Docs/Basics/BigQuery updates
Basics

BigQuery updates

Google's BigQuery release notes since 2011, cut to the changes a SQL user can act on. Entries from August 2026 on also say whether the Querylab.io editor supports the feature.

This page follows Google’s BigQuery release notes back to the first one in 2011 and keeps the entries a SQL user can act on: syntax, functions, INFORMATION_SCHEMA views, model names, editions and pricing. Drivers, SDKs, agents, conversational analytics, pipelines, admin consoles and data transfer connectors are left out. Entries run newest first, and each date links to Google’s note. A launch stage appears only when Google’s note states one.

From August 2026 on, each entry also has a Querylab.io line that says whether the editor knows the feature, such as completions, hover and signature help. Queries run on BigQuery, so this line is about editing help only. Older entries list what BigQuery added and nothing more.

Last updated 18 November 2026, covering Google’s notes through 17 November.

November 2026

  • 17 November, GA. AI.RANK. A table function that sorts the rows of a table or query by how well the text in one column answers a search query. It returns a rank and a score beside the input columns. BigQuery bills it through the ranking API, not by bytes read. Querylab.io: supported. The editor offers AI.RANK after AI., completes rank, score and the input columns, and shows hover and signature help with the cost note. When the arguments are literals and the editor can see the input’s columns, it reports a column_to_rank that is missing or not a STRING, an empty query, an input column called rank or score, and a top_n below 1, with BigQuery’s own messages. It does not check the model name or the argument count and types, and the cost estimate still shows bytes only. The local engine used by Learn does not run it.
  • 17 November, GA. Reranking in AI.SEARCH. With rerank => TRUE, AI.SEARCH reorders the rows it found by how well their text answers the query, the same way AI.RANK does. The result gets a rerank_score column, and options takes a reranker object with num_candidates and model. Querylab.io: supported. The editor offers rerank => with TRUE and FALSE, and shows the reranker options, the requirements of rerank => TRUE and the cost note on hover. When the arguments are literals, it reports with BigQuery’s own messages: an ObjectRef column, a query_value that is empty or is not a string literal or parameter, top_k => 0, a reranker without rerank => TRUE, and a num_candidates that does not fit top_k. It judges only a call over a named table, not over a CTE or a subquery, because BigQuery checks that input first. It does not check the model name, the cap of 10,000,000 candidates (BigQuery checks it only when the query runs), or the rerank_score column, which is not offered in completions yet.
  • 17 November, GA. Optimized mode for AI.IF and AI.CLASSIFY. With optimization_mode => 'MINIMIZE_COST' and an embeddings column, BigQuery calls the model on a sample of rows and answers the rest with a smaller model it trains on that sample. AI.CLASSIFY can do this with several labels per row. Querylab.io: AI.IF, AI.CLASSIFY and AI.SCORE know the new named arguments (examples, embeddings, optimization_mode, max_error_ratio; AI.SCORE only has max_error_ratio). The editor offers the values of optimization_mode and output_mode, and marks a value BigQuery refuses: an unknown optimization_mode or output_mode, an embeddings array written out as a literal, and a max_error_ratio outside 0 to 1. AI.CLASSIFY with output_mode is typed ARRAY<STRING>, and output_mode takes single or multi. The local engine does not run these functions.
  • 13 November, Preview. MySQL, Oracle and PostgreSQL connections. Federated queries can read from these databases through a BigQuery connection. Querylab.io: no change in the editor.
  • 12 November, GA. Model gemini-4-hydrogen. A new Gemini model for the generative AI functions. Querylab.io: the editor offers the name in completions for the endpoint of the generative AI functions and for the ENDPOINT option of a remote model in CREATE MODEL. It still accepts any name.
  • 11 November, GA. JSON_FLATTEN. Returns every value in a JSON value that is not an array, looking inside nested arrays. Querylab.io: supported, with completions, hover and signature help.
  • 10 November, GA. Graph measures. A graph property can be a measure, an aggregate tied to a key so it does not overcount after joins. You read measures with the GRAPH_EXPAND table function and the AGG function. GRAPH_EXPAND now takes a start node, so one graph can hold several fact tables. Querylab.io: the editor parses measures, GRAPH_EXPAND and AGG, and shows hover on measure properties. It offers start_node => and flatten_results =>, and the node tables of the graph as start_node values. When the graph is loaded, SELECT * and column completion over a start_node call list only the tables reachable from that node, and flatten_results => FALSE lists the start node’s columns and the nested STRUCT per joined table. CALL BQ.SHOW_GRAPH_EXPAND_SCHEMA and the five other CALL BQ.* system procedures get completions, hover and signature help, and a wrong number of arguments, a literal of the wrong type or a value where the output variable goes is marked with BigQuery’s own message. A graph with several roots and no start_node, and a start_node that is not a node table, are marked with BigQuery’s own messages. See Graph queries.
  • 10 November, GA. extensions in graph OPTIONS. Labels and properties of a graph take an extensions option, a list of string pairs for custom metadata, beside description and synonyms. Querylab.io: supported. The editor parses OPTIONS after DEFAULT LABEL and after a property, offers the three option names, and reports an unknown option, a value of the wrong type and a repeated option with BigQuery’s own wording.
  • 10 November, GA. Model gemini-4-carbon. A new Gemini model for the generative AI functions. Querylab.io: the editor offers the name in completions for the endpoint of the generative AI functions and for the ENDPOINT option of a remote model in CREATE MODEL. It still accepts any name.
  • 9 November, Preview. Compute cost attribution. Editions customers can opt in to have billed compute charges attributed to the projects that run queries on a reservation. Querylab.io: no change in the editor.
  • 9 November, Preview. Compression functions. ZSTD_COMPRESS compresses a STRING or BYTES value with Zstandard. ZSTD_DECOMPRESS_TO_BYTES and ZSTD_DECOMPRESS_TO_STRING turn it back into BYTES or STRING. Querylab.io: supported, with completions, hover and signature help.
  • 6 November, Announcement. Graph queries on on-demand pricing. GQL queries run on on-demand pricing and on Enterprise and Enterprise Plus reservations. Standard edition reservations cannot run them. Querylab.io: supported. Auto routing can send a graph query to on-demand or an Enterprise reservation, never to a Standard one. See Graph queries.
  • 3 November. SQL UDFs in continuous queries. Continuous queries can call persistent and temporary SQL user-defined functions. Other kinds of UDF are still not allowed. Querylab.io: no change in the editor.
  • 2 November, GA. AI.PARSE_DOCUMENT. Reads a document with the Document AI layout parser and returns its titles, headings, paragraphs, lists and tables as JSON. Querylab.io: supported, with completions, hover and signature help. See AI functions.
  • 2 November, Preview. SECURE_CONTEXT. Returns a value passed in at query time, so a logical view can filter rows by user or tenant. Querylab.io: supported, with completions, hover and signature help. The editor also reports the call when it is used outside a logical view or row access policy definition, as BigQuery does.
  • 2 November, GA. APPENDS and continuous queries on Iceberg managed tables. The APPENDS table function lists the rows added to an Iceberg managed table in a time range. Continuous queries can read from and write to these tables. Querylab.io: the editor parses APPENDS and does not check its arguments.

October 2026

  • 27 October, GA. AI.COUNT_TOKENS. Estimates how many tokens a text string uses. Querylab.io: supported, with completions, hover and signature help. See AI functions.
  • 26 October, Preview. REGEXP_EXTRACT_GROUPS. Returns the text matched by each capturing group of a regular expression. Querylab.io: supported, with completions, hover and signature help.
  • 22 October, GA. AI.TRANSLATE. Translates text between languages with a translation model or an LLM. Querylab.io: supported, with completions, hover and signature help. See AI functions.
  • 21 October, GA. AI.CHUNK_DOCUMENT. A table function that splits documents into text chunks for embeddings. Querylab.io: supported, with completions, hover and signature help.
  • 15 October, GA. Reusable column aliases. A SELECT alias can be used by later items in the same list and by the WHERE of the same query. It is off until you turn on the project option. Querylab.io: supported. The editor checks the project option and follows it. See Reusable column aliases.
  • 14 October, GA. output_schema for AI.AGG. Sets the fields of the function’s result as a list of names and types. Querylab.io: supported.
  • 13 October, GA. AI.AGG. Combines many rows of unstructured input into one answer, following an instruction in plain language. Querylab.io: supported, with completions, hover and signature help. See AI functions.
  • 13 October, GA. Model gemini-4-argon. A new Gemini model for the generative AI functions. Querylab.io: the editor offers the name in completions for the endpoint of the generative AI functions and for the ENDPOINT option of a remote model in CREATE MODEL. It still accepts any name.
  • 7 October, Preview. TimesFM 3.0. A time series model. With AI.FORECAST it can forecast several series together and use covariates. AI.EVALUATE and AI.DETECT_ANOMALIES take it for a single series. Querylab.io: 'TimesFM 3.0' is offered in completions for model in all three functions. See Forecasting.
  • 5 October, GA. Model gemini-embedding-2. An embedding model for AI.EMBED, AI.SIMILARITY and AI.GENERATE_EMBEDDING. Querylab.io: the editor offers the name in completions for the endpoint of AI.EMBED and AI.SIMILARITY. AI.GENERATE_EMBEDDING takes a model, not a name, so there is nothing to offer. The editor still accepts any name.
  • 5 October, GA. Flexible column names on Iceberg tables. Iceberg external and managed tables allow column names with spaces and non-ASCII characters by default. Querylab.io: supported. Such names are quoted with backticks in completions, in Expand SELECT * and in the SQL converters.

September 2026

  • 30 September, GA. OBJ.LIST. A table function that lists files in Cloud Storage with their metadata and ObjectRef values, without an object table. Querylab.io: supported, with completions, hover and signature help. See Object tables.
  • 29 September, GA. AI.KEY_DRIVERS. Shows which combinations of dimension values explain a change in a metric you can sum. Querylab.io: supported, with completions, hover and signature help.
  • 29 September, Preview. INFORMATION_SCHEMA.FAILOVER_HISTORY. Lists failover events for reservations that use managed disaster recovery. It needs a region qualifier. Querylab.io: supported. The editor accepts the view with a region qualifier and flags it without one. See Replication and disaster recovery.
  • 28 September. Continuous queries can INSERT into Iceberg managed tables. A continuous query can write its output rows to an Iceberg managed table with an INSERT statement. Querylab.io: no change in the editor.
  • 21 September. Model gemini-3.8-flash. A new Gemini model for the generative AI functions. Querylab.io: the editor offers the name in completions for the endpoint of the generative AI functions and for the ENDPOINT option of a remote model in CREATE MODEL. It still accepts any name.
  • 14 September, Preview. WHERE inside an aggregate call. SUM(x WHERE x > 0) filters the rows that feed one aggregate. Querylab.io: supported, including with DISTINCT. See Aggregate functions.
  • 10 September, Preview. ML.METRICS. Computes classification or regression metrics from a table of actual and predicted values, with no stored model. Querylab.io: supported, with completions, hover and signature help.
  • 10 September, Preview. AI.CAUSAL_EFFECT. Estimates how much an event, such as a campaign, changed a time series. Querylab.io: supported, with completions, hover and signature help.
  • 9 September. Models gemini-3.5-flash-lite, gemini-3.6-flash and gemini-3.7-flash. Three new Gemini models for the generative AI functions. Querylab.io: the editor offers the names in completions for the endpoint of the generative AI functions and for the ENDPOINT option of a remote model in CREATE MODEL. It still accepts any name.

August 2026

  • 31 August, GA. BigQuery Graph. Graph is generally available. GQL gains the CALL statement and the path functions IS_ACYCLIC, IS_SIMPLE and IS_TRAIL. Querylab.io: supported. Completions offer the three path functions. See Graph queries.
  • 31 August, Preview. Identity columns. GENERATED ALWAYS AS IDENTITY (...) and GENERATED BY DEFAULT AS IDENTITY (...) make BigQuery generate a unique integer for each inserted row. Querylab.io: supported, including the START WITH and INCREMENT BY options. See Defaults and identity columns.
  • 31 August, Preview. TabFM. A pretrained model for tabular data. AI.PREDICT and AI.EVALUATE select it with model => 'TabFM 1.0'. Querylab.io: 'TabFM 1.0' is offered in completions for model.
  • 27 August, GA. XGBOOST_VERSION. A CREATE MODEL option that picks the XGBoost version (0.9, 1.1 or 2.1) for boosted tree and random forest models. Querylab.io: supported. The option appears in completions. See BigQuery ML.
  • 27 August, Announcement. Graph processing and editions. Google said core graph processing needs an Enterprise or Enterprise Plus reservation, and that allowlisted users could keep Standard edition or on-demand pricing until 26 April 2027. Graph measures stay available on on-demand pricing. On-demand pricing is allowed for all graph queries since 6 November, see above. The same rule was first announced on 20 August. See On-demand vs capacity.
  • 20 August, Preview. ML.TREND, ML.SEASONALITY and ML.DETECT_CHANGE_POINTS. Table functions that report the trend, the seasonal pattern and the points where a time series changes level or slope. Querylab.io: supported, with completions, hover and signature help.
  • 11 August, GA. ANY TABLE parameters. A table function can take a table of any shape as a parameter. Querylab.io: supported. See Table functions.
  • 10 August, GA. Models gemini-3.1-flash-lite and gemini-3.5-flash. Two Gemini models for the generative AI functions. Querylab.io: the editor offers the names in completions for the endpoint of the generative AI functions and for the ENDPOINT option of a remote model in CREATE MODEL. It still accepts any name.
  • 3 August, Announcement. Hybrid search restored. VECTOR_SEARCH can combine semantic and keyword search, and AI.SEARCH accepts HYBRID mode. Querylab.io: supported, including the lexical arguments of VECTOR_SEARCH. See Vector search.

July 2026

  • 27 July, GA. APPENDS and CHANGES. The change history functions return the rows that were appended to or changed in a table during a time range.
  • 15 July, Preview. Data governance tags. A Resource Manager tag can be attached to sensitive columns. Data policies use the tag to enforce column-level security and data masking.
  • 13 July, GA. Iceberg managed tables. Table partitioning, multi-statement transactions and advanced runtime work on Apache Iceberg managed tables.
  • 13 July, Preview. Project caps. A reservation can limit the maximum slots and the concurrency of each project in it. Google also calls them scheduling policies.
  • 13 July, Preview. ALTER SEARCH INDEX. Changes the configuration of an existing search index.
  • 9 July. Hybrid search in VECTOR_SEARCH disabled. Google turned off the combined semantic and keyword search for now and says it is working to restore it.
  • 8 July, Preview. Multi-level aggregation. An aggregate function can be used as an argument of another aggregate function in GoogleSQL.

