Skip to main content

CREATE VIEW

Creates a new view. Views can be normal, materialized, refreshable materialized, and window (refreshable materialized view and window view are experimental features).

Normal View

Syntax:

CREATE [OR REPLACE] VIEW [IF NOT EXISTS] [db.]table_name [ON CLUSTER cluster_name]
[DEFINER = { user | CURRENT_USER }] [SQL SECURITY { DEFINER | INVOKER | NONE }]
AS SELECT ...
[COMMENT 'comment']

Normal views do not store any data. They just perform a read from another table on each access. In other words, a normal view is nothing more than a saved query. When reading from a view, this saved query is used as a subquery in the FROM clause.

As an example, assume you’ve created a view:

CREATE VIEW view AS SELECT ...

and written a query:

SELECT a, b, c FROM view

This query is fully equivalent to using the subquery:

SELECT a, b, c FROM (SELECT ...)

Parameterized View

Parametrized views are similar to normal views, but can be created with parameters which are not resolved immediately. These views can be used with table functions, which specify the name of the view as function name and the parameter values as its arguments.

CREATE VIEW view AS SELECT * FROM TABLE WHERE Column1={column1:datatype1} and Column2={column2:datatype2} ...

The above creates a view for table which can be used as table function by substituting parameters as shown below.

SELECT * FROM view(column1=value1, column2=value2 ...)

Materialized View

CREATE MATERIALIZED VIEW [IF NOT EXISTS] [db.]table_name [ON CLUSTER] [TO[db.]name] [ENGINE = engine] [POPULATE]
[DEFINER = { user | CURRENT_USER }] [SQL SECURITY { DEFINER | INVOKER | NONE }]
AS SELECT ...
[COMMENT 'comment']
Tip

Here is a step by step guide on using Materialized views.

Materialized views store data transformed by the corresponding SELECT query.

When creating a materialized view without TO [db].[table], you must specify ENGINE – the table engine for storing data.

When creating a materialized view with TO [db].[table], you can't also use POPULATE.

A materialized view is implemented as follows: when inserting data to the table specified in SELECT, part of the inserted data is converted by this SELECT query, and the result is inserted in the view.

Note

Materialized views in ClickHouse use column names instead of column order during insertion into destination table. If some column names are not present in the SELECT query result, ClickHouse uses a default value, even if the column is not Nullable. A safe practice would be to add aliases for every column when using Materialized views.

Materialized views in ClickHouse are implemented more like insert triggers. If there’s some aggregation in the view query, it’s applied only to the batch of freshly inserted data. Any changes to existing data of source table (like update, delete, drop partition, etc.) does not change the materialized view.

Materialized views in ClickHouse do not have deterministic behaviour in case of errors. This means that blocks that had been already written will be preserved in the destination table, but all blocks after error will not.

By default if pushing to one of views fails, then the INSERT query will fail too, and some blocks may not be written to the destination table. This can be changed using materialized_views_ignore_errors setting (you should set it for INSERT query), if you will set materialized_views_ignore_errors=true, then any errors while pushing to views will be ignored and all blocks will be written to the destination table.

Also note, that materialized_views_ignore_errors set to true by default for system.*_log tables.

If you specify POPULATE, the existing table data is inserted into the view when creating it, as if making a CREATE TABLE ... AS SELECT ... . Otherwise, the query contains only the data inserted in the table after creating the view. We do not recommend using POPULATE, since data inserted in the table during the view creation will not be inserted in it.

Note

Given that POPULATE works like CREATE TABLE ... AS SELECT ... it has limitations:

  • It is not supported with Replicated database
  • It is not supported in ClickHouse cloud

Instead a separate INSERT ... SELECT can be used.

A SELECT query can contain DISTINCT, GROUP BY, ORDER BY, LIMIT. Note that the corresponding conversions are performed independently on each block of inserted data. For example, if GROUP BY is set, data is aggregated during insertion, but only within a single packet of inserted data. The data won’t be further aggregated. The exception is when using an ENGINE that independently performs data aggregation, such as SummingMergeTree.

The execution of ALTER queries on materialized views has limitations, for example, you can not update the SELECT query, so this might be inconvenient. If the materialized view uses the construction TO [db.]name, you can DETACH the view, run ALTER for the target table, and then ATTACH the previously detached (DETACH) view.

