Snowflake is a cloud-based database and is currently offered as a pay-as-you-go service in the Amazon cloud. It is developed by Snowflake Computing. Snowflake adopts a shared-nothing architecture. It uses Amazon S3 for its underlying data storage. It performs query execution within in elastic clusters of virtual machines, called virtual warehouse. The Cloud Service layer stores the collection of services that manage computation clusters, queries, transactions, and all the metadata like database catalogs, access control information and ect. in a key-value store (FoundationDB).
It uses Amazon S3 for its underlying data storage. It performs query execution within in elastic clusters of virtual machines, called virtual warehouse. Upon query execution, virtual warehouse use HTTP-based interface to read/write data from S3. The Cloud Service layer stores the collection of services that manage computation clusters, queries, transactions, and all the metadata like database catalogs, access control information and ect. in a key-value store.
Snowflake is relational as it supports ANSI SQL and ACID transactions. It offers built-in functions and SQL extensions for traversing, flattening, and nesting of semi-structured data, with support for popular formats such as JSON and Avro. When storing semi-structured data, Snowflake can perform automatic type inference to find the most common types and store them using the same compressed columnar format as native relational data. Thus it can accelerate query execution on them.
Snowflake horizontally partitions data into large immutable files which are equivalent to blocks or pages in a traditional database system. Within each file, the values of each attribute or column are grouped together and heav- ily compressed, a well-known scheme called PAX or hybrid columnar. Each table file has a header which, among other metadata, contains the offsets of each column within the file.
Multi-version Concurrency Control (MVCC)
Snowflake supports MVCC. As Snowflake's underlying data storage is done by Amazon S3, each write operation instead of performing writes in place, it creates a new entire file including the changes. The stale version of data is replaced by the newly created file, but is not deleted immediately. Snowflake allows users to define how long the stale version will be kept in S3, which is up to 90 days. Based on MVCC, Snowflake also supports time travel query.
https://docs.snowflake.net/manuals/index.html
Snowflake Computing
2013