June 2026

  • 30 June. principal on reservation assignments. An optional principal property routes queries to a reservation based on the user, service account or third-party identity that runs the job.
  • 25 June, Preview. Hybrid search in VECTOR_SEARCH. VECTOR_SEARCH can combine a semantic search with a keyword search. AI.SEARCH takes a HYBRID mode on tables with autonomous embedding generation, and a vector index can include keyword information to speed up the keyword part.
  • 25 June, Preview. DDL and DML on Lakehouse Iceberg tables. Apache Iceberg tables in the Lakehouse runtime catalog accept CREATE, ALTER and DROP statements and INSERT, UPDATE, DELETE and MERGE from BigQuery.
  • 17 June, GA. Autonomous embedding generation. CREATE TABLE and ALTER TABLE can turn on an embedding column that BigQuery fills and updates from a source column.
  • 15 June. Daily token quotas turned off. Configuring daily token quotas for the generative AI functions is disabled for now. Google says it is working to restore it.
  • 12 June, GA. ObjectRef input for AI functions. The AI functions take ObjectRef values directly, without a call to OBJ.GET_ACCESS_URL.
  • 11 June, Preview. AI.KEY_DRIVERS restored. The function finds the data segments that cause statistically significant changes in a summable metric.
  • 10 June, Preview. ARRAY_AGG and STRING_AGG in continuous queries. Continuous queries support both aggregate functions.
  • 8 June, GA. Daily token quotas. Daily token quotas limit the cost of the generative AI functions.
  • 3 June, GA. Fluid scaling. Autoscaling reservations are billed per second with no minimum duration.
  • 2 June, GA. Custom path in remote function endpoints. Several remote functions can share one Cloud Run service by using different path suffixes on the same endpoint URL.

May 2026

  • 20 May, GA. Python UDFs. A scalar function written in Python can be called from a SQL query. It can install packages from PyPI and reach external services through a Cloud resource connection.
  • 20 May, Preview. AI.AGG. Aggregates unstructured input data by following an instruction in natural language.
  • 18 May, GA. Reservation groups. Reservations in a group share idle slots with each other before the slots go to other reservations in the project.
  • 15 May, GA. Single input table for AI.DETECT_ANOMALIES. The function takes one table that holds both the historical and the target data.
  • 14 May. AI.KEY_DRIVERS turned off. The Preview of the function is disabled for now. Google says it is working to restore it.
  • 12 May, Preview. AI.COUNT_TOKENS. Estimates the token count of a text input. For some generative AI functions, a query can also show the input, output, thought and cache tokens for each modality.

April 2026

  • 29 April, GA. PARTITION BY on CREATE VECTOR INDEX. Partitions a TreeAH vector index, which allows partition pruning and can lower I/O cost.
  • 29 April, GA. VECTOR_INDEX.STATISTICS and ALTER VECTOR INDEX REBUILD. VECTOR_INDEX.STATISTICS measures how far the table data has drifted since the index was created. ALTER VECTOR INDEX REBUILD rebuilds the index without downtime.
  • 28 April, GA. Materialized views over CDC tables. A materialized view can be created over a table with active change data capture.
  • 22 April, Preview. Graph measures. Some graphs can define measures that lock an aggregation to a key, which avoids overcounting. The GRAPH_EXPAND table function flattens the graph into a table, and the AGG function queries the measures there. Labels and properties of a graph can also carry descriptions and synonyms.
  • 22 April, Preview. Model gemini-embedding-2-preview. Works in AI.EMBED, AI.SIMILARITY and AI.GENERATE_EMBEDDING and builds one embedding from a mix of text, image, audio, video and PDF input.
  • 20 April, Preview. Python UDF options and cost column. CREATE FUNCTION takes a container_request_concurrency option that caps concurrent requests per Python UDF container instance. Python UDF cost appears in the external_service_costs column of INFORMATION_SCHEMA.JOBS. Quotas apply: 10 GiB image storage and 30 mutations per minute, per project per region.
  • 16 April, Preview. AI.KEY_DRIVERS. Identifies the segments of data that cause statistically significant changes in a summable metric.
  • 15 April, Preview. EXPORT DATA to AlloyDB. EXPORT DATA statements can reverse ETL BigQuery data into AlloyDB.
  • 15 April, Preview. Iceberg version 3 external tables. Apache Iceberg external tables support Iceberg version 3, including binary deletion vectors.
  • 13 April. AI.AGG turned off. The Preview of the function is disabled for now. Google says it is working to restore it.
  • 13 April, Preview. Optimized mode for AI.IF and AI.CLASSIFY. Optimized mode cuts LLM token use and query latency on large datasets.
  • 13 April, GA. AI.IF, AI.SCORE and AI.CLASSIFY. AI.IF filters or joins text and unstructured data on a condition written in natural language. AI.SCORE rates data to rank it. AI.CLASSIFY sorts it into categories you define.
  • 9 April, Preview. Stateful continuous queries. Continuous queries can keep information across rows or time intervals with JOINs and windowing aggregations.
  • 9 April, Preview. BigQuery Graph. A graph is created from existing tables that hold entities and relationships, and Graph Query Language (GQL) queries it.
  • 8 April, GA. @@session_id in routines and views. The system variable works in SQL user-defined functions, table functions and logical views.
  • 7 April, Preview. Model embeddinggemma-300m. A built-in text embedding model for AI.EMBED and AI.SIMILARITY. It generates embeddings on your BigQuery slots.
  • 6 April, Preview. AI.AGG. Aggregates unstructured input data by following an instruction in natural language.
  • 2 April, GA. CREATE CONNECTION, ALTER CONNECTION SET OPTIONS and DROP CONNECTION. DDL statements manage Cloud resource connections. GRANT and REVOKE also take the connection user type and the PROJECT resource type.
  • 2 April, GA. Column granularity for search indexes. CREATE SEARCH INDEX can store extra column information in the index to speed up search queries.

March 2026

  • 31 March, GA. ObjectRef updates. The ObjectRef functions run with direct or delegated access. OBJ.MAKE_REF fetches the latest Cloud Storage metadata into ref.details. OBJ.GET_READ_URL returns a STRUCT with a read URL and status columns.
  • 30 March, GA. Forecasting and anomaly detection updates. AI.DETECT_ANOMALIES and AI.EVALUATE take a custom context window of recent data points. AI.FORECAST takes the latest timestamp to forecast from. AI.EVALUATE also returns the mean absolute scaled error.
  • 30 March, GA. Materialized views over Spanner data. Non-incremental materialized views cache the results of queries on Spanner data.
  • 26 March, GA. EXPORT DATA to Spanner with a Cloud resource connection. The statement reverse ETLs BigQuery data into Spanner through the connection.
  • 23 March, GA. AI.EMBED and AI.SIMILARITY. AI.EMBED creates embeddings from text or images. AI.SIMILARITY computes the semantic similarity between pairs of text, pairs of images, or text and images.
  • 17 March, GA. Automatically deployed open models. BigQuery ML can deploy open models to Vertex AI endpoints, manage the Vertex AI resources, reserve capacity with Compute Engine reservations, and undeploy the model automatically or immediately.
  • 16 March, GA. Global default location. The default location applies when a request sets no location and none can be inferred. It can be set at the organization or project level.
  • 12 March. Advanced runtime is the default. Advanced runtime is turned on for all projects.
  • 6 March, GA. Embedding remote models. A remote model can be based on Vertex AI gemini-embedding-001, or on an open embedding model from Model Garden or Hugging Face deployed to Vertex AI. AI.GENERATE_EMBEDDING uses these models, and AI.EMBED can call the gemini-embedding-001 endpoint directly.
  • 5 March, Preview. Single-vector syntax for VECTOR_SEARCH. An alternate call syntax speeds up a search for one vector.
  • 4 March, GA. Continuous queries to Spanner. A continuous query can stream BigQuery data to Spanner as it arrives.

February 2026

  • 25 February. Legacy SQL limited from 1 June 2026. An organization or project that does not use legacy SQL between 1 November 2025 and 1 June 2026 cannot use it afterwards. Those that do use it keep their existing workloads, and new ones might not run.
  • 23 February, GA. Undelete a dataset. A dataset inside its time travel window can be restored to the state it had when it was deleted.
  • 17 February, Preview. Global queries. One query can reference data stored in more than one region.
  • 12 February, GA. Field descriptions in AI.GENERATE and AI.GENERATE_TABLE. A custom output schema can carry a description for each field.
  • 12 February, Preview. Multiple categories in AI.CLASSIFY. The function can put an input into more than one category.
  • 4 February, GA. Data policies on columns. A data policy can be attached directly to a column to control access and apply masking or transformation rules.

January 2026

  • 29 January. RANDOM_HASH masking rule. A predefined masking rule that returns a salted hash of the column value. Google describes it as stronger than the Hash (SHA-256) rule.
  • 23 January. endpoint argument for AI.IF, AI.SCORE and AI.CLASSIFY. An optional endpoint argument picks the model.
  • 22 January. End-user credentials for AI.IF, AI.SCORE and AI.CLASSIFY. The functions can run with your end-user credentials instead of a BigQuery connection.
  • 22 January. Table parameters in table-valued functions. Support was restored.
  • 6 January, GA. Options for CREATE EXTERNAL TABLE and LOAD DATA. time_zone sets the time zone used when loading. date_format, datetime_format, time_format and timestamp_format define how date and time values look in the source files. null_markers sets the strings that mean NULL in CSV files. source_column_match matches loaded columns to the schema by position or by name.

