject technology performs an accumulate operation on the internal state to generate intermediate data; and the subject technology performs an externalize operation on the intermediate data to generate an output data type.

Framework for providing intermediate aggregation operators in a query plan

Joining data using a disjunctive operator is described. An example computer-implemented method can include generating a query plan for a query, wherein there is a join operator expression for each of a plurality of disjunctive predicates and each join operator expression includes at least a conjunctive predicate and a disjunctive operator. The method may also include generating a bloom filter for each of the plurality of disjunctive operators. The method may further include evaluating each of the plurality of join operator expressions using a corresponding one of the plurality of disjunctive operators and bloom filter for each of the plurality of disjunctive predicates to generate a result set.

Growing community of inventors

San Mateo, CA, United States of America

Florian Andreas Funke