Overview of BigQuery storage
This page describes the storage component of BigQuery.
BigQuery storage is optimized for running analytic queries over large datasets. It also supports high-throughput streaming ingestion and high-throughput reads. Understanding BigQuery storage can help you to optimize your workloads.
One of the key features of BigQuery's architecture is the separation of storage and compute. This allows BigQuery to scale both storage and compute independently, based on demand.
When you run a query, the query engine distributes the work in parallel across multiple workers, which scan the relevant tables in storage, process the query, and then gather the results. BigQuery executes queries completely in memory, using a petabit network to ensure that data moves extremely quickly to the worker nodes.
Here are some key features of BigQuery storage:
Managed. BigQuery storage is a completely managed service. You don't need to provision storage resources or reserve units of storage. BigQuery automatically allocates storage for you when you load data into the system. You only pay for the amount of storage that you use. The BigQuery pricing model charges for compute and storage separately. For pricing details, see BigQuery pricing.
Durable. BigQuery storage is designed for 99.999999999% (11 9's) annual durability. BigQuery replicates your data across multiple availability zones to protect from data loss due to machine-level failures or zonal failures. For more information, see Availability and durability.
Encrypted. BigQuery automatically encrypts all data before it is written to disk. You can provide your own encryption key or let Google manage the encryption key. For more information, see Encryption at rest.
Efficient. BigQuery storage uses an efficient encoding format that is optimized for analytic workloads. If you want to learn more about BigQuery's storage format, see the blog post Inside Capacitor, BigQuery's next-generation columnar storage format.
The majority of the data that you store in BigQuery is table data. Table data includes standard tables, table clones, table snapshots, and materialized views. You are billed for the storage that you use for these resources.
Table snapshots are point-in-time copies of tables. Table snapshots are read-only, but you can restore a table from a table snapshot. BigQuery only stores the delta between a table snapshot and its base table.
Materialized views are precomputed views that periodically cache the results of the view query. The cached results are stored in BigQuery storage.
In addition, cached query results are stored as temporary tables. You aren't charged for this storage.
External tables are a special type of table, where the data resides in a data store that is external to BigQuery, such as Cloud Storage. An external table has a table schema, just like a standard table, but the table definition points to the external data store. In this case, only the table metadata is kept in BigQuery storage. BigQuery does not charge for external table storage, although the external data store might charge for storage.
BigQuery organizes tables and other resources into logical containers called datasets. How you group your BigQuery resources affects permissions, quotas, billing, and other aspects of your BigQuery workloads. For more information and best practices, see Organizing BigQuery resources.
BigQuery storage also holds metadata about your BigQuery resources. You aren't charged for metadata storage.
When you create any persistent entity in BigQuery, such as a table, view, or user-defined function (UDF), BigQuery stores metadata about the entity. This is true even for resources that don't contain any table data, such as UDFs and logical views.
Metadata includes information such as the table schema, partitioning and clustering specifications, table expiration times, and other information. This type of metadata is visible to the user and can be configured when you create the resource. In addition, BigQuery stores metadata that it uses internally to optimize queries. This metadata is not directly visible to users.
Many traditional database systems store their data in row-oriented format, meaning rows are stored together, with the fields in each row appearing sequentially on disk. Row-oriented databases are efficient at looking up individual records. However, they can be less efficient at performing analytical functions across many records, because the system has to read every field when accessing a record.
BigQuery stores table data in columnar format, meaning it stores each column separately. Column-oriented databases are particularly efficient at scanning individual columns over an entire dataset.
Column-oriented databases are optimized for analytic workloads that aggregate data over a very large number of records. Often, an analytic query only needs to read a few columns from a table. For example, if you want to compute the sum of a column over millions of rows, BigQuery can read that column data without reading every field of every row.
Another advantage of column-oriented databases is that data within a column typically has more redundancy than data across a row. This characteristic allows for greater data compression by using techniques such as run-length encoding, which can improve read performance.
BigQuery does not support foreign keys. This makes BigQuery more suitable for OLAP and data warehouse workloads than OLTP workloads. For OLTP workloads, consider using either Cloud Spanner or Cloud SQL. BigQuery supports federated queries with both of these database services.
Optimizing tables for your data
When you design your tables, consider the following ways that you can optimize their performance.
Nested and repeated fields
BigQuery's internal storage format supports both nested fields and repeated fields.
A nested field is a complex data type that contains one or more strongly typed subfields. In standard SQL, nested fields are represented by the
A repeated field is an ordered list of zero or more values, all with the same data type. In standard SQL, a repeated field is represented as an
For more information, see Specify nested and repeated columns in table schemas.