2025

  • 2 December 2025, Preview. Autonomous embedding generation. A table made with CREATE TABLE can keep a column of embeddings based on a source column. BigQuery generates or updates the embeddings when the source column changes. AI.SEARCH runs semantic search on such tables.
  • 19 November 2025, Preview. JSON_FLATTEN. Extracts all non-array values that sit directly in a JSON value or in arrays nested inside it.
  • 18 November 2025. Gemini 3.0 in generative AI functions. Functions such as AI.GENERATE accept gemini-3-pro-preview. The call needs the full global endpoint argument, https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/google/models/gemini-3-pro-preview.
  • 18 November 2025, GA. AI.GENERATE. Generates free text from text, images, audio, video or documents, and can extract entities or return structured output. AI.GENERATE_BOOL, AI.GENERATE_DOUBLE and AI.GENERATE_INT return scalar values for filtering, scoring and counting. These functions support authentication with end-user credentials.
  • 18 November 2025, Preview. AI.EMBED and AI.SIMILARITY. AI.EMBED turns text, images, audio, video or documents into embeddings. AI.SIMILARITY computes semantic similarity between pairs of text, pairs of images, or text and images.
  • 18 November 2025, GA. AI.GENERATE_TABLE, AI.GENERATE_TEXT and AI.GENERATE_EMBEDDING. AI.GENERATE_TABLE returns a table of structured output from unstructured data. AI.GENERATE_TEXT and AI.GENERATE_EMBEDDING are the preferred versions of ML.GENERATE_TEXT and ML.GENERATE_EMBEDDING, with the same functionality and simpler output column names.
  • 10 November 2025. TimesFM 2.5. The time series model works in AI.FORECAST, AI.EVALUATE and AI.DETECT_ANOMALIES, with better forecasting accuracy and lower latency.
  • 10 November 2025, Preview. AI.DETECT_ANOMALIES. Finds anomalies in time series data with a TimesFM model, using historical data as the baseline.
  • 10 November 2025, Preview. Partitioning for BigLake tables for Apache Iceberg in BigQuery.
  • 5 November 2025, GA. MATCH_RECOGNIZE. A clause that filters and aggregates matches across rows in a table.
  • 29 October 2025, Preview. Reservation groups. Reservations in a group share idle slots with each other before the slots go to other reservations in the project.
  • 28 October 2025, GA. job_principal_subject in INFORMATION_SCHEMA.SHARED_DATASET_USAGE. Holds the principal of the user who ran a job or query against a linked dataset, when subscriber email logging is on for the listing or data exchange.
  • 27 October 2025, Preview. AI.IF, AI.SCORE and AI.CLASSIFY. Managed AI functions that use Gemini. AI.IF filters and joins text or multimodal data by a condition written in natural language. AI.SCORE rates input to rank data. AI.CLASSIFY sorts text into categories you define.
  • 22 October 2025, GA. Built-in TimesFM model. AI.FORECAST and AI.EVALUATE can use the built-in TimesFM univariate model. AI.FORECAST supports a larger context window.
  • 22 October 2025, Issue. Table parameters in table-valued functions. Support was temporarily disabled.
  • 20 October 2025, Preview. BigQuery-managed open models. Open models run as Vertex AI endpoints that BigQuery manages. SQL queries manage the Vertex AI resources, and the models can be undeployed automatically or immediately.
  • 9 October 2025, GA. Maximum slot limit for a reservation. The maximum size is set when the reservation is created or updated.
  • 9 October 2025, GA. Fair idle slot allocation. Idle slots are shared about equally across the reservations in one admin project.
  • 8 October 2025, Breaking change. Default QueryUsagePerDay limit. The default limit for on-demand pricing is 200 TiB for all new projects. Existing projects got a default based on their usage over the last 30 days. Projects with custom cost controls or reservations are not affected.
  • 8 October 2025, GA. Reservation chosen at query time. A query can name the reservation it uses at runtime, and IAM policies can be set directly on reservations.
  • 8 October 2025, GA. Labels on reservations. Labels organize reservations and help with billing analysis.
  • 6 October 2025, GA. Columns added to INFORMATION_SCHEMA.SHARED_DATASET_USAGE. shared_resource_id and shared_resource_type describe the queried resource. referenced_tables holds project_id, dataset_id, table_id and processed_bytes of the base table. They report usage of external tables and routines.
  • 29 September 2025, Preview. Non-incremental materialized views over Spanner data. The view caches results periodically to speed up queries.
  • 29 September 2025, Announcement. History-based query optimizations. They are on by default. A project or organization that turned them off can turn them back on.
  • 25 September 2025, GA. ARRAY_FIRST, ARRAY_LAST and ARRAY_SLICE.
  • 22 September 2025, GA. Federated queries to Spanner PostgreSQL dialect databases. They run through BigQuery external datasets with GoogleSQL, including cross-region federated queries.
  • 16 September 2025, GA. FOR SYSTEM_TIME AS OF on Iceberg external tables. The clause reads snapshots that are kept in the Iceberg metadata.
  • 16 September 2025, GA. JSON_KEYS and JSONPath modes. JSON_KEYS extracts the unique keys of a JSON expression. Some JSON functions that take a JSONPath accept a mode that sets how the path matches the JSON structure.
  • 15 September 2025, GA. BigQuery advanced runtime. It improves query execution time and slot usage. Between 15 September 2025 and early 2026 it becomes the default runtime for all projects.
  • 9 September 2025. Supervised tuning of gemini-2.5-pro and gemini-2.5-flash-lite remote models. It applies to BigQuery ML remote models based on those Vertex AI models.
  • 8 September 2025, Preview. Table parameters in table-valued functions. A TVF can take a table as a parameter.
  • 3 September 2025, GA. per_second_details in INFORMATION_SCHEMA.RESERVATIONS_TIMELINE. Reports reservation capacity and usage per second, including autoscale utilization.
  • 2 September 2025, Preview. gemini-embedding-001 remote model. ML.GENERATE_EMBEDDING generates embeddings with a remote model based on it.
  • 2 September 2025, Preview. Remote models on open embedding models. The model comes from Vertex Model Garden or Hugging Face and is deployed to Vertex AI, such as E5 Embedding. ML.GENERATE_EMBEDDING generates embeddings with it.
  • 25 August 2025, GA. ST_REGIONSTATS. Combines raster data from Earth Engine with vector data stored in BigQuery.
  • 22 August 2025, Preview. Multi-statement transactions on BigLake Iceberg tables in BigQuery.
  • 14 August 2025, GA. Cross region federated queries to Spanner. They query Spanner tables from regions other than the BigQuery region and incur Spanner network egress charges.
  • 11 August 2025, GA. WITH expressions. Create temporary variables inside a GoogleSQL expression.
  • 11 August 2025, GA. Chained function call syntax. Makes deeply nested function calls easier to read.
  • 6 August 2025, Preview. Short query optimizations. Turning on the advanced runtime includes them.
  • 31 July 2025. Higher throughput for generative AI functions. Throughput of ML.GENERATE_TEXT, AI.GENERATE_TABLE, AI.GENERATE, AI.GENERATE_BOOL, AI.GENERATE_DOUBLE and AI.GENERATE_INT improved by more than 100x. A typical 6-hour job can process millions of rows.
  • 31 July 2025, Preview. Continuous queries export to Spanner. A continuous query can export BigQuery data to Spanner in real time.
  • 28 July 2025, Preview. Data policies on columns. A data policy can be attached directly to a column to control access and apply masking and transformation rules.
  • 22 July 2025, Preview. VECTOR_INDEX.STATISTICS. Calculates how far an indexed table’s data has drifted since the vector index was created. ALTER VECTOR INDEX REBUILD rebuilds the index when the drift is large enough.
  • 22 July 2025, Preview. null_markers and source_column_match options. CREATE EXTERNAL TABLE and LOAD DATA accept both. null_markers defines the strings that stand for NULL in CSV files. source_column_match matches loaded columns to the schema by position or by name.
  • 22 July 2025, Preview. MATCH_RECOGNIZE. A clause that filters and aggregates matches across rows in a table.
  • 21 July 2025, GA. DISTINCT pipe operator. Selects distinct rows in pipe syntax queries.
  • 17 July 2025, GA. WITH pipe operator and named windows. The WITH pipe operator defines common table expressions in pipe syntax queries. Named windows also work in pipe syntax queries.
  • 8 July 2025, Announcement. GoogleSQL as the default dialect for the CLI and API. From 1 August 2025 queries run from the command line or the API use GoogleSQL unless the request asks for legacy SQL. Legacy SQL can also be chosen with the default_sql_dialect_option setting set to 'default_legacy_sql' at project or organization level.
  • 1 July 2025, GA. @@location. The system variable sets the location in which a query runs.
  • 26 June 2025, Preview. PARTITION BY in CREATE VECTOR INDEX. Partitions TreeAH vector indexes. Queries can then prune partitions, which can lower I/O costs.
  • 26 June 2025, GA. INFORMATION_SCHEMA.SEARCH_INDEXES_BY_ORGANIZATION. Shows how much of the free search index management limit the organization has used, by project and table.
  • 16 June 2025, GA. TIME_SERIES_ID_COL for ARIMA_PLUS_XREG. The option forecasts several time series at once with a multivariate model.
  • 16 June 2025, GA. IAM tags in SQL. IAM tags on datasets and tables can be managed with SQL.
  • 11 June 2025, Preview. ARRAY_FIRST, ARRAY_LAST and ARRAY_SLICE. ARRAY_FIRST returns the first element of an array and ARRAY_LAST the last. ARRAY_SLICE returns consecutive elements.
  • 10 June 2025, GA. Provisioned Throughput for ML.GENERATE_TEXT and AI.GENERATE. Vertex AI Provisioned Throughput gives these functions consistent high throughput on supported Gemini models.
  • 9 June 2025, GA. Iceberg external tables in materialized views. A materialized view can reference them, so the data does not have to move into BigQuery storage.
  • 4 June 2025. Organization-level default_sql_dialect_option and query_runtime. These configuration settings are no longer supported at organization level.
  • 3 June 2025, Preview. BigQuery advanced runtime. It improves query execution time and slot usage.
  • 2 June 2025, Preview. Date and time options for CREATE EXTERNAL TABLE and LOAD DATA. time_zone sets the time zone used when loading data. date_format, datetime_format, time_format and timestamp_format define how values are formatted in the source files.
  • 27 May 2025, GA. Optional job creation mode. Speeds up small queries in dashboards and data exploration. Eligible queries are optimized automatically and use a cache to cut latency.
  • 22 May 2025. INFORMATION_SCHEMA.OBJECT_PRIVILEGES permission. From 17 March 2026 viewing a dataset’s access controls and querying this view needs bigquery.datasets.getIamPolicy.
  • 21 May 2025. Supervised tuning of gemini-2.0-flash-001 and gemini-2.0-flash-lite-001 remote models. It applies to BigQuery ML remote models based on those Vertex AI models.
  • 19 May 2025, GA. Continuous queries. Long-lived SQL statements that analyze and process incoming data, and run ML inference on it, in real time. They can use slot autoscaling and a custom job ID prefix.
  • 19 May 2025, Preview. Cross regional federated queries to Spanner. Spanner tables in regions other than the BigQuery region can be queried. There are no Spanner network egress charges during the preview.
  • 14 May 2025, GA. TreeAH vector index. Vector indexes support the TreeAH index type, which uses Google’s ScaNN algorithm. It suits batch processing of a few thousand to hundreds of thousands of embeddings at once.
  • 14 May 2025, GA. Cross-region load and export. A single LOAD DATA or EXPORT DATA statement can move data between any two regions or multi-regions.
  • 13 May 2025, GA. GROUP BY STRUCT, GROUP BY ARRAY and GROUP BY ALL. STRUCT and ARRAY values can be grouped and used with SELECT DISTINCT. GROUP BY ALL groups by all non-aggregated select items.
  • 12 May 2025, Preview. Multimodal data with ObjectRef. ObjectRef values bring unstructured data into standard tables, and ObjectRefRuntime values work with it in queries. BigQuery ML SQL functions with Gemini and multimodal embedding models analyze it. Python UDFs can transform images and chunk PDF files.
  • 23 April 2025, Preview. Reservation features. A query can name the reservation it uses at runtime, and IAM policies can be set on reservations. A reservation can have a maximum slot limit. Idle slots can be allocated fairly across reservations in one admin project.
  • 21 April 2025, Announcement. Spend-based committed use discounts. A commitment to spend a minimum amount per hour on pay-as-you-go compute earns a discount. The commitment period is one or three years.
  • 9 April 2025, Preview. ST_REGIONSTATS. Combines raster and vector data for geospatial analysis.
  • 8 April 2025, Preview. Built-in TimesFM model. AI.FORECAST can forecast with the built-in TimesFM univariate model, with no model to create or train.
  • 7 April 2025, GA. Llama and Mistral AI remote models. ML.GENERATE_TEXT works with remote models based on Llama and Mistral AI models in Vertex AI.
  • 7 April 2025, GA. Smart-tuning for materialized views. It applies when the view is in the same project as one of its base tables, or in the project running the query.
  • 7 April 2025. Token-based batching for embedding generation. BigQuery ML puts as many rows as possible into one embedding request, with an average 10x improvement against query-per-minute quotas.
  • 4 April 2025, Preview. AI.GENERATE, AI.GENERATE_BOOL, AI.GENERATE_INT and AI.GENERATE_DOUBLE. Analyze text with a Vertex AI Gemini model. The output matches the type in the function name.
  • 3 April 2025, Preview. AI.GENERATE_TABLE and output_schema. Returns structured data from Gemini 1.5 Pro, Gemini 1.5 Flash and Gemini 2.0 Flash. The output_schema argument takes a SQL schema like the one in CREATE TABLE.
  • 2 April 2025, Preview. Python UDFs. User-defined functions written in Python can use additional libraries and external APIs.
  • 1 April 2025, GA. Contribution analysis models. CREATE MODEL builds a contribution analysis model, with the top_k_insights_by_apriori_support and pruning_method options. ML.GET_INSIGHTS reports changes in key metrics for summable, summable ratio and summable by category metrics.
  • 1 April 2025, GA. Pipe syntax. A linear query structure meant to make queries easier to read, write and maintain.
  • 31 March 2025, GA. Merge-on-read for Iceberg external tables. Queries can read Iceberg tables with position deletes and equality deletes.
  • 31 March 2025, GA. Subqueries in row-level access policies. They work with BigLake managed tables and the BigQuery Storage Read API too.
  • 31 March 2025, GA. _CHANGE_SEQUENCE_NUMBER. Sets the order of streaming UPSERT changes in BigQuery change data capture.
  • 26 March 2025, Preview. Column granularity for search indexes. Set when the search index is created, it stores more column information in the index to speed up search queries.
  • 24 March 2025, Preview. KLL quantile functions. Compute approximate quantiles efficiently.
  • 24 March 2025, Preview. Labels on reservations. Labels organize reservations and help with billing analysis.
  • 20 March 2025, GA. Anthropic Claude remote models. ML.GENERATE_TEXT works with remote models based on Claude in Vertex AI, and ML.EVALUATE can evaluate them.
  • 17 March 2025, GA. EXPORT DATA to Spanner. Reverse ETL from BigQuery to Spanner.
  • 17 March 2025, GA. Spanner external datasets. An external dataset in BigQuery links to an existing Spanner database.
  • 17 March 2025, GA. TYPEOF. Returns the data type of an expression.
  • 3 March 2025, GA. SQL CREATE AGGREGATE FUNCTION. Creates a SQL user-defined aggregate function.
  • 24 February 2025, Preview. @@location. The system variable sets the location in which a query runs.
  • 6 February 2025, GA. JavaScript CREATE AGGREGATE FUNCTION. Creates a JavaScript user-defined aggregate function.
  • 3 February 2025, GA. BY NAME and CORRESPONDING. Set operations match columns by name instead of by position.
  • 3 February 2025. output_format for ML.BUCKETIZE and ML.QUANTILE_BUCKETIZE. The argument formats the output as a string like bin_<bucket_index>, as a string in interval notation, or as a JSON-formatted string.
  • 27 January 2025, GA. Remote models on open models. The model comes from Vertex Model Garden or Hugging Face and is deployed to Vertex AI, such as Llama and Gemma. ML.GENERATE_TEXT and ML.EVALUATE work with these remote models.
  • 21 January 2025, Preview. ML.EVALUATE for Claude models. Evaluates Anthropic Claude remote models. The Claude quotas in BigQuery ML match the Vertex AI quotas.
  • 13 January 2025, Preview. TIME_SERIES_ID_COL for ARIMA_PLUS_XREG. The option forecasts several time series at once with a multivariate model.