Note that materialized view is influenced by optimize_on_insert setting. The data is merged before the insertion into a view.

Views look the same as normal tables. For example, they are listed in the result of the SHOW TABLES query.

To delete a view, use DROP VIEW. Although DROP TABLE works for VIEWs as well.

SQL security

DEFINER and SQL SECURITY allow you to specify which ClickHouse user to use when executing the view's underlying query. SQL SECURITY has three legal values: DEFINER, INVOKER, or NONE. You can specify any existing user or CURRENT_USER in the DEFINER clause.

The following table will explain which rights are required for which user in order to select from view. Note that regardless of the SQL security option, in every case it is still required to have GRANT SELECT ON <view> in order to read from it.

SQL security optionViewMaterialized View
DEFINER alicealice must have a SELECT grant for the view's source table.alice must have a SELECT grant for the view's source table and an INSERT grant for the view's target table.
INVOKERUser must have a SELECT grant for the view's source table.SQL SECURITY INVOKER can't be specified for materialized views.
NONE--
Note

SQL SECURITY NONE is a deprecated option. Any user with the rights to create views with SQL SECURITY NONE will be able to execute any arbitrary query. Thus, it is required to have GRANT ALLOW SQL SECURITY NONE TO <user> in order to create a view with this option.

If DEFINER/SQL SECURITY aren't specified, the default values are used:

If a view is attached without DEFINER/SQL SECURITY specified, the default value is SQL SECURITY NONE for the materialized view and SQL SECURITY INVOKER for the normal view.

To change SQL security for an existing view, use

ALTER TABLE MODIFY SQL SECURITY { DEFINER | INVOKER | NONE } [DEFINER = { user | CURRENT_USER }]

Examples

CREATE VIEW test_view
DEFINER = alice SQL SECURITY DEFINER
AS SELECT ...
CREATE VIEW test_view
SQL SECURITY INVOKER
AS SELECT ...

Live View [Deprecated]

This feature is deprecated and will be removed in the future.

For your convenience, the old documentation is located here

Refreshable Materialized View [Experimental]

CREATE MATERIALIZED VIEW [IF NOT EXISTS] [db.]table_name
REFRESH EVERY|AFTER interval [OFFSET interval]
RANDOMIZE FOR interval
DEPENDS ON [db.]name [, [db.]name [, ...]]
SETTINGS name = value [, name = value [, ...]]
[APPEND]
[TO[db.]name] [(columns)] [ENGINE = engine] [EMPTY]
AS SELECT ...
[COMMENT 'comment']

where interval is a sequence of simple intervals:

number SECOND|MINUTE|HOUR|DAY|WEEK|MONTH|YEAR

Periodically runs the corresponding query and stores its result in a table.

  • If the query says APPEND, each refresh inserts rows into the table without deleting existing rows. The insert is not atomic, just like a regular INSERT SELECT.
  • Otherwise each refresh atomically replaces the table's previous contents.

Differences from regular non-refreshable materialized views:

  • No insert trigger. I.e. when new data is inserted into the table specified in SELECT, it's not automatically pushed to the refreshable materialized view. The periodic refresh runs the entire query.
  • No restrictions on the SELECT query. Table functions (e.g. url()), views, UNION, JOIN, are all allowed.
Note

The settings in the REFRESH ... SETTINGS part of the query are refresh settings (e.g. refresh_retries), distinct from regular settings (e.g. max_threads). Regular settings can be specified using SETTINGS at the end of the query.

Refresh Schedule

Example refresh schedules:

REFRESH EVERY 1 DAY -- every day, at midnight (UTC)
REFRESH EVERY 1 MONTH -- on 1st day of every month, at midnight
REFRESH EVERY 1 MONTH OFFSET 5 DAY 2 HOUR -- on 6th day of every month, at 2:00 am
REFRESH EVERY 2 WEEK OFFSET 5 DAY 15 HOUR 10 MINUTE -- every other Saturday, at 3:10 pm
REFRESH EVERY 30 MINUTE -- at 00:00, 00:30, 01:00, 01:30, etc
REFRESH AFTER 30 MINUTE -- 30 minutes after the previous refresh completes, no alignment with time of day
-- REFRESH AFTER 1 HOUR OFFSET 1 MINUTE -- syntax errror, OFFSET is not allowed with AFTER
REFRESH EVERY 1 WEEK 2 DAYS -- every 9 days, not on any particular day of the week or month;
-- specifically, when day number (since 1969-12-29) is divisible by 9
REFRESH EVERY 5 MONTHS -- every 5 months, different months each year (as 12 is not divisible by 5);
-- specifically, when month number (since 1970-01) is divisible by 5

