SELECT Queries Syntax

SELECT performs data retrieval.

[WITH expr_list|(subquery)]
[FROM [db.]table | (subquery) | table_function] [FINAL]
[SAMPLE sample_coeff]
[WHERE expr]
[GROUP BY expr_list] [WITH TOTALS]
[HAVING expr]
[ORDER BY expr_list]
[LIMIT [offset_value, ]n BY columns]
[LIMIT [n, ]m]
[INTO OUTFILE filename]
[FORMAT format]

All the clauses are optional, except for the required list of expressions immediately after SELECT. The clauses below are described in almost the same order as in the query execution conveyor.

If the query omits the DISTINCT, GROUP BY and ORDER BY clauses and the IN and JOIN subqueries, the query will be completely stream processed, using O(1) amount of RAM. Otherwise, the query might consume a lot of RAM if the appropriate restrictions are not specified: max_memory_usage, max_rows_to_group_by, max_rows_to_sort, max_rows_in_distinct, max_bytes_in_distinct, max_rows_in_set, max_bytes_in_set, max_rows_in_join, max_bytes_in_join, max_bytes_before_external_sort, max_bytes_before_external_group_by. For more information, see the section "Settings". It is possible to use external sorting (saving temporary tables to a disk) and external aggregation. The system does not have "merge join".

WITH Clause

This section provides support for Common Table Expressions (CTE), with some limitations: 1. Recursive queries are not supported 2. When subquery is used inside WITH section, it's result should be scalar with exactly one row 3. Expression's results are not available in subqueries Results of WITH clause expressions can be used inside SELECT clause.

Example 1: Using constant expression as "variable"

WITH '2019-08-01 15:23:00' as ts_upper_bound
FROM hits
    EventDate = toDate(ts_upper_bound) AND
    EventTime <= ts_upper_bound

Example 2: Evicting sum(bytes) expression result from SELECT clause column list

WITH sum(bytes) as s
GROUP BY table

Example 3: Using results of scalar subquery

/* this example would return TOP 10 of most huge tables */
        SELECT sum(bytes)
        WHERE active
    ) AS total_disk_usage
    (sum(bytes) / total_disk_usage) * 100 AS table_disk_usage,
GROUP BY table
ORDER BY table_disk_usage DESC

Example 4: Re-using expression in subquery As a workaround for current limitation for expression usage in subqueries, you may duplicate it.

WITH ['hello'] AS hello
    WITH ['hello'] AS hello
    SELECT hello
│ ['hello'] │ ['hello'] │

FROM Clause

If the FROM clause is omitted, data will be read from the table. The table contains exactly one row (this table fulfills the same purpose as the DUAL table found in other DBMSs).

The FROM clause specifies the source to read data from:

ARRAY JOIN and the regular JOIN may also be included (see below).

Instead of a table, the SELECT subquery may be specified in parenthesis. In contrast to standard SQL, a synonym does not need to be specified after a subquery.

To execute a query, all the columns listed in the query are extracted from the appropriate table. Any columns not needed for the external query are thrown out of the subqueries. If a query does not list any columns (for example, SELECT count() FROM t), some column is extracted from the table anyway (the smallest one is preferred), in order to calculate the number of rows.

FINAL Modifier

Applicable when selecting data from tables from the MergeTree-engine family other than GraphiteMergeTree. When FINAL is specified, ClickHouse fully merges the data before returning the result and thus performs all data transformations that happen during merges for the given table engine.

Also supported for: - Replicated versions of MergeTree engines. - View, Buffer, Distributed, and MaterializedView engines that operate over other engines, provided they were created over MergeTree-engine tables.

Queries that use FINAL are executed not as fast as similar queries that don't, because:

  • Query is executed in a single thread and data is merged during query execution.
  • Queries with FINAL read primary key columns in addition to the columns specified in the query.

In most cases, avoid using FINAL.


The SAMPLE clause allows for approximated query processing.

When data sampling is enabled, the query is not performed on all the data, but only on a certain fraction of data (sample). For example, if you need to calculate statistics for all the visits, it is enough to execute the query on the 1/10 fraction of all the visits and then multiply the result by 10.

Approximated query processing can be useful in the following cases:

  • When you have strict timing requirements (like <100ms) but you can't justify the cost of additional hardware resources to meet them.
  • When your raw data is not accurate, so approximation doesn't noticeably degrade the quality.
  • Business requirements target approximate results (for cost-effectiveness, or in order to market exact results to premium users).


You can only use sampling with the tables in the MergeTree family, and only if the sampling expression was specified during table creation (see MergeTree engine).

The features of data sampling are listed below:

  • Data sampling is a deterministic mechanism. The result of the same SELECT .. SAMPLE query is always the same.
  • Sampling works consistently for different tables. For tables with a single sampling key, a sample with the same coefficient always selects the same subset of possible data. For example, a sample of user IDs takes rows with the same subset of all the possible user IDs from different tables. This means that you can use the sample in subqueries in the IN clause. Also, you can join samples using the JOIN clause.
  • Sampling allows reading less data from a disk. Note that you must specify the sampling key correctly. For more information, see Creating a MergeTree Table.

For the SAMPLE clause the following syntax is supported:

SAMPLE Clause Syntax Description
SAMPLE k Here k is the number from 0 to 1.
The query is executed on k fraction of data. For example, SAMPLE 0.1 runs the query on 10% of data. Read more
SAMPLE n Here n is a sufficiently large integer.
The query is executed on a sample of at least n rows (but not significantly more than this). For example, SAMPLE 10000000 runs the query on a minimum of 10,000,000 rows. Read more
SAMPLE k OFFSET m Here k and m are the numbers from 0 to 1.
The query is executed on a sample of k fraction of the data. The data used for the sample is offset by m fraction. Read more


Here k is the number from 0 to 1 (both fractional and decimal notations are supported). For example, SAMPLE 1/2 or SAMPLE 0.5.

In a SAMPLE k clause, the sample is taken from the k fraction of data. The example is shown below:

    count() * 10 AS PageViews
FROM hits_distributed
    CounterID = 34

In this example, the query is executed on a sample from 0.1 (10%) of data. Values of aggregate functions are not corrected automatically, so to get an approximate result, the value count() is manually multiplied by 10.


Here n is a sufficiently large integer. For example, SAMPLE 10000000.

In this case, the query is executed on a sample of at least n rows (but not significantly more than this). For example, SAMPLE 10000000 runs the query on a minimum of 10,000,000 rows.

Since the minimum unit for data reading is one granule (its size is set by the index_granularity setting), it makes sense to set a sample that is much larger than the size of the granule.

When using the SAMPLE n clause, you don't know which relative percent of data was processed. So you don't know the coefficient the aggregate functions should be multiplied by. Use the _sample_factor virtual column to get the approximate result.

The _sample_factor column contains relative coefficients that are calculated dynamically. This column is created automatically when you create a table with the specified sampling key. The usage examples of the _sample_factor column are shown below.

Let's consider the table visits, which contains the statistics about site visits. The first example shows how to calculate the number of page views:

SELECT sum(PageViews * _sample_factor)
FROM visits
SAMPLE 10000000

The next example shows how to calculate the total number of visits:

SELECT sum(_sample_factor)
FROM visits
SAMPLE 10000000

The example below shows how to calculate the average session duration. Note that you don't need to use the relative coefficient to calculate the average values.

SELECT avg(Duration)
FROM visits
SAMPLE 10000000


Here k and m are numbers from 0 to 1. Examples are shown below.

Example 1


In this example, the sample is 1