2024

  • 11 December 2024, Preview. Model gemini-2.0-flash-exp. A remote model can be based on this Vertex AI model. ML.GENERATE_TEXT runs it on text in tables and on image, video, audio, PDF or text content in object tables.
  • 11 December 2024, GA. Cross-region dataset replication. A dataset can be replicated from its source region to one or more other regions.
  • 11 December 2024, GA. Managed disaster recovery. Gives managed failover and redundant compute capacity for a total region outage. It needs the Enterprise Plus edition.
  • 19 November 2024, GA. Search indexes on INT64 and TIMESTAMP columns. CREATE SEARCH INDEX accepts these columns, and BigQuery can optimize predicates that use them.
  • 14 November 2024, GA. Models gemini-1.5-flash and gemini-1.5-pro. Remote models can be based on these Vertex AI models. ML.GENERATE_TEXT runs them on text in tables and on image, video, audio, PDF or text content in object tables.
  • 11 November 2024, GA. Tuning and evaluating LLMs. A remote model based on Gemini 1.5 flash or pro can be tuned with supervised tuning. ML.EVALUATE evaluates pre-trained PaLM and Gemini models and tuned Gemini models.
  • 14 October 2024, Preview. Fine-grained DML. Optimizes the execution of UPDATE, DELETE and MERGE statements on tables.
  • 10 October 2024, Preview. Tables for Apache Iceberg. BigQuery storage optimization applies to Iceberg tables that stay in your own Cloud Storage buckets, so the data does not move.
  • 8 October 2024, Preview. Pipe syntax. A linear query structure that can be used anywhere GoogleSQL is written.
  • 3 October 2024, Preview. Spanner external datasets. An external dataset links BigQuery to an existing Spanner database.
  • 30 September 2024, GA. History-based optimizations. They can be turned on and off, and their effect on queries can be analyzed.
  • 30 September 2024, GA. Flexible column names. Tables and views accept more characters in column names for extracting, loading, streaming and querying.
  • 24 September 2024, GA. ML.GENERATE_EMBEDDING for more models. It works with a remote model based on Vertex AI multimodalembedding to embed text, images and video in one space. It also works with PCA, autoencoder and matrix factorization models.
  • 24 September 2024, GA. ML.PROCESS_DOCUMENT and ML.TRANSCRIBE. ML.PROCESS_DOCUMENT processes documents in object tables with a remote model based on Document AI. ML.TRANSCRIBE transcribes audio files in object tables with a remote model based on Speech-to-Text.
  • 19 September 2024, GA. Model monitoring functions. ML.DESCRIBE_DATA, ML.VALIDATE_DATA_SKEW, ML.VALIDATE_DATA_DRIFT, ML.TFDV_DESCRIBE and ML.TFDV_VALIDATE compute statistics for training and serving data and compare them to find anomalous differences.
  • 16 September 2024, Preview. Contribution analysis models. CREATE MODEL creates a contribution analysis model. ML.GET_INSIGHTS uses it to report insights about changes to key metrics in multi-dimensional data.
  • 16 September 2024, GA. Stored columns in vector indexes. Columns stored in a vector index can be used to pre-filter VECTOR_SEARCH.
  • 4 September 2024, GA. VECTOR_SEARCH and vector indexes. VECTOR_SEARCH searches embeddings for semantically similar entities. A vector index makes it more efficient and returns more approximate results.
  • 28 August 2024, GA. Phrase search in SEARCH. SEARCH supports phrase matching.
  • 28 August 2024, Preview. GROUP BY and SELECT DISTINCT on ARRAY and STRUCT. Both clauses accept columns of these types.
  • 28 August 2024, GA. AlloyDB federated queries. A federated query can read data in AlloyDB.
  • 26 August 2024, Preview. Anthropic Claude models. A remote model can be based on a Claude model in Vertex AI. ML.GENERATE_TEXT runs it on text in tables.
  • 26 August 2024, GA. EXPORT DATA to Bigtable. EXPORT DATA writes BigQuery data directly to Bigtable.
  • 20 August 2024, GA. Anomaly detection with ARIMA_PLUS_XREG. Multivariate time series models can detect anomalies in historical data or in new data with several feature columns.
  • 19 August 2024, Preview. INSIGHTS, RECOMMENDATIONS and RECOMMENDATIONS_BY_ORGANIZATION. These INFORMATION_SCHEMA views list BigQuery insights and recommendations.
  • 14 August 2024, Preview. Short query optimized mode. Lowers latency for small queries. BigQuery decides which queries to accelerate, and the rest run as before.
  • 12 August 2024, GA. Time series and range functions. They support time series analysis.
  • 8 August 2024, Preview. JSON_KEYS. Extracts the unique keys from a JSON expression.
  • 8 August 2024, Preview. JSONPath modes. Some JSON functions that take a JSONPath accept a mode that sets how strictly the path matches the JSON structure.
  • 7 August 2024, Preview. Materialized views over Iceberg tables. The view can be partition aligned with the base table. Only time-based partition transforms are supported: YEAR, MONTH, DAY and HOUR.
  • 31 July 2024, GA. Autoscaler changes. The autoscaler scales up immediately and to the nearest multiple of 50 slots instead of 100. Capacity commitments, baseline slots and autoscale max slots are set in steps of 50. After one minute or more since the last increase, capacity can be reduced several times in a row without waiting a minute between decreases.
  • 30 July 2024, Preview. output_dimensionality for ML.GENERATE_EMBEDDING. Sets the number of dimensions of the embeddings when the function uses a remote model based on Vertex AI multimodalembedding.
  • 29 July 2024, GA. RANGE as a JSON encoding. The RANGE type can be encoded to JSON.
  • 29 July 2024, Preview. TreeAH vector index. A vector index can use the TreeAH type, which uses Google’s ScaNN algorithm. It suits batch queries with hundreds or more query vectors.
  • 22 July 2024, Preview. Continuous queries. Long-lived SQL statements that process incoming data in BigQuery in real time and can run ML inference on it. They can write results to a Pub/Sub topic, a Bigtable instance or another BigQuery table.
  • 22 July 2024, GA. DML on recently streamed rows. UPDATE, DELETE and MERGE can modify rows that the Storage Write API wrote recently.
  • 22 July 2024, Preview. CHANGES. A table function that returns the history of changes to a table over a window of time. It covers CREATE TABLE, INSERT, MERGE, UPDATE, DELETE and TRUNCATE TABLE, loads, streaming ingestion, WRITE_TRUNCATE jobs and partition deletions.
  • 17 July 2024, GA. Default storage billing model. The default storage billing model for new datasets can be configured.
  • 11 July 2024, Preview. EXPORT DATA to Spanner. EXPORT DATA writes BigQuery data to Spanner.
  • 18 June 2024, GA. Collation in NULLIF. NULLIF follows collation and can be used in collation-aware comparisons with STRUCT.
  • 17 June 2024, Preview. Tuning gemini-1.0-pro-002. A remote model based on this model supports supervised tuning.
  • 5 June 2024, GA. Slot recommender. Analyzes historical usage and recommends capacity to buy for edition and on-demand workloads.
  • 5 June 2024. output_dimensionality for text-embedding models. ML.GENERATE_EMBEDDING accepts it for text-embedding and text-multilingual-embedding models to set the number of dimensions.
  • 29 May 2024. Partition limit raised. A partitioned table can have 10,000 partitions, up from 4,000.
  • 28 May 2024, Preview. Model gemini-1.5-flash. A remote model can be based on this Vertex AI model. ML.GENERATE_TEXT runs it on text in tables and on image, video, audio, PDF or text content in object tables.
  • 23 May 2024. FORECAST_LIMIT_LOWER_BOUND and FORECAST_LIMIT_UPPER_BOUND with TIME_SERIES_ID_COL. These options of univariate time series models work together with TIME_SERIES_ID_COL.
  • 23 May 2024. Grounding and safety settings in ML.GENERATE_TEXT. For Vertex AI Gemini models, ground_with_google_search lets the model use information from the internet. safety_settings sets the attributes the model uses to filter responses.
  • 23 May 2024, Preview. Video embeddings. ML.GENERATE_EMBEDDING creates multimodal embeddings that include video with a remote model based on Vertex AI multimodalembedding.
  • 22 May 2024, Preview. AlloyDB federated queries. A federated query can read data in AlloyDB.
  • 21 May 2024, Preview. Model gemini-1.5-pro. A remote model can be based on this Vertex AI model. ML.GENERATE_TEXT runs it on text in tables and on image, video, audio, PDF or text content in object tables.
  • 20 May 2024, GA. DLP functions. They encrypt and decrypt data between BigQuery and Sensitive Data Protection using AES-SIV.
  • 20 May 2024, Preview. Search indexes on INT64 and TIMESTAMP columns. A search index can optimize lookups on these types.
  • 13 May 2024, Preview. Phrase search in SEARCH. SEARCH supports phrase matching.
  • 7 May 2024, Preview. JavaScript user-defined aggregate functions. CREATE AGGREGATE FUNCTION creates a JavaScript UDAF.
  • 7 May 2024, Preview. Stored columns in vector indexes. Columns stored in a vector index can be used to pre-filter VECTOR_SEARCH.
  • 6 May 2024, Preview. Managed disaster recovery. Gives managed failover and redundant compute capacity for a total region outage. It needs the Enterprise Plus edition.
  • 30 April 2024, GA. ML.TRANSFORM and transform-only models. ML.TRANSFORM applies the transformations captured in the TRANSFORM clause of an existing model to input data. A transform-only model applies preprocessing functions and returns the processed data, so preprocessing is separate from training.
  • 30 April 2024, Preview. Iceberg tables in materialized views. A materialized view can reference Iceberg tables instead of data moved into BigQuery storage.
  • 24 April 2024, Preview. SQL user-defined aggregate functions. CREATE AGGREGATE FUNCTION creates a UDAF written with SQL expressions.
  • 18 April 2024, GA. Quantified LIKE. LIKE ANY matches when at least one pattern matches, and LIKE SOME is a synonym. LIKE ALL matches when every pattern matches. The patterns can be a list or an array.
  • 16 April 2024, Preview. Subqueries in row-level access policies. A row-level access policy can use a subquery.
  • 9 April 2024, Preview. More ML.GENERATE_EMBEDDING models. It works with a remote model based on Vertex AI multimodalembedding to embed text and images in one space. It also works with PCA, autoencoder and matrix factorization models.
  • 5 April 2024, GA. Differential privacy.
  • 4 April 2024, GA. allow_non_incremental_definition and max_staleness for materialized views. The first option allows a wider range of queries in a materialized view. The second keeps performance steady and cost controlled on large tables that change often.
  • 4 April 2024, Preview. Model monitoring functions. ML.DESCRIBE_DATA, ML.VALIDATE_DATA_SKEW, ML.VALIDATE_DATA_DRIFT, ML.TFDV_DESCRIBE and ML.TFDV_VALIDATE compute statistics for training and serving data and compare them to find anomalous differences.
  • 3 April 2024. Materialized views on CDC tables. A materialized view can be built on a table that has change data capture enabled for streaming updates and deletes.
  • 3 April 2024, GA. More collation support. Collation applies to the underscore in LIKE. STRUCT values can be compared with =, !=, IN and CASE under collation.
  • 2 April 2024, Preview. Tuning and evaluating LLMs. A remote model based on a text-bison model supports supervised tuning. ML.EVALUATE evaluates pre-trained text-bison, text unicorn and gemini-pro models and tuned text-bison models.
  • 1 April 2024, Preview. History-based optimizations. They can be turned on and off, and their effect on queries can be analyzed.
  • 28 March 2024, Preview. SAP Datasphere federated queries. A federated query can read data in SAP Datasphere.
  • 28 March 2024, GA. Search index optimization for string literals. Queries are optimized when they compare string literals with indexed data using =, IN, LIKE or STARTS_WITH.
  • 26 March 2024, Preview. Model gemini-pro-vision. A remote model can be based on this Vertex AI model. ML.GENERATE_TEXT runs it on visual content in object tables, for tasks such as image or video captioning and visual Q&A.
  • 21 March 2024, Preview. LEFT OUTER JOIN and UNION ALL in incremental materialized views. Incremental materialized views accept both.
  • 19 March 2024, GA. Spark stored procedures. Stored procedures can be written in Python, Java and Scala and run on Apache Spark.
  • 18 March 2024, Preview. Undelete a dataset. A dataset within the time travel window can be recovered to the state it had when it was deleted.
  • 18 March 2024, GA. Text analysis and distance functions. CREATE SEARCH INDEX takes text analysis options. The LOG_ANALYZER and PATTERN_ANALYZER analyzers work in functions such as SEARCH. TEXT_ANALYZE, ML.BAG_OF_WORDS, ML.TF_IDF, BAG_OF_WORDS, TF_IDF, COSINE_DISTANCE, EUCLIDEAN_DISTANCE and EDIT_DISTANCE are also generally available.
  • 18 March 2024, GA. Hierarchical forecasts. Time series models can forecast hierarchically, aggregating and rolling up values for all series in the model.
  • 6 March 2024, GA. INFORMATION_SCHEMA.WRITE_API_TIMELINE*. These views hold per-minute aggregated statistics for Storage Write API ingestion.
  • 4 March 2024, GA. BQ.REFRESH_EXTERNAL_METADATA_CACHE. A system procedure that refreshes the metadata cache of selected BigLake tables.
  • 29 February 2024, GA. Cross-cloud materialized views and joins. Materialized views can sit over Amazon S3 metadata cache-enabled BigLake tables, and replicas of those views avoid data egress costs. Information about the replicas is available through SQL. Cross-cloud joins run queries that span Google Cloud and BigQuery Omni regions.
  • 28 February 2024, Preview. Materialized views over logical views. A materialized view can reference logical views.
  • 28 February 2024, GA. Filtering Amazon S3 and Azure Blob Storage data before loading. CREATE TABLE AS SELECT, CREATE TABLE IF NOT EXISTS AS SELECT, CREATE OR REPLACE TABLE AS SELECT and INSERT INTO SELECT can filter files before the results go into BigQuery tables. Billing is enabled.
  • 28 February 2024, Preview. Anomaly detection with ARIMA_PLUS_XREG. Multivariate time series models can detect anomalies in historical data or in new data with several feature columns.
  • 27 February 2024, GA. DELETE of whole partitions. A DELETE that targets all rows in a partition deletes the partition without scanning bytes or using slots.
  • 27 February 2024, Preview. Time series and range functions. They support time series analysis.
  • 26 February 2024, Preview. GROUP BY ALL. Groups rows by grouping keys inferred from the SELECT items.
  • 26 February 2024, GA. GROUP BY GROUPING SETS, GROUP BY CUBE and GROUPING. GROUPING SETS aggregates over chosen grouping sets, CUBE over all permutations of the grouping sets. GROUPING reports whether a value in GROUP BY is aggregated.
  • 22 February 2024, GA. Text embeddings. A remote model can reference a Vertex AI textembedding-gecko* model, and ML.GENERATE_EMBEDDING embeds text with it. ML.GENERATE_EMBEDDING also works with the NNLM, SWIVEL and BERT TensorFlow models.
  • 15 February 2024, GA. Model gemini-pro. A remote model can be based on this Vertex AI model. ML.GENERATE_TEXT runs it on text in tables.
  • 8 February 2024, GA. Custom data masking. Custom masking routines, such as a salted hash, can be defined. It needs the Enterprise Plus edition.
  • 6 February 2024, Announcement. Spark stored procedure billing. Billing begins on 12 March 2024. Until then Spark stored procedures cost nothing extra.
  • 31 January 2024, Preview. ORGANIZATION_OPTIONS_CHANGES and PROJECT_OPTIONS_CHANGES. These INFORMATION_SCHEMA views show the history of configuration changes to the options of an organization and of a project.
  • 31 January 2024, Preview. VECTOR_SEARCH and vector indexes. VECTOR_SEARCH searches embeddings for semantically similar entities. A vector index makes it more efficient and returns more approximate results.
  • 24 January 2024. APPROX_GLOBAL_FEATURE_CONTRIB default. For boosted tree and random forest models it defaults to TRUE when ENABLE_GLOBAL_EXPLAIN is TRUE and NUM_PARALLEL_TREE is above 10 (boosted tree) or 50 (random forest). This speeds up training.
  • 24 January 2024. residual column in ML.EXPLAIN_FORECAST. For ARIMA_PLUS and ARIMA_PLUS_XREG models, it holds the difference between the actual and the fitted series for the historical data.
  • 24 January 2024, GA. ST_LINEINTERPOLATEPOINT. Returns the point at a given fraction along a linestring.
  • 16 January 2024, Preview. Cross-cloud joins. Queries can span Google Cloud and BigQuery Omni regions.