RANDOMIZE FOR randomly adjusts the time of each refresh, e.g.:

REFRESH EVERY 1 DAY OFFSET 2 HOUR RANDOMIZE FOR 1 HOUR -- every day at random time between 01:30 and 02:30

At most one refresh may be running at a time, for a given view. E.g. if a view with REFRESH EVERY 1 MINUTE takes 2 minutes to refresh, it'll just be refreshing every 2 minutes. If it then becomes faster and starts refreshing in 10 seconds, it'll go back to refreshing every minute. (In particular, it won't refresh every 10 seconds to catch up with a backlog of missed refreshes - there's no such backlog.)

Additionally, a refresh is started immediately after the materialized view is created, unless EMPTY is specified in the CREATE query. If EMPTY is specified, the first refresh happens according to schedule.

In Replicated DB

If the refreshable materialized view is in a Replicated database, the replicas coordinate with each other such that only one replica performs the refresh at each scheduled time. ReplicatedMergeTree table engine is required, so that all replicas see the data produced by the refresh.

In APPEND mode, coordination can be disabled using SETTINGS all_replicas = 1. This makes replicas do refreshes independently of each other. In this case ReplicatedMergeTree is not required.

In non-APPEND mode, only coordinated refreshing is supported. For uncoordinated, use Atomic database and CREATE ... ON CLUSTER query to create refreshable materialized views on all replicas.

The coordination is done through Keeper. The znode path is determined by default_replica_path server setting.

Dependencies

DEPENDS ON synchronizes refreshes of different tables. By way of example, suppose there's a chain of two refreshable materialized views:

CREATE MATERIALIZED VIEW source REFRESH EVERY 1 DAY AS SELECT * FROM url(...)
CREATE MATERIALIZED VIEW destination REFRESH EVERY 1 DAY AS SELECT ... FROM source

Without DEPENDS ON, both views will start a refresh at midnight, and destination typically will see yesterday's data in source. If we add dependency:

CREATE MATERIALIZED VIEW destination REFRESH EVERY 1 DAY DEPENDS ON source AS SELECT ... FROM source

then destination's refresh will start only after source's refresh finished for that day, so destination will be based on fresh data.

Alternatively, the same result can be achieved with:

CREATE MATERIALIZED VIEW destination REFRESH AFTER 1 HOUR DEPENDS ON source AS SELECT ... FROM source

where 1 HOUR can be any duration less than source's refresh period. The dependent table won't be refreshed more frequently than any of its dependencies. This is a valid way to set up a chain of refreshable views without specifying the real refresh period more than once.

A few more examples:

  • REFRESH EVERY 1 DAY OFFSET 10 MINUTE (destination) depends on REFRESH EVERY 1 DAY (source)
    If source refresh takes more than 10 minutes, destination will wait for it.
  • REFRESH EVERY 1 DAY OFFSET 1 HOUR depends on REFRESH EVERY 1 DAY OFFSET 23 HOUR
    Similar to the above, even though the corresponding refreshes happen on different calendar days. destination's refresh on day X+1 will wait for source's refresh on day X (if it takes more than 2 hours).
  • REFRESH EVERY 2 HOUR depends on REFRESH EVERY 1 HOUR
    The 2 HOUR refresh happens after the 1 HOUR refresh for every other hour, e.g. after the midnight refresh, then after the 2am refresh, etc.
  • REFRESH EVERY 1 MINUTE depends on REFRESH EVERY 2 HOUR
    REFRESH AFTER 1 MINUTE depends on REFRESH EVERY 2 HOUR
    REFRESH AFTER 1 MINUTE depends on REFRESH AFTER 2 HOUR
    destination is refreshed once after every source refresh, i.e. every 2 hours. The 1 MINUTE is effectively ignored.
  • REFRESH AFTER 1 HOUR depends on REFRESH AFTER 1 HOUR
    Currently this is not recommended.
Note

DEPENDS ON only works between refreshable materialized views. Listing a regular table in the DEPENDS ON list will prevent the view from ever refreshing (dependencies can be removed with ALTER, see below).

