BigQuery Storage Read API overview

The BigQuery Storage Read API provides fast access to BigQuery-managed storage by using an rpc-based protocol.

Background

Historically, users of BigQuery have had two mechanisms for accessing BigQuery-managed table data:

  • Record-based paginated access by using the tabledata.list or jobs.getQueryResults REST API methods. The BigQuery API provides structured row responses in a paginated fashion appropriate for small result sets.

  • Bulk data export using BigQuery extract jobs that export table data to Cloud Storage in a variety of file formats such as CSV, JSON, and Avro. Table exports are limited by daily quotas and by the batch nature of the export process.

The BigQuery Storage Read API provides a third option that represents an improvement over prior options. When you use the Storage Read API, structured data is sent over the wire in a binary serialization format. This allows for additional parallelism among multiple consumers for a set of results.

The Storage Read API does not provide functionality related to managing BigQuery resources such as datasets, jobs, or tables.

Key Features

  • Multiple Streams: The Storage Read API allows consumers to read disjoint sets of rows from a table using multiple streams within a session. This facilitates consumption from distributed processing frameworks or from independent consumer threads within a single client.

  • Column Projection: At session creation, users can select an optional subset of columns to read. This allows efficient reads when tables contain many columns.

  • Column Filtering: Users may provide simple filter predicates to enable filtration of data on the server side before transmission to a client.

  • Snapshot Consistency: Storage sessions read based on a snapshot isolation model. All consumers read based on a specific point in time. The default snapshot time is based on the session creation time, but consumers may read data from an earlier snapshot.

Enabling the API

The Storage Read API is distinct from the BigQuery API, and shows up separately in the Google Cloud Console as the BigQuery Storage API. However, the Storage Read API is enabled in all projects in which the BigQuery API is enabled; no additional activation steps are required.

Permissions

Establishing a read session to a BigQuery table requires permissions to two distinct resources within BigQuery: the project that controls the session and the table from which the data is read. Note that permission to the table can be granted through any of the following:

More detailed information about granular BigQuery permissions can be found on the Predefined roles and permissions page.

Basic API flow

This section describes the basic flow of using the Storage Read API. For examples, see the libraries and samples page.

Create a session

Storage Read API usage begins with the creation of a read session. The maximum number of streams, the snapshot time, the set of columns to return, and the predicate filter are all specified as part of the ReadSession message supplied to the CreateReadSession RPC.

The ReadSession response contains a set of Stream identifiers. When a read session is created, the server determines the amount of data that can be read in the context of the session and creates one or more streams, each of which represents approximately the same amount of table data to be scanned. This means that, to read all the data from a table, callers must read from all Stream identifiers returned in the ReadSession response. This is a change from earlier versions of the API, in which no limit existed on the amount of data that could be read in a single stream context.

The ReadSession response contains a reference schema for the session and a list of available Stream identifiers. Sessions expire automatically and do not require any cleanup or finalization. The expiration time is returned as part of the ReadSession response and is guaranteed to be at least 6 hours from session creation time.

Read from a session stream

Data from a given stream is retrieved by invoking the ReadRows streaming RPC. Once the read request for a Stream is initiated, the backend will begin transmitting blocks of serialized row data. If there is an error, you can restart reading a stream at a particular point by supplying the row offset when you call ReadRows.

To support dynamic work rebalancing, the Storage Read API provides an additional method to split a Stream into two child Stream instances whose contents are, together, equal to the contents of the parent Stream. For more information, see the API reference.

Decode row blocks

Row blocks must be deserialized once they are received. Currently, users of the Storage Read API may specify all data in a session to be serialized using either Apache Avro format, or Apache Arrow.

The reference schema is sent as part of the initial ReadSession response, appropriate for the data format selected. In most cases, decoders can be long-lived because the schema and serialization are consistent among all streams and row blocks in a session.

Schema conversion

Avro Schema Details

Due to type system differences between BigQuery and the Avro specification, Avro schemas may include additional annotations that identify how to map the Avro types to BigQuery representations. When compatible, Avro base types and logical types are used. The Avro schema may also include additional annotations for types present in BigQuery that do not have a well defined Avro representation.

To represent nullable columns, unions with the Avro NULL type are used.

BigQuery Standard SQL type Avro type Avro schema annotations
BOOLEAN boolean
INT64 long
FLOAT64 double
BYTES bytes
STRING string
DATE int logicalType: date
DATETIME string sqlType: DATETIME
TIMESTAMP long logicalType: timestamp-micros
TIME long logicalType: time-millis
NUMERIC bytes logicalType: decimal (precision = 38, scale = 9)
NUMERIC(P[, S]) bytes logicalType: decimal (precision = P, scale = S)
BIGNUMERIC bytes logicalType: decimal (precision = 77, scale = 38)
BIGNUMERIC(P[, S]) bytes logicalType: decimal (precision = P, scale = S)
GEOGRAPHY string sqlType: GEOGRAPHY
ARRAY array
STRUCT record

Arrow Schema Details

The Apache Arrow format works well with Python data science workloads.

For cases where multiple BigQuery types converge on a single Arrow data type, the metadata property of the Arrow schema field indicates the original data type.

If you're working in an older version of the Storage Read API, then use the appropriate version of Arrow as follows:

  • v1beta1: Arrow 0.14 and earlier
  • v1: Arrow 0.15 and later

Regardless of API version, to access API functions, we recommend that you use the BigQuery Storage API client libraries. The libraries can be used with any version of Arrow and don't obstruct its updates.

BigQuery Standard SQL type Arrow logical type Notes
BOOLEAN Boolean
INT64 Int64
FLOAT64 Double
BYTES Binary
STRING Utf8
DATE Date 32-bit days since epoch
DATETIME Timestamp Microsecond precision, no timezone
TIMESTAMP Timestamp Microsecond precision, UTC timezone
TIME Time Microsecond precision
NUMERIC Decimal Precision = 38, scale = 9
NUMERIC(P[, S]) Decimal Precision = P, scale = S
BIGNUMERIC Decimal256 Precision = 76, scale = 38
BIGNUMERIC(P[, S]) Decimal256 Precision = P, scale = S
GEOGRAPHY Utf8
ARRAY List
STRUCT Struct

Limitations

  • Because the Storage Read API operates on managed storage, you cannot use the Storage Read API to directly read other types of BigQuery tables such as federated tables, logical views, or materialized views.

  • There are restrictions on the ability to reorder projected columns and the complexity of row filter predicates. Currently, filtering support when serializing data using Apache Avro is more mature than when using Apache Arrow.

Supported regions

The Storage Read API is supported in the same regions as BigQuery. See the Dataset locations page for a complete list of supported regions and multi-regions.

Quotas and limits

For Storage Read API quotas and limits, see Storage Read API limits.

Pricing

For information on Storage Read API pricing, see the Pricing page.