2023

  • 12 December 2023, Preview. Cross-cloud materialized views. Materialized views can sit over Amazon S3 metadata cache-enabled BigLake tables. Replicas of those views avoid data egress costs.
  • 7 December 2023, Preview. ML.TRANSFORM and transform-only models. ML.TRANSFORM applies the transformations captured in the TRANSFORM clause of an existing model to input data. A transform-only model applies preprocessing functions and returns the processed data.
  • 30 November 2023, Preview. Slot estimator for on-demand workloads. It gives project-level recommendations for cost-optimal commitments and autoscaling.
  • 16 November 2023, GA. Remote model syntax for LLMs. CREATE MODEL for remote models reaches all text generation and text embedding LLMs, such as text-bison-32k and textembedding-gecko-multilingual, and supports different LLM versions.
  • 13 November 2023, GA. ML.FEATURES_AT_TIME and ML.ENTITY_FEATURES_AT_TIME. They fetch features as of a point in time to avoid data leakage. ML.FEATURES_AT_TIME uses one cutoff for all entities. ML.ENTITY_FEATURES_AT_TIME reads features from several points in time for several entities.
  • 13 November 2023, Preview. ML.PROCESS_DOCUMENT and ML.TRANSCRIBE. ML.PROCESS_DOCUMENT processes documents in object tables with a remote model based on Document AI. ML.TRANSCRIBE transcribes audio files in object tables with a remote model based on Speech-to-Text.
  • 6 November 2023, GA. Forecast limits and custom holidays. FORECAST_LIMIT_LOWER_BOUND and FORECAST_LIMIT_UPPER_BOUND in CREATE MODEL bound the forecasted values. CREATE MODEL can also specify custom holidays for time series models. ML.HOLIDAY_INFO lists the holidays an ARIMA_PLUS or ARIMA_PLUS_XREG model uses, and ML.EXPLAIN_FORECAST explains the effect of each holiday.
  • 2 November 2023, Preview. Text analysis and distance functions. CREATE SEARCH INDEX takes text analysis options. The LOG_ANALYZER and PATTERN_ANALYZER analyzers work in functions such as SEARCH. TEXT_ANALYZE, ML.BAG_OF_WORDS, ML.TF_IDF, BAG_OF_WORDS, TF_IDF, COSINE_DISTANCE, EUCLIDEAN_DISTANCE and EDIT_DISTANCE are added.
  • 1 November 2023, Preview. TABLE_STORAGE_USAGE_TIMELINE and TABLE_STORAGE_USAGE_TIMELINE_BY_ORGANIZATION. These INFORMATION_SCHEMA views give the total billable bytes per table per day, for a project and for an organization.
  • 1 November 2023, GA. Cached results shared between users. On the Enterprise and Enterprise Plus editions, a query can use the cached result of the same query run by another user in the same project.
  • 31 October 2023, Preview. DML on recently streamed rows. DML can modify rows that the Storage Write API wrote recently.
  • 23 October 2023, Preview. More functions in custom data masking. Custom masking routines can use more functions, including SHA hash functions with salt.
  • 19 October 2023, Preview. Spark stored procedures without enrollment. Stored procedures for Apache Spark no longer need enrollment.
  • 16 October 2023, Preview. DLP functions. They encrypt and decrypt data between BigQuery and DLP using AES-SIV.
  • 12 October 2023, GA. ST_LINESUBSTRING and ST_HAUSDORFFDISTANCE. ST_LINESUBSTRING returns a segment of a linestring between a start and an end fraction. ST_HAUSDORFFDISTANCE returns the discrete Hausdorff distance between two geometries.
  • 9 October 2023, Preview. Grouping sets. GROUP BY GROUPING SETS and GROUP BY CUBE are added, and GROUP BY ROLLUP accepts sets of groupable items. The GROUPING function reports whether a value in GROUP BY is aggregated.
  • 9 October 2023, GA. Column descriptions on views. CREATE VIEW and ALTER COLUMN can set descriptions on the columns of a view.
  • 28 September 2023, Preview. ML.FEATURES_AT_TIME and ML.ENTITY_FEATURES_AT_TIME. They fetch features as of a point in time to avoid data leakage. ML.FEATURES_AT_TIME uses one cutoff for all entities. ML.ENTITY_FEATURES_AT_TIME reads features from several points in time for several entities.
  • 25 September 2023, GA. Authorized stored procedures. A stored procedure can be shared with users or groups who have no direct access to the underlying tables.
  • 25 September 2023, GA. Materialized views over BigLake metadata cache-enabled tables. They can reference structured data in Cloud Storage and work like materialized views over BigQuery-managed tables, including automatic refresh and smart tuning.
  • 12 September 2023, GA. Struct subscript operator. Accesses a STRUCT field by index, offset or ordinal.
  • 12 September 2023, GA. Array subscript by index. The array subscript operator returns a value directly by index. Before, only offset and ordinal were available.
  • 30 August 2023, Preview. EXPORT DATA to Bigtable. EXPORT DATA writes BigQuery data directly to Bigtable.
  • 29 August 2023, GA. Model text-bison. A remote model can be based on this Vertex AI LLM. ML.GENERATE_TEXT runs it on text in tables.
  • 29 August 2023, Preview. Data clean rooms. Several parties can share, join and analyze data without moving it or revealing the underlying data. The release also lists an aggregation threshold clause for queries and views.
  • 24 August 2023, Preview. Text embeddings. A remote model can reference the Vertex AI textembedding-gecko model, and ML.GENERATE_TEXT_EMBEDDING embeds text with it. ARRAY<NUMERIC> works as an input feature type for other models, and the NNLM, SWIVEL and BERT TensorFlow models can generate text embeddings.
  • 24 August 2023, GA. BigQuery ML inference. ONNX, XGBoost and TensorFlow Lite models can be imported and run in BigQuery ML. Models can be hosted remotely on Vertex AI Prediction. BigQuery ML functions can run inference on pretrained Vertex AI models for natural language, translation and computer vision tasks.
  • 22 August 2023, Preview. Custom masking routines. Masking rules written with REGEX_REPLACE can obfuscate sensitive data.
  • 17 August 2023, Preview. Cross-region dataset replication. A dataset can be replicated from its source region to one or more other regions.
  • 14 August 2023. Network egress pricing. From 15 September 2023, prices apply to network egress from a BigQuery Google Cloud region to another Google Cloud region, on the same continent and between continents.
  • 10 August 2023, GA. Protobuf export. User-defined functions can export BigQuery data as Protocol Buffer columns.
  • 8 August 2023, GA. ANY_VALUE with HAVING MAX and HAVING MIN, MAX_BY and MIN_BY. They work in queries and in materialized views. MAX_BY is a synonym for ANY_VALUE(x HAVING MAX y), and MIN_BY for ANY_VALUE(x HAVING MIN y).
  • 7 August 2023, GA. JSON functions. JSON_ARRAY, JSON_ARRAY_APPEND, JSON_ARRAY_INSERT, JSON_OBJECT, JSON_REMOVE, JSON_SET, JSON_STRIP_NULLS, LAX_BOOL, LAX_FLOAT64, LAX_INT64 and LAX_STRING.
  • 7 August 2023, Preview. Quantified LIKE. LIKE ANY matches when at least one pattern matches, and LIKE SOME is a synonym. LIKE ALL matches when every pattern matches.
  • 3 August 2023, GA. GRANT and REVOKE on materialized views. Access to materialized views can be granted and revoked with a SQL statement.
  • 28 July 2023, GA. Query queues. BigQuery sets query concurrency from available slots instead of a fixed limit and queues queries once the maximum is reached. Queues are on by default. The maximum concurrency target can be set for a reservation, and the queue timeout for interactive and batch queries comes from default configurations.
  • 20 July 2023, GA. ARIMA_PLUS_XREG. Multivariate time series forecasting with extra feature columns.
  • 20 July 2023. BigQuery ML explainability. ML.EXPLAIN_FORECAST works with ARIMA_PLUS_XREG models and explains the holiday effect for ARIMA_PLUS and ARIMA_PLUS_XREG. ML.GLOBAL_EXPLAIN works with AutoML Tables models. The approx_global_feature_contrib training option and the approx_feature_contrib option of ML.EXPLAIN_PREDICT approximate feature contribution for boosted tree and random forest models.
  • 19 July 2023, Preview. Search index optimization for string literals. Some queries are optimized when they compare string literals with indexed data using =, IN, LIKE or STARTS_WITH.
  • 17 July 2023, GA. Primary and foreign key constraints.
  • 12 July 2023, GA. TRANSFORM preprocessing functions. ML.IMPUTER, ML.LABEL_ENCODER, ML.MAX_ABS_SCALER, ML.MULTI_HOT_ENCODER, ML.NORMALIZER, ML.ONE_HOT_ENCODER and ML.ROBUST_SCALER preprocess features. Models that use the TRANSFORM clause can be exported to the TensorFlow SavedModel format.
  • 12 July 2023, Preview. Custom holiday modeling. CREATE MODEL can specify custom holidays for time series models. ML.HOLIDAY_INFO lists the holidays an ARIMA_PLUS or ARIMA_PLUS_XREG model uses, and ML.EXPLAIN_FORECAST explains the effect of each holiday. The public table bigquery-public-data.ml_datasets.holidays_and_events_for_forecasting lists the built-in holidays.
  • 5 July 2023, GA. Fail-safe period. Seven more days of storage after the time travel window keep data available for emergency recovery. Fail-safe bytes are not billed until 17 July 2023.
  • 5 July 2023, GA. Time travel window. It can be set from a minimum of two days to a maximum of seven.
  • 5 July 2023, GA. Physical bytes storage billing. When a dataset uses the physical bytes billing model, active storage cost includes the bytes used for time travel and fail-safe.
  • 5 July 2023, GA. LOAD DATA. Loads Avro, CSV, newline-delimited JSON, JSON, ORC or Parquet files into a table.
  • 5 July 2023, Preview. Slot estimator. Gives cost-optimal commitment and autoscale recommendations based on editions pricing and historical performance.
  • 5 July 2023, Announcement. Capacity commitments. Annual commitments are only available in the Enterprise or Enterprise Plus edition, and flat-rate annual commitments are gone. Monthly and flex commitments are no longer available.
  • 26 June 2023, Preview. Spark stored procedures in Java and Scala. Stored procedures for Apache Spark can be written in Java or Scala.
  • 21 June 2023, GA. TRUNCATE TABLE in transactions. Multi-statement transactions support TRUNCATE TABLE.
  • 20 June 2023, GA. Metadata caching. Cached metadata can speed up queries on BigLake tables and object tables that reference many objects, because the query skips listing objects in Cloud Storage. Metadata cache usage is billed from this date.
  • 20 June 2023, GA. Apache Iceberg tables. BigQuery queries Iceberg tables created by open source engines.
  • 15 June 2023, Preview. Model text-bison. A remote model can be based on this Vertex AI LLM, and ML.GENERATE_TEXT runs it on text in tables. Access needs an allowlist.
  • 14 June 2023, GA. TABLE_STORAGE and TABLE_STORAGE_BY_ORGANIZATION. These INFORMATION_SCHEMA views show a snapshot of current table storage usage for a project and for an organization.
  • 23 May 2023. DML and table modification limits. DML statements no longer count toward the daily limit on table and partitioned table modifications. The limit itself is unchanged.
  • 19 May 2023, GA. EXTERNAL_QUERY SQL pushdown. Column pruning from SELECT and filter pushdown from WHERE reduce the data fetched from sources such as Cloud SQL and Spanner. It applies to SELECT * FROM T queries, and not every data type supports filter pushdown.
  • 11 May 2023, GA. Object tables. Read-only tables of metadata for unstructured data in Cloud Storage. The release adds ML.DECODE_IMAGE, ML.CONVERT_COLOR_SPACE, ML.CONVERT_IMAGE_TYPE, ML.RESIZE_IMAGE, ML.DISTANCE and ML.LP_NORM.
  • 9 May 2023, GA. EXTERNAL_QUERY SQL pushdown. Column pruning from SELECT and filter pushdown from WHERE reduce the data fetched from sources such as Cloud SQL and Spanner. It applies to SELECT * FROM T queries, and not every data type supports filter pushdown.
  • 8 May 2023, Preview. Differential privacy. The differential privacy clause and four differentially private aggregate functions anonymize data: AVG, COUNT, SUM and PERCENTILE_CONT.
  • 8 May 2023, Preview. INFORMATION_SCHEMA.MATERIALIZED_VIEW. This view and enhanced job statistics show materialized view usage and refresh jobs.
  • 5 May 2023, Preview. INSERT INTO SELECT from Amazon S3 and Azure Blob Storage. The statement filters data from files in S3 and Blob Storage and appends it to BigQuery tables.
  • 3 May 2023, GA. Table clones.
  • 1 May 2023, Preview. Query queue timeouts. With query queues, the interactive and batch queue timeouts can be set in the default configuration.
  • 1 May 2023, Preview. Column descriptions on views. CREATE VIEW and ALTER COLUMN can set descriptions on the columns of a view.
  • 27 April 2023, GA. JSON mapping in Spanner federated queries.
  • 25 April 2023, GA. Dynamic data masking on repeated records. Masking works on RECORD columns in REPEATED mode. Before, querying such columns with masking applied returned internal errors.
  • 13 April 2023, GA. Rounding mode. A parameterized NUMERIC or BIGNUMERIC column can use ROUND_HALF_EVEN or ROUND_HALF_AWAY_FROM_ZERO. A default can be set on a table or dataset and applies to columns added later. ROUND() takes the mode as an optional argument.
  • 10 April 2023. Cached results for table snapshots. Queries against table snapshots can return results from the cache.
  • 10 April 2023, GA. Limits when querying Azure and Amazon S3 data. The maximum result size is 20 GiB of logical bytes and the maximum row size is 10 MiB. A query with ORDER BY fails when its result is larger than 256 MB, where the limit was 2 MB.
  • 5 April 2023, Preview. Non-incremental materialized views. They support most SQL queries, including OUTER JOIN, UNION, HAVING and analytic functions.
  • 29 March 2023, GA. Editions and autoscaling slots. Compute is generally available in three editions: Standard, Enterprise and Enterprise Plus, which use the slot autoscaling model. Autoscaling reservations and commitments created in the preview became Enterprise edition.
  • 28 March 2023, Preview. Model import and remote hosting. ONNX, XGBoost and TensorFlow Lite models can be imported for inference, and models can be hosted remotely on Vertex AI Prediction. Table-valued functions run inference with the Cloud Vision, Cloud Natural Language and Cloud Translation APIs. Enrollment is required.
  • 28 March 2023. tf_version, xgboost_version and instance_weight_col. tf_version picks the TensorFlow version for training, ‘1.15’ by default, and ‘2.8.0’ for TF2 with Keras. xgboost_version picks the XGBoost version, ‘0.9’ by default, or ‘1.1’. instance_weight_col names the column of per-row weights, for boosted tree and random forest models with non-array features.
  • 22 March 2023, Preview. Unicode column names. Column names can use international character sets, alphanumeric and special characters. Existing columns get such names with RENAME.
  • 20 March 2023, GA. AutoML Tables OPTIMIZATION_OBJECTIVE options. MAXIMIZE_PRECISION_AT_RECALL and MAXIMIZE_RECALL_AT_PRECISION are accepted.
  • 10 March 2023, Preview. CREATE TABLE AS SELECT from Amazon S3 and Azure Blob Storage. The statement filters data from files in S3 and Blob Storage before the results go into BigQuery tables.
  • 7 March 2023, GA. Case-insensitive collation. The release adds MIN, MAX, COUNT with DISTINCT and PERCENTILE_DISC as window functions, ORDER BY and PARTITION BY in the WINDOW clause, the LIKE operator with limits, views, materialized views, table functions and BI Engine.
  • 2 March 2023, GA. WITH RECURSIVE. A query can include one or more recursive common table expressions.
  • 27 February 2023. ARIMA_PLUS_XREG for on-demand users. The multivariate time series model is available to on-demand users.
  • 27 February 2023, GA. Column default values. Columns in tables can have default values.
  • 23 February 2023, Preview. Authorized stored procedures. A stored procedure can be shared with users or groups who have no direct access to the underlying tables.
  • 22 February 2023, Preview. Primary and foreign key constraints. CREATE TABLE, ALTER TABLE ADD PRIMARY KEY and ALTER TABLE ADD CONSTRAINT define them.
  • 21 February 2023, GA. ALTER TABLE RENAME COLUMN and ALTER TABLE DROP COLUMN.
  • 15 February 2023, GA. Case-insensitive datasets. CREATE SCHEMA and ALTER SCHEMA can make a dataset and its tables case-insensitive.
  • 13 February 2023, Preview. Materialized views over BigLake metadata cache-enabled tables. They reference structured data in Cloud Storage.
  • 9 February 2023, Preview. Autoscaling slot reservations. Autoscaling reservations and their commitments can be created.
  • 9 February 2023, GA. More dynamic data masking types. Date year, email, first four characters and last four characters masks.
  • 6 February 2023, GA. query_info column. INFORMATION_SCHEMA.JOBS, JOBS_BY_FOLDER and JOBS_BY_ORGANIZATION show query processing information in this column.
  • 6 February 2023, Preview. ANY_VALUE with HAVING MAX and HAVING MIN.
  • 19 January 2023, Preview. Preprocessing functions and TRANSFORM export. ML.ROBUST_SCALER, ML.NORMALIZER, ML.ONE_HOT_ENCODER, ML.IMPUTER, ML.MAX_ABS_SCALER and ML.LABEL_ENCODER are added. Models trained with the TRANSFORM clause can be exported in the TensorFlow SavedModel format.
  • 17 January 2023, GA. Sparse input. BigQuery ML model training supports sparse input, for data whose values are mostly zero or empty.
  • 17 January 2023, Preview. ARIMA_PLUS_XREG. Multivariate time series forecasting with extra feature columns.
  • 10 January 2023, GA. ALTER CAPACITY SET OPTIONS and ALTER RESERVATION SET OPTIONS. CREATE CAPACITY, CREATE RESERVATION and CREATE ASSIGNMENT accept an OPTIONS clause.
  • 9 January 2023, GA. Sessions. Temporary functions last until the session ends. Statements with TEMP can also use OR REPLACE and IF NOT EXISTS.
  • 4 January 2023, Preview. Vision models in BigQuery ML. Imported vision models run inference on images in object tables, for tasks such as object detection and OCR. Access needs a sign-up form.