Settings

Available refresh settings:

  • refresh_retries - How many times to retry if refresh query fails with an exception. If all retries fail, skip to the next scheduled refresh time. 0 means no retries, -1 means infinite retries. Default: 0.
  • refresh_retry_initial_backoff_ms - Delay before the first retry, if refresh_retries is not zero. Each subsequent retry doubles the delay, up to refresh_retry_max_backoff_ms. Default: 100 ms.
  • refresh_retry_max_backoff_ms - Limit on the exponential growth of delay between refresh attempts. Default: 60000 ms (1 minute).

Changing Refresh Parameters

To change refresh parameters:

ALTER TABLE [db.]name MODIFY REFRESH EVERY|AFTER ... [RANDOMIZE FOR ...] [DEPENDS ON ...] [SETTINGS ...]
Note

This replaces all refresh parameters at once: schedule, dependencies, settings, and APPEND-ness. E.g. if the table had a DEPENDS ON, doing a MODIFY REFRESH without DEPENDS ON will remove the dependencies.

Other operations

The status of all refreshable materialized views is available in table system.view_refreshes. In particular, it contains refresh progress (if running), last and next refresh time, exception message if a refresh failed.

To manually stop, start, trigger, or cancel refreshes use SYSTEM STOP|START|REFRESH|CANCEL VIEW.

To wait for a refresh to complete, use SYSTEM WAIT VIEW. In particular, useful for waiting for initial refresh after creating a view.

Note

Fun fact: the refresh query is allowed to read from the view that's being refreshed, seeing pre-refresh version of the data. This means you can implement Conway's game of life: https://pastila.nl/?00021a4b/d6156ff819c83d490ad2dcec05676865#O0LGWTO7maUQIA4AcGUtlA==

Window View [Experimental]

Info

This is an experimental feature that may change in backwards-incompatible ways in the future releases. Enable usage of window views and WATCH query using allow_experimental_window_view setting. Input the command set allow_experimental_window_view = 1.

CREATE WINDOW VIEW [IF NOT EXISTS] [db.]table_name [TO [db.]table_name] [INNER ENGINE engine] [ENGINE engine] [WATERMARK strategy] [ALLOWED_LATENESS interval_function] [POPULATE]
AS SELECT ...
GROUP BY time_window_function
[COMMENT 'comment']

Window view can aggregate data by time window and output the results when the window is ready to fire. It stores the partial aggregation results in an inner(or specified) table to reduce latency and can push the processing result to a specified table or push notifications using the WATCH query.

Creating a window view is similar to creating MATERIALIZED VIEW. Window view needs an inner storage engine to store intermediate data. The inner storage can be specified by using INNER ENGINE clause, the window view will use AggregatingMergeTree as the default inner engine.

When creating a window view without TO [db].[table], you must specify ENGINE – the table engine for storing data.

Time Window Functions

Time window functions are used to get the lower and upper window bound of records. The window view needs to be used with a time window function.

TIME ATTRIBUTES

Window view supports processing time and event time process.

Processing time allows window view to produce results based on the local machine's time and is used by default. It is the most straightforward notion of time but does not provide determinism. The processing time attribute can be defined by setting the time_attr of the time window function to a table column or using the function now(). The following query creates a window view with processing time.

CREATE WINDOW VIEW wv AS SELECT count(number), tumbleStart(w_id) as w_start from date GROUP BY tumble(now(), INTERVAL '5' SECOND) as w_id

Event time is the time that each individual event occurred on its producing device. This time is typically embedded within the records when it is generated. Event time processing allows for consistent results even in case of out-of-order events or late events. Window view supports event time processing by using WATERMARK syntax.

Window view provides three watermark strategies:

  • STRICTLY_ASCENDING: Emits a watermark of the maximum observed timestamp so far. Rows that have a timestamp smaller to the max timestamp are not late.
  • ASCENDING: Emits a watermark of the maximum observed timestamp so far minus 1. Rows that have a timestamp equal and smaller to the max timestamp are not late.
  • BOUNDED: WATERMARK=INTERVAL. Emits watermarks, which are the maximum observed timestamp minus the specified delay.