A repeated, nested field is an efficient way to store denormalized data in
BigQuery. For example, suppose that a customer order can have
multiple line items. The line items contain information such as product name,
quantity, and total cost. In a traditional RDBMS, you might create an
table and an
OrderItems table, and join them with a foreign key.
In BigQuery, you could represent the line items as an array of
nested fields inside the
Orders table. Running a query on a single table with
repeated, nested fields is often more efficient than joining multiple tables.
For historical data, such as customer transactions, you can denormalize the data
as part of an ETL or ELT process when loading the data into
For more information about this scenario, see the blog post BigQuery explained: Working with joins, nested and repeated data.
A partitioned table is a table that is divided into smaller parts, called partitions, based on the value of one column in the table, called the partition key. The partition key can be:
- A time-unit column, such as a
- An integer column
- The time when the data was ingested. In this case, BigQuery automatically stores the ingestion time in a pseudo-column that is not part of the table schema.
Partitions are divided into buckets. For partitions based on time-unit columns or ingestion time, the bucket size can be hourly, daily, monthly, or yearly. For integer columns, you specify the size of the buckets as an integer range. When you load data into a partitioned table, each row is placed into the partition that corresponds to the value of the partition key.
The following diagram shows a table that uses daily partitioning on a column
transaction_date. Three partitions are shown.
Partitioning can speed up a query that filters on the partition key, because BigQuery only has to scan the partitions that match the filter, and can skip partitions that don't match. This process is called pruning.
Partitioning works best when you have a low number of distinct values in the partitioning key — a few thousand or less. For time-based partitioning, you can set a partition expiration time, so that older partitions are automatically deleted.
For more information, see Introduction to partitioned tables.
Clustering organizes a table based on the contents of one or more columns, so that similar values are colocated. When a table is clustered, BigQuery sorts the table data using the values in the clustering columns. These values are used to organize the data into multiple blocks in BigQuery storage. For each block, BigQuery tracks the minimum and maximum values of the clustered columns within that block.
The following diagram shows a table that is clustered on a column named
When new data is added to a table or a specific partition, BigQuery performs automatic re-clustering in the background to restore the sort property of the table or partition.
Clustering can improve query performance in several ways:
When a query contains a filter clause based on the clustering columns, BigQuery uses the sorted blocks to eliminate scans of unnecessary data.
When a query aggregates data based on the values in the clustering columns, performance is improved because similar values are colocated, which can reduce data shuffling.
Clustering can improve joins over star schemas, when a large fact table is joined with smaller dimension tables.
Clustering works best when you have a large number of distinct values in one or
more columns, and your queries typically filter or aggregate across a small
range of those values at a time. You can combine clustering with partitioning.
For example, your data can be partitioned by a
TIMESTAMP column and
then clustered on a different set of columns. Data within each partition is
clustered according to the clustering columns. The following diagram shows a
table that is partitioned on the
transaction_date column and clustered on the
For more information, see Introduction to clustered tables.
Ingesting data into BigQuery storage
There are several basic patterns for ingesting data into BigQuery.
Batch load: Load your source data into a BigQuery table in a single batch operation. This can be a one-time operation or you can automate it to occur on a schedule. A batch load operation can create a new table or append data into an existing table.
Streaming: Continually stream smaller batches of data, so that the data is available for querying in near-real-time.
Generated data: Use SQL statements to insert rows into an existing table or write the results of a query to a table.
Reading data from BigQuery storage
Most of the time, you store data in BigQuery in order to run analytical queries on that data. However, sometimes you might want to read records directly from a table. BigQuery provides several ways to read table data:
Paginate through the records in a table by calling the
tabledata.listREST API method. This method returns structured row responses in a paginated fashion, appropriate for small result sets. However, this method has relatively low throughput and supports only one reader at a time. For more information, see Browsing table data.
Export a table to Cloud Storage by running an extract job. You can save the data to a variety of file formats such as CSV, JSON, and Avro. You can export up to 1 GB of table data to a single file. For larger tables, the results are written to multiple files.
Use the BigQuery Storage Read API. The Storage Read API provides direct read access to data stored in BigQuery tables and supports parallel reads. When you use the Storage Read API, structured data is sent over the wire in a binary serialization format. Callers can read data from BigQuery tables in multiple parallel streams.
For pricing information, see Data extraction pricing.
- Learn about working with tables.
- Learn how to optimize storage.
- Learn how to query data in BigQuery.
- Learn about data security and governance.