2022

  • 15 December 2022, GA. Cloud SQL federated queries work over a private connection.
  • 1 December 2022, Preview. Apache Iceberg tables. Queries can read Iceberg tables created by open source engines.
  • 17 November 2022, Preview. Metadata caching. Queries on BigLake tables and object tables can use cached metadata, which avoids listing objects in Cloud Storage. It helps tables that reference many objects.
  • 17 November 2022, Preview. Object tables. Object tables are read-only tables holding metadata for unstructured data in Cloud Storage. BigQuery ML and remote functions can analyze the files and run inference on images, audio, documents and other file types.
  • 9 November 2022, GA. LOAD DATA from Amazon S3 and Azure Blob Storage. The statement loads hive-partitioned files, loads semi-structured JSON into JSON columns without a schema, supports customer-managed encryption keys on the destination table, and transfers to the US multi-region and US-EAST-4.
  • 3 November 2022, GA. Wrapped keyset functions. KEYS.NEW_WRAPPED_KEYSET, KEYS.ROTATE_WRAPPED_KEYSET and KEYS.REWRAP_KEYSET create, rotate and rewrap wrapped keysets in SQL. Columns can be encrypted and decrypted with a wrapped keyset.
  • 31 October 2022, Preview. max_staleness materialized view option. The option sets how stale a materialized view may be, to keep performance steady and cost controlled on large tables that change often.
  • 31 October 2022, GA. Column-level data masking. Masking hides column data from selected groups of users while still giving them access to the column.
  • 27 October 2022, GA. Search indexes and SEARCH. Search indexes and the SEARCH() function find specific values in unstructured text and semi-structured data.
  • 26 October 2022, GA. ST_ISCLOSED, ST_ISRING. ST_ISCLOSED returns TRUE for a non-empty geography where each element has an empty boundary. ST_ISRING checks whether a geography is a linestring that is closed and simple.
  • 18 October 2022, GA. Remote functions. Queries can call functions in Cloud Functions or Cloud Run.
  • 11 October 2022, Preview. Stored procedures for Apache Spark.
  • 10 October 2022, GA. Multi-statement transactions.
  • 10 October 2022, Preview. Physical bytes for storage billing. A dataset can be billed on physical bytes.
  • 3 October 2022, GA. JSON column type.
  • 29 September 2022, Preview. ROUND_HALF_EVEN rounding mode. Parameterized NUMERIC and BIGNUMERIC columns accept it, and ROUND() takes it as an optional argument.
  • 21 September 2022, GA. BigQuery Omni on-demand pricing. The on-demand model is offered for a limited time.
  • 16 September 2022, BigQuery Omni limits. The quota for total query result size per project is 1 TB per day. The maximum result size of a query rises from 2 MB to 10 GB in Preview.
  • 14 September 2022, Preview. is_case_insensitive schema option. It makes a dataset and its table names case-insensitive.
  • 9 September 2022, GA. ARIMA_PLUS options. HOLIDAY_REGION takes more than one region string, and the holidays of all listed regions are used in the model. TREND_SMOOTHING_WINDOW_SIZE smooths the trend component with a centered moving average.
  • 1 September 2022, GA. Random forest models.
  • 29 August 2022, GA. Community contributed UDFs. They live in the bigquery-utils GitHub repository and in the bigquery-public-data.persistent_udfs public dataset.
  • 19 August 2022, Preview. ALTER TABLE RENAME COLUMN.
  • 17 August 2022, Preview. Column default values.
  • 15 August 2022, GA. BigQuery Omni reservation DDL and access control DCL.
  • 10 August 2022, GA. Default configurations. Defaults can be set at the project or organization level.
  • 10 August 2022, GA. Query priority for Cloud Spanner federated queries.
  • 9 August 2022, GA. Cloud Bigtable external data sources.
  • 3 August 2022, Preview. max_staleness materialized view option. It sets a staleness limit for large, frequently changing tables. A note on 15 August says an issue was found and the feature is not available.
  • 1 August 2022, GA. CBRT. CBRT computes the cube root of a value.
  • 1 August 2022, Preview. LOAD DATA. The statement loads data from one or more files into a table.
  • 27 July 2022, GA. COT, COTH, CSC, CSCH, SEC, SECH. They compute the cotangent, hyperbolic cotangent, cosecant, hyperbolic cosecant, secant and hyperbolic secant of an angle.
  • 25 July 2022, GA. BigLake. BigQuery ML models can be created from data in Cloud Storage through BigLake tables.
  • 6 July 2022, Preview. APPENDS. This change history table-valued function returns the appends to a table over a window of time.
  • 21 June 2022, Preview. Query queues. BigQuery sets query concurrency itself instead of using a fixed limit, for on-demand and flat-rate customers. Flat-rate customers can set their own concurrency target, and queries beyond it wait until resources free up.
  • 15 June 2022, GA. Deterministic encryption functions. DETERMINISTIC_ENCRYPT, DETERMINISTIC_DECRYPT_BYTES and DETERMINISTIC_DECRYPT_STRING encrypt and decrypt columns in a way that still supports aggregation and joins.
  • 13 June 2022, GA. @@dataset_project_id. The system variable sets the default project for datasets that do not name one in the query. It is also a connection property.
  • 10 June 2022, Multi-statement query time limit. The cumulative time limit rises from 6 hours to 24 hours.
  • 3 June 2022, Preview. BigQuery Omni reservation DDL and access control DCL.
  • 31 May 2022, Preview. Column-level data masking. Masking hides column data from selected groups of users while still giving them access to the column. Combined with column-level access control, access can range from full to none per group.
  • 5 May 2022, GA. %J format element. It gives the ISO 8601 day of the year, starting at 1, in DATE, TIME, DATETIME and TIMESTAMP functions.
  • 5 May 2022. PARSE_DATE, PARSE_TIME, PARSE_DATETIME, PARSE_TIMESTAMP format elements. They accept %a, %A, %g, %G, %j, %u, %U, %V, %w and %W.
  • 3 May 2022, GA. ARIMA_PLUS forecasting metrics and fast training. ML.EVALUATE returns MAPE, SMAPE and MSE. The options TIME_SERIES_LENGTH_FRACTION, MIN_TIME_SERIES_LENGTH and MAX_TIME_SERIES_LENGTH speed up training with little or no loss of accuracy.
  • 2 May 2022, Preview. Collation. Case-insensitive collation controls how strings compare and sort. It applies to comparison operators, joins, ORDER BY, GROUP BY, set operations and some scalar and aggregate functions. The COLLATE function returns a STRING with a collation specification. The DEFAULT COLLATE clause of CREATE SCHEMA, ALTER SCHEMA, CREATE TABLE and ALTER TABLE sets a default for all columns that support collation. The COLLATE clause of ALTER TABLE ADD COLUMN and ALTER COLUMN SET DATA TYPE sets it on one column.
  • 25 April 2022, Preview. TABLE_STORAGE view. It shows current storage usage for tables and materialized views.
  • 25 April 2022, Preview. Configurable time travel window. The window can be set from two to seven days.
  • 7 April 2022, Preview. BigLake. BigLake tables query data across cloud data lakes and warehouses, with access policies on columns or rows.
  • 7 April 2022, Preview. Search indexes and SEARCH. They find data elements in unstructured text and semi-structured data.
  • 1 April 2022, Preview. BigQuery ML and Vertex AI Model Registry. BigQuery ML models can be sent to the Vertex AI Model Registry, organized there and deployed to endpoints.
  • 28 March 2022, GA. Wide-and-Deep models.
  • 16 March 2022. Explicit schema for external tables. External tables over Parquet, ORC and Avro files can take a schema you give them. Before, the schema was always detected from the last file in lexicographic order.
  • 4 March 2022, GA. Sessions. A session can be terminated automatically or by hand, and a label can be set for all queries in it. INFORMATION_SCHEMA.SESSIONS_BY_PROJECT and INFORMATION_SCHEMA.SESSIONS_BY_USER return session metadata.
  • 16 February 2022, Preview. Remote functions. They let a function be written in languages other than SQL and JavaScript, or use libraries and services that user-defined functions do not allow.
  • 15 February 2022, Preview. Table clones. A clone is a lightweight, writable copy of a table. Storage is charged only for data that differs from the base table.
  • 14 February 2022. ARIMA_PLUS training speed. BigQuery ML trains ARIMA_PLUS models 5 times faster than before.
  • 14 February 2022, GA. INFORMATION_SCHEMA.STREAMING_TIMELINE_* views.
  • 14 February 2022, GA. QUALIFY. The clause filters the results of analytic functions.
  • 10 February 2022, Preview. BigQuery Omni INFORMATION_SCHEMA views. Omni supports the INFORMATION_SCHEMA.JOBS_* and INFORMATION_SCHEMA.RESERVATION* views.
  • 3 February 2022, GA. BigQuery ML hyperparameter tuning. CREATE MODEL searches for the best hyperparameters while training.
  • 2 February 2022, Preview. WITH RECURSIVE. A query in a WITH clause can refer to itself or to queries defined later in the clause.
  • 31 January 2022, GA. Materialized views without aggregation and with inner joins.
  • 25 January 2022, GA. Explainable AI in BigQuery ML. It explains prediction and forecasting results.
  • 6 January 2022, Preview. JSON data type.