The following queries are examples of creating a window view with WATERMARK:

CREATE WINDOW VIEW wv WATERMARK=STRICTLY_ASCENDING AS SELECT count(number) FROM date GROUP BY tumble(timestamp, INTERVAL '5' SECOND);
CREATE WINDOW VIEW wv WATERMARK=ASCENDING AS SELECT count(number) FROM date GROUP BY tumble(timestamp, INTERVAL '5' SECOND);
CREATE WINDOW VIEW wv WATERMARK=INTERVAL '3' SECOND AS SELECT count(number) FROM date GROUP BY tumble(timestamp, INTERVAL '5' SECOND);

By default, the window will be fired when the watermark comes, and elements that arrived behind the watermark will be dropped. Window view supports late event processing by setting ALLOWED_LATENESS=INTERVAL. An example of lateness handling is:

CREATE WINDOW VIEW test.wv TO test.dst WATERMARK=ASCENDING ALLOWED_LATENESS=INTERVAL '2' SECOND AS SELECT count(a) AS count, tumbleEnd(wid) AS w_end FROM test.mt GROUP BY tumble(timestamp, INTERVAL '5' SECOND) AS wid;

Note that elements emitted by a late firing should be treated as updated results of a previous computation. Instead of firing at the end of windows, the window view will fire immediately when the late event arrives. Thus, it will result in multiple outputs for the same window. Users need to take these duplicated results into account or deduplicate them.

You can modify SELECT query that was specified in the window view by using ALTER TABLE ... MODIFY QUERY statement. The data structure resulting in a new SELECT query should be the same as the original SELECT query when with or without TO [db.]name clause. Note that the data in the current window will be lost because the intermediate state cannot be reused.

Monitoring New Windows

Window view supports the WATCH query to monitoring changes, or use TO syntax to output the results to a table.

WATCH [db.]window_view
[EVENTS]
[LIMIT n]
[FORMAT format]

WATCH query acts similar as in LIVE VIEW. A LIMIT can be specified to set the number of updates to receive before terminating the query. The EVENTS clause can be used to obtain a short form of the WATCH query where instead of the query result you will just get the latest query watermark.

Settings

  • window_view_clean_interval: The clean interval of window view in seconds to free outdated data. The system will retain the windows that have not been fully triggered according to the system time or WATERMARK configuration, and the other data will be deleted.
  • window_view_heartbeat_interval: The heartbeat interval in seconds to indicate the watch query is alive.
  • wait_for_window_view_fire_signal_timeout: Timeout for waiting for window view fire signal in event time processing.

Example

Suppose we need to count the number of click logs per 10 seconds in a log table called data, and its table structure is:

CREATE TABLE data ( `id` UInt64, `timestamp` DateTime) ENGINE = Memory;

First, we create a window view with tumble window of 10 seconds interval:

CREATE WINDOW VIEW wv as select count(id), tumbleStart(w_id) as window_start from data group by tumble(timestamp, INTERVAL '10' SECOND) as w_id

Then, we use the WATCH query to get the results.

WATCH wv

When logs are inserted into table data,

INSERT INTO data VALUES(1,now())

The WATCH query should print the results as follows:

┌─count(id)─┬────────window_start─┐
│ 1 │ 2020-01-14 16:56:40 │
└───────────┴─────────────────────┘

Alternatively, we can attach the output to another table using TO syntax.

CREATE WINDOW VIEW wv TO dst AS SELECT count(id), tumbleStart(w_id) as window_start FROM data GROUP BY tumble(timestamp, INTERVAL '10' SECOND) as w_id

Additional examples can be found among stateful tests of ClickHouse (they are named *window_view* there).

Window View Usage

The window view is useful in the following scenarios:

  • Monitoring: Aggregate and calculate the metrics logs by time, and output the results to a target table. The dashboard can use the target table as a source table.
  • Analyzing: Automatically aggregate and preprocess data in the time window. This can be useful when analyzing a large number of logs. The preprocessing eliminates repeated calculations in multiple queries and reduces query latency.