2021

  • 16 December 2021. Row-level security supports administrator access to historical data for tables with row-level access policies.
  • 14 December 2021, GA. BI Engine SQL interface.
  • 6 December 2021, GA. ML.DETECT_ANOMALIES. The function detects anomalies in time-series data with ARIMA_PLUS models. With K-means, autoencoder or PCA models it detects anomalies in independent and identically distributed data.
  • 3 December 2021, GA. PCA and autoencoder models. They serve dimensionality reduction, feature embedding and unsupervised anomaly detection.
  • 9 November 2021, GA. Scripting statements. CASE runs the first statement list whose boolean expression is TRUE. CASE search_expression runs the first list whose WHEN expression matches the search expression. Labels give an unconditional jump to the end of the labeled block or loop. REPEAT runs a statement list until the condition at its end is TRUE. FOR...IN loops over every row of a table expression.
  • 8 November 2021, GA. DDL column. The ROUTINES, SCHEMATA and TABLES views in INFORMATION_SCHEMA have a DDL column holding the statement that creates the resource.
  • 2 November 2021, GA. Parameterized types. STRING(L), BYTES(L), NUMERIC(P), NUMERIC(P, S), BIGNUMERIC(P) and BIGNUMERIC(P, S).
  • 28 October 2021, GA. Table snapshots. A snapshot can be created in a different project from its base table.
  • 27 October 2021, GA. SQL column-level encryption with Cloud KMS. The AEAD encryption functions work with keysets that are encrypted by Cloud KMS.
  • 25 October 2021, GA. BigQuery Omni. Omni is the multi-cloud analytics offering.
  • 1 October 2021, Preview. ST_BUFFER, ST_BUFFERWITHTOLERANCE. Both return a GEOGRAPHY for the buffer around the input. ST_BUFFER takes a number of segments and ST_BUFFERWITHTOLERANCE takes a tolerance, which set how far the result may deviate from the ideal radius.
  • 28 September 2021, GA. Table functions. Authorized table functions are supported.
  • 27 September 2021, GA. ST_BOUNDINGBOX, ST_EXTENT, S2_COVERINGCELLIDS, S2_CELLIDFROMPOINT. ST_BOUNDINGBOX returns a STRUCT for the bounding box of a geography and ST_EXTENT does so for a set of geographies. S2_COVERINGCELLIDS returns an array of S2 cell IDs covering a geography. S2_CELLIDFROMPOINT returns the S2 cell ID covering a point.
  • 17 September 2021, GA. ST_EXTERIORRING, ST_INTERIORRINGS, ST_ANGLE, ST_AZIMUTH, ST_NUMGEOMETRIES, ST_GEOMETRYTYPE. ST_EXTERIORRING returns the outer ring of a polygon as a linestring and ST_INTERIORRINGS returns the inner rings. ST_ANGLE returns the angle between two intersecting lines. ST_AZIMUTH returns the azimuth of a segment between two points. ST_NUMGEOMETRIES counts the geometries in a geography. ST_GEOMETRYTYPE returns the OGC geometry type as a string.
  • 8 September 2021, Preview. Sessions. A session ties SQL activity across scripts and multi-statement transactions to one session identifier. Session variables such as the default time zone or dataset and temporary tables last for the whole session. The jobs INFORMATION_SCHEMA views have a SESSION_ID column.
  • 19 August 2021, GA. Cloud Spanner federated queries.
  • 6 August 2021, Preview. PCA models. CREATE MODEL supports principal component analysis.
  • 3 August 2021, GA. ALTER COLUMN SET DATA TYPE.
  • 3 August 2021, GA. CREATE TABLE LIKE and CREATE TABLE COPY.
  • 28 July 2021, Preview. Wide-and-Deep models. 'DNN_LINEAR_COMBINED_CLASSIFIER' and 'DNN_LINEAR_COMBINED_REGRESSOR' create Wide-and-Deep classifier and regressor models.
  • 27 July 2021, Preview. Explainable AI in BigQuery ML. It shows how each feature contributed to a prediction for classification and regression models.
  • 27 July 2021, Preview. INTERVAL type. The type represents a duration or an amount of time.
  • 26 July 2021, GA. Holiday effects for weekly time series. Time series models handled daily series before.
  • 26 July 2021, GA. DML statistics. DML jobs return the number of rows inserted, deleted and updated, and the INFORMATION_SCHEMA.JOBS_BY_* views carry the same statistics.
  • 19 July 2021, GA. Workload management statements. CREATE CAPACITY, CREATE RESERVATION, CREATE ASSIGNMENT, DROP CAPACITY, DROP RESERVATION and DROP ASSIGNMENT.
  • 19 July 2021, GA. PIVOT and UNPIVOT.
  • 19 July 2021, GA. CONTAINS_SUBSTR.
  • 7 July 2021, Preview. Materialized views without aggregation and with inner joins.
  • 29 June 2021, Preview. Multi-statement transactions. They run mutating operations on one or more tables and commit or roll back the changes atomically.
  • 28 June 2021, GA. GRANT and REVOKE. The INFORMATION_SCHEMA.OBJECT_PRIVILEGES view is in Preview.
  • 28 June 2021, Preview. Table functions. These user-defined table-valued functions return a table.
  • 28 June 2021, GA. Casting features. PARSE_BIGNUMERIC, PARSE_NUMERIC, the FORMAT clause of CAST for STRING, DATE, DATETIME, TIME, TIMESTAMP, numeric and BYTES types, the INT64 aliases INT, SMALLINT, INTEGER, BIGINT, TINYINT and BYTEINT, and ST_GEOGFROM.
  • 28 June 2021, GA. ALTER COLUMN SET OPTIONS.
  • 25 June 2021, Preview. Table snapshots. A snapshot is a low-cost, read-only copy of a table as it was at a point in time.
  • 22 June 2021, Preview. BigQuery ML anomaly detection, autoencoder and hyperparameter tuning. ML.DETECT_ANOMALIES runs against time series with ARIMA_PLUS models and against independent and identically distributed data with AUTOENCODER and KMEANS models. The AUTOENCODER model type is available in CREATE MODEL, and ML.PREDICT can use it to reduce the dimensions of query results. Hyperparameter tuning searches for the best hyperparameters while CREATE MODEL trains.
  • 21 June 2021, GA. Row-level security.
  • 7 June 2021, Preview. Parameterized types. STRING(L), BYTES(L), NUMERIC(P), NUMERIC(P, S), BIGNUMERIC(P) and BIGNUMERIC(P, S).
  • 20 May 2021, GA. ST_STARTPOINT, ST_ENDPOINT, ST_POINTN. They return a point of a linestring geography as a point geography.
  • 19 May 2021, GA. ALTER TABLE RENAME TO.
  • 18 May 2021, GA. CREATE MODEL for AutoML Tables models.
  • 10 May 2021, Preview. PIVOT, UNPIVOT and QUALIFY.
  • 30 April 2021, GA. CREATE VIEW with a column name list, ALTER COLUMN DROP NOT NULL.
  • 21 April 2021. Changing clustering. A non-clustered table can become clustered and the reverse, and the clustered columns of a clustered table can change. Google had documented this since October 2020 without a release note.
  • 19 April 2021. ARIMA_PLUS models. They replace the ARIMA model type, which is deprecated. ARIMA_PLUS adds ML.EXPLAIN_FORECAST, a fuller ML.ARIMA_EVALUATE, several ID columns through time_series_id_col, up to 500,000 time series forecast at once, the training options clean_spikes_and_dips and adjust_step_changes, and the per_minute data frequency.
  • 2 April 2021, Preview. ALTER TABLE DROP COLUMN.
  • 2 April 2021. Name length limit. The maximum length of table column names, column alias names and user-defined function names rises from 128 to 300 characters.
  • 31 March 2021, GA. JSON_EXTRACT_STRING_ARRAY, JSON_QUERY_ARRAY, JSON_VALUE_ARRAY.
  • 31 March 2021, Preview. Partitions view. INFORMATION_SCHEMA has a view for table partitions.
  • 31 March 2021, Preview. TABLESAMPLE. The operator queries random subsets of large tables.
  • 31 March 2021, GA. CREATE SCHEMA, ALTER SCHEMA, DROP SCHEMA.
  • 31 March 2021, Preview. DDL column in INFORMATION_SCHEMA.TABLES. It holds the statement that recreates the table.
  • 31 March 2021, GA. BIGNUMERIC type.
  • 25 February 2021, GA. Materialized views. They are precomputed views that cache the results of a query periodically.
  • 12 February 2021. Cached results for Cloud Storage data. Standard SQL queries on Cloud Storage data can use cached query results.
  • 29 January 2021, GA. DATETIME clustering columns.

2020

  • 21 December 2020, Preview. BIGNUMERIC type. It is meant for high-precision computations.
  • 23 November 2020, GA. EXPORT MODEL. BigQuery ML models can be exported to Cloud Storage and used for online prediction.
  • 23 November 2020, GA. Boosted tree and DNN models. CREATE MODEL trains boosted tree models with XGBoost and deep neural network models through AI Platform. ML.FEATURE_IMPORTANCE covers the boosted tree models.
  • 26 October 2020, GA. SUBSTRING.
  • 22 October 2020, Beta. make_valid and planar parameters. ST_GEOGFROMGEOJSON and ST_GEOGFROMTEXT take make_valid, which tries to correct polygon problems on import when set to TRUE. ST_GEOGFROMTEXT also takes planar, which treats imported WKT geometries as having planar edges when set to TRUE.
  • 14 October 2020, GA. Dynamic SQL. EXECUTE IMMEDIATE generates and runs SQL statements at runtime.
  • 14 October 2020, GA. String and date functions. ASCII, CHR, INITCAP, INSTR, LAST_DAY, LEFT, OCTET_LENGTH, REGEXP_INSTR, REGEXP_SUBSTR, RIGHT, SOUNDEX, TRANSLATE and UNICODE. REGEXP_EXTRACT takes two more parameters, position and occurrence.
  • 14 October 2020, GA. Statements. CREATE EXTERNAL TABLE, DROP EXTERNAL TABLE, ALTER TABLE ADD COLUMN, EXPORT DATA and TRUNCATE TABLE.
  • 14 October 2020. DATE arithmetic operators.
  • 14 October 2020. Unicode table names.
  • 14 October 2020, GA. INFORMATION_SCHEMA views. TABLES, TABLE_OPTIONS, COLUMNS, COLUMN_FIELD_PATHS, VIEWS, ROUTINES, ROUTINE_OPTIONS, PARAMETERS, SCHEMATA and SCHEMATA_OPTIONS.
  • 14 October 2020. Duplicate column names. A query result can have duplicate column names.
  • 6 October 2020, GA. Authorized user-defined functions. They share query results without giving access to the underlying tables.
  • 29 September 2020. DATA_FREQUENCY for multiple time series. Time series models accept a DATA_FREQUENCY other than the default AUTO_FREQUENCY when TIME_SERIES_ID_COL forecasts several series.
  • 21 September 2020, GA. Time-unit partitioning. Tables can be partitioned by hour, month or year, and by a DATETIME column.
  • 11 September 2020. BQ.JOBS.CANCEL. The system procedure cancels a job.
  • 27 August 2020, GA. Time series models. The AUTO_ARIMA_MAX_ORDER training option is added.
  • 26 August 2020, GA. Reservations views. INFORMATION_SCHEMA has views for BigQuery reservations.
  • 25 August 2020. Jobs timeline views. INFORMATION_SCHEMA has views for jobs metadata by timeslice.
  • 19 August 2020. Consecutive ON and USING clauses. Parentheses around the nested join are optional, so FROM A JOIN B JOIN C ON B.x = C.y USING (z) works like FROM A JOIN (B JOIN C ON B.x = C.y) USING (z).
  • 18 August 2020. On-demand pricing with reservations. A project can be assigned to none, which keeps it on on-demand pricing.
  • 17 August 2020, GA. Matrix factorization models. CREATE MODEL and ML.RECOMMEND.
  • 11 August 2020. Flat-rate minimum. The minimum slot purchase is 100 slots, in 100-slot increments.
  • 27 July 2020, Alpha. Streaming views. INFORMATION_SCHEMA has views with historical and real-time information about streaming into BigQuery.
  • 16 July 2020. ST_CONVEXHULL, ST_DUMP. ST_CONVEXHULL returns the smallest convex GEOGRAPHY that covers the input. ST_DUMP returns an ARRAY of simple GEOGRAPHY values, one per component of the input.
  • 15 July 2020. AutoML Tables data split options. Data split and validation options are available for AutoML Tables model training.
  • 13 July 2020. ASSERT. The statement checks that data matches an expectation.
  • 1 July 2020, Beta. Time series models. CREATE MODEL supports time series models.
  • 29 June 2020, GA. Flex slots.
  • 26 June 2020, Beta. Region-qualified INFORMATION_SCHEMA views.
  • 26 June 2020. Unqualified SCHEMATA queries. From mid-August, an unqualified INFORMATION_SCHEMA query on SCHEMATA or SCHEMATA_OPTIONS returns metadata from the US multi-region.
  • 16 June 2020, GA. Jobs views. INFORMATION_SCHEMA has views for jobs.
  • 16 June 2020, Preview. BigQuery ML and AI Platform. CREATE MODEL supports AutoML Tables models, boosted tree models with XGBoost and deep neural network models.
  • 9 June 2020. Clustering for non-partitioned tables.
  • 2 June 2020. ST_SIMPLIFY. It returns a simplified GEOGRAPHY by replacing sections of the input with straight lines.
  • 2 June 2020. References in UDF and view bodies. The body of a standard SQL user-defined function can reference tables and views. The body of a standard SQL view can contain references without a project qualifier when the view is created through the tables.insert API or lives in the project that runs the CREATE VIEW query.
  • 20 May 2020, Beta. Hourly partitioned tables.
  • 20 May 2020, GA. Cloud SQL federated queries.
  • 20 May 2020, Beta. Dynamic SQL. EXECUTE IMMEDIATE generates and runs SQL statements at runtime, in all regions.
  • 1 May 2020. JSON_EXTRACT_ARRAY. It extracts the contents of a JSON document as a string array.
  • 1 May 2020. NULLS FIRST and NULLS LAST. ORDER BY accepts both to place null and non-null values.
  • 22 April 2020, Beta. Exporting BigQuery ML models. Models can be exported to Cloud Storage and used for online prediction.
  • 17 April 2020, Beta. Matrix factorization models. CREATE MODEL builds recommendation models.
  • 16 April 2020, GA. BigQuery Reservations. Reservations buy slots for flat-rate pricing and assign them to workloads.
  • 16 April 2020. Dataset metadata views. Around the end of April, the INFORMATION_SCHEMA dataset views return metadata for all datasets in a region instead of all datasets in the project across regions. A specific region can be queried, for example region-us.INFORMATION_SCHEMA.SCHEMATA. Filtering on the LOCATION column of SCHEMATA gives the same result before the change.
  • 8 April 2020, Beta. Materialized views.
  • 31 March 2020, Alpha. Reservations views. INFORMATION_SCHEMA has views for BigQuery reservations.
  • 30 March 2020, GA. Scripting and stored procedures.
  • 2 March 2020, GA. External data in ORC and Parquet, and hive partitioning. Queries can read ORC and Parquet files in Cloud Storage. Externally partitioned data in Cloud Storage with the default hive layout can be loaded and queried, in Avro, Parquet, ORC, CSV or JSON format.
  • 26 February 2020. DML quotas. DML statements have no quota limits.
  • 26 February 2020, Beta. Flex slots. They are available in all Reservations regions.
  • 18 February 2020, GA. Integer range partitioning. Tables can be partitioned on an integer column.
  • 14 February 2020, Beta. Jobs views. INFORMATION_SCHEMA has views for live and historical jobs metadata.

2019

  • 20 December 2019, GA. ALTER TABLE and ALTER VIEW. They set options on tables and views.
  • 19 December 2019, GA. BigQuery ML preprocessing. The preprocessing functions and the TRANSFORM clause handle feature engineering.
  • 13 December 2019, Beta. BigQuery Reservations. Reservations buy slots for flat-rate pricing and assign them to workloads, in the US and EU.
  • 11 December 2019, Beta. Integer range partitioning. Tables can be partitioned on an integer column.
  • 10 December 2019. On-demand query acceleration. On-demand queries can burst beyond the default slot capacity at no extra charge.
  • 4 December 2019. KMEANS++. It initializes the clusters of a k-means model and trains a better model than random initialization.
  • 26 November 2019. INFORMATION_SCHEMA permission checks. A query against an INFORMATION_SCHEMA view returns an accessDenied (403) error when the user lacks the IAM permission to read the metadata. Before, the results were filtered by those permissions. INFORMATION_SCHEMA is in beta.
  • 22 November 2019, GA. Persistent standard SQL UDFs.
  • 22 November 2019, Beta. Routines metadata. INFORMATION_SCHEMA views return routine metadata.
  • 21 November 2019, Beta. BigQuery ML preprocessing.
  • 21 November 2019. BigQuery ML encryption keys. ML models can be encrypted with your own Cloud KMS keys.
  • 12 November 2019, GA. Omitting column names in INSERT and MERGE.
  • 31 October 2019, Beta. External data in ORC and Parquet, and hive partitioning. Queries can read ORC and Parquet files in Cloud Storage. Externally partitioned data in Cloud Storage with the default hive layout can be loaded and queried, in Avro, Parquet, ORC, CSV or JSON format.
  • 9 October 2019. RANGE_BUCKET. The function scans a sorted array and returns the 0-based position of the upper bound of a given point. It helps group data into partitions, histograms and business-defined rules.
  • 3 October 2019, Beta. Scripting, stored procedures and temporary tables.
  • 30 September 2019, GA. Importing TensorFlow models.
  • 23 September 2019, GA. K-means clustering models.
  • 27 August 2019. Google Sheets external tables. An external table linked to a Google Sheets file can name a single sheet or a cell range, not only the first sheet.
  • 22 August 2019. Automatic re-clustering. BigQuery re-clusters all clustered tables itself, so manual re-clustering is no longer needed.
  • 7 August 2019, Beta. Cloud SQL federated queries. A query can combine BigQuery and Cloud SQL data.
  • 2 July 2019, Beta. Importing TensorFlow models.
  • 1 July 2019, Beta. Persistent user-defined functions.
  • 3 June 2019, Beta. Omitting column names in INSERT and MERGE.
  • 29 May 2019, GA. BigQuery ML.
  • 28 May 2019. Maximum query response size. It rises from 128 MB to 10 GB.
  • 15 May 2019, GA. Clustering by GEOGRAPHY. A table can be clustered by a GEOGRAPHY column.
  • 14 May 2019. DROP MODEL. The DDL statement deletes a BigQuery ML model.
  • 13 May 2019. Concurrent query limit. The rate limit for on-demand interactive queries rises from 50 to 100 concurrent queries.
  • 10 April 2019. AEAD encryption functions.
  • 10 April 2019, GA. Clustered tables.
  • 10 April 2019. K-means models. BigQuery ML supports the k-means model type for clustering and customer segmentation.
  • 4 April 2019, GA. BigQuery GIS.
  • 18 March 2019. CREATE MODEL query limit. The limit on CREATE MODEL queries rises from 100 to 1,000.
  • 18 March 2019. ML.ROC_CURVE, ML.CONFUSION_MATRIX. Both work without input data.
  • 25 February 2019, GA. MERGE.
  • 11 February 2019. DML quotas. The limit on combined UPDATE, DELETE and MERGE statements per table per day rises from 200 to 1,000. The limit of 10,000 such statements per project per day is removed.
  • 4 February 2019, Beta. ALTER TABLE and ALTER VIEW set options.
  • 4 February 2019, Beta. INFORMATION_SCHEMA. The views return resource metadata.
  • 29 January 2019. Linear regression optimization strategies. BigQuery ML supports automatic, batch gradient descent and normal equation strategies.

2018

  • 13 December 2018. ML.PREDICT thresholds. ML.PREDICT supports thresholds for binary logistic regression models.
  • 13 December 2018. ML.WEIGHTS standardization. ML.WEIGHTS supports standardization.
  • 12 December 2018. Query length. The maximum length of a standard SQL query is 1 MB.
  • 5 December 2018. Default partition expiration. A dataset can have a default partition expiration time, set or updated later.
  • 19 October 2018. Label values in CREATE MODEL. Multiclass logistic regression models support up to 50 unique label values, up from 10.
  • 11 October 2018. Model training and evaluation. The random data split method is deterministic, so later training runs give the same split while the input data is unchanged. Input data for ML.EVALUATE is optional.
  • 10 October 2018, GA. DML on partitioned tables.
  • 10 October 2018, GA. Tables partitioned on a DATE or TIMESTAMP column.
  • 9 October 2018. Job location. A job that references resources outside the US and EU regions no longer needs a location.
  • 25 September 2018. GENERATE_UUID.
  • 13 September 2018. Multiclass logistic regression. BigQuery ML supports multiclass logistic regression models and ML.CONFUSION_MATRIX.
  • 10 September 2018, Beta. BigQuery GIS.
  • 20 August 2018, GA. NUMERIC type.
  • 30 July 2018, Beta. Clustered tables.
  • 25 July 2018, Beta. BigQuery ML.
  • 25 July 2018, GA. Data definition language.
  • 19 July 2018, Alpha. Clustered tables.
  • 28 June 2018. Templated SQL UDFs.
  • 28 June 2018. GENERATE_TIMESTAMP_ARRAY.
  • 28 June 2018. FROM_BASE32, TO_BASE32.
  • 15 May 2018, Beta. NUMERIC type.
  • 4 May 2018. Quotas. A partitioned table can have up to 4,000 partitions, up from 2,500. The limit on combined UPDATE, DELETE and MERGE statements per table per day rises from 96 to 200.
  • 10 April 2018, Beta. MERGE.
  • 30 March 2018. ERROR and the SAFE prefix. ERROR returns an error from an expression. The SAFE prefix suppresses errors in a function call.
  • 5 March 2018. Create a table from a query. The data definition language can create a table from the result of a query.
  • 5 March 2018, Beta. DML on partitioned tables.
  • 8 February 2018, Beta. Tables partitioned on a TIMESTAMP or DATE column.
  • 17 January 2018, Beta. Data definition language.
  • 5 January 2018. ISOWEEK and ISOYEAR. DATE_TRUNC, DATETIME_TRUNC, TIMESTAMP_TRUNC, DATE_DIFF and DATETIME_DIFF accept them.

2017

  • 30 November 2017. FARM_FINGERPRINT in legacy SQL.
  • 30 November 2017. _PARTITIONTIME in wildcard tables. The pseudo column works when querying partitioned tables through a wildcard table.
  • 30 November 2017. WEEK(<WEEKDAY>). It is a valid date part for EXTRACT.
  • 30 November 2017. INTERSECT DISTINCT and EXCEPT DISTINCT. Standard SQL supports both set operators.
  • 14 November 2017. High compute queries. Opting in is no longer required, and a query under tier 100 is billed as tier 1 without a threshold flag. A query above tier 100 fails with a RESOURCES_EXCEEDED_PER_BYTE error unless it runs in a reserved instance. Users who ran one above tier 100 in the previous 30 days keep the higher billing tier until 31 October 2018.
  • 2 November 2017. IGNORE NULLS and RESPECT NULLS. FIRST_VALUE, LAST_VALUE and NTH_VALUE accept them.
  • 2 November 2017. FOR SYSTEM_TIME AS OF. The expression is allowed in the FROM clause.
  • 29 September 2017. GROUP BY ROLLUP. Standard SQL supports it.
  • 28 September 2017. Ingestion-time partitioned tables. Partitioning options given when query results or loaded data go into a new destination table create an ingestion-time partitioned table.
  • 28 September 2017. NORMALIZE, NORMALIZE_AND_CASEFOLD.
  • 1 September 2017. PERCENTILE_CONT, PERCENTILE_DISC. Standard SQL supports both analytic functions.
  • 11 August 2017. OVER for statistical aggregates. COVAR_SAMP, COVAR_POP and CORR accept an OVER clause.
  • 2 August 2017. EXTRACT date parts. Standard SQL EXTRACT supports WEEK, ISOWEEK and ISOYEAR.
  • 26 July 2017. TO_HEX, FROM_HEX. Standard SQL supports both.
  • 13 July 2017. DML on tables with required fields. DML statements can modify such tables. A SELECT that writes to a destination table can also update them.
  • 5 June 2017. OVER for statistical aggregates. STDDEV, STDDEV_POP, STDDEV_SAMP, VAR_POP, VAR_SAMP and VARIANCE accept an OVER clause.
  • 2 June 2017. TO_JSON_STRING. Standard SQL supports it.
  • 16 March 2017. REVERSE, ARRAY_REVERSE. They reverse STRING, BYTES and ARRAY values.
  • 15 March 2017. HyperLogLog++ functions. They estimate the number of unique values in a large dataset.
  • 15 March 2017. ARRAY_AGG and STRING_AGG as window functions. Both accept an OVER clause.
  • 8 March 2017, Beta. Cloud Bigtable external data sources.
  • 8 March 2017, GA. IGNORE NULLS and RESPECT NULLS in ARRAY_AGG and APPROX_QUANTILES.
  • 8 March 2017. DML limits. The daily limits for update and delete statements are higher.
  • 8 March 2017, GA. DML for nonpartitioned tables.
  • 1 February 2017. Standard SQL additions. Bitwise operators for BYTES, BIT_COUNT for BYTES, LPAD, RPAD, REPEAT, FARM_FINGERPRINT, and the NET functions NET.IP_FROM_STRING, NET.SAFE_IP_FROM_STRING, NET.IP_TO_STRING, NET.IP_NET_MASK, NET.IP_TRUNC, NET.IPV4_FROM_INT64 and NET.IPV4_TO_INT64.

2016

  • 22 December 2016. Hash functions. Standard SQL supports them.
  • 22 December 2016. TO_CODE_POINTS, CODE_POINTS_TO_BYTES, CODE_POINTS_TO_STRING.
  • 22 December 2016. URL functions. Standard SQL supports them.
  • 22 December 2016. Parameterized queries. They help prevent SQL injection when a query is built from user input.
  • 22 December 2016. ORDER BY and LIMIT inside aggregates. ARRAY_AGG, ARRAY_CONCAT_AGG and STRING_AGG accept both clauses.
  • 22 December 2016, Issue. Legacy SQL OMIT...IF. A query returns an error when OMIT...IF excludes part of a record in a repeating field and the query also selects other independently repeating fields. Migrating the query to standard SQL avoids the error.
  • 30 November 2016. JSON functions. Standard SQL supports them.
  • 1 November 2016. SQL variant prefix. A query prefix sets the SQL variant, legacy SQL or standard SQL.
  • 1 November 2016. Safe casting functions. They return NULL where a normal cast would raise an error.
  • 1 November 2016. Bracket notation in legacy SQL JSON functions.
  • 29 September 2016, GA. Standard SQL. It follows the SQL 2011 standard, with extensions for nested and repeated data.
  • 29 September 2016. DML for standard SQL. DML inserts, updates and deletes rows.
  • 29 September 2016. Flat-rate pricing. It gives unlimited querying for a fixed monthly price.
  • 2 June 2016. Time-partitioned tables.
  • 2 June 2016, Beta. Standard SQL. It follows the SQL 2011 standard, with extensions for nested and repeated data.
  • 6 May 2016. Google Drive data. A Google Drive document can be a federated data source, and a query can read one through a temporary table.
  • 6 May 2016. Legacy SQL functions. PARSE_PACKED_IP returns BYTES. FROM_BASE64, SHA1 and TO_BASE64 are added.
  • 23 March 2016. BYTES type.

2015

  • 25 August 2015. User-defined functions.
  • 25 August 2015. Slots and query pricing tiers. Slots guarantee resources regardless of demand on the shared pool.
  • 25 August 2015. Quotas. The concurrent query rate limit rises to 50 queries and the daily query limit to 100,000.
  • 25 August 2015. Cloud Storage data. Queries can read CSV and JSON data in Google Cloud Storage directly, as a federated data source.
  • 16 April 2015. Undelete a table. A deleted table can be restored with table decorators.
  • 16 April 2015. Query reference additions. COALESCE, GROUP_CONCAT_UNQUOTED, RIGHT OUTER JOIN, FULL OUTER JOIN, and the ROLLUP and GROUPING modifiers.

2014

  • 25 March 2014. Table wildcard functions. They query a specific set of tables at lower cost.
  • 25 March 2014. Joins. CROSS JOIN and multiple joins in one query are supported.
  • 25 March 2014. Views. A view is a virtual table defined by a SQL query.
  • 25 March 2014. Batch queries. They no longer wait 30 minutes. A batch job starts as soon as idle resources are available, and becomes interactive if it has not started within 3 hours.
  • 25 March 2014. JSON functions.
  • 25 March 2014. Pricing model. Rates are lower and reserved capacity pricing is added.

2013

2012

  • 28 August 2012. Batch queries. They run between 30 minutes and 3 hours later and cost less than interactive queries.
  • 1 March 2012. VARIANCE, STDDEV, AVG. The aggregate functions are supported.
  • 5 April 2012. Join behavior. Outer joins are not supported, and only the left-most table of a join can be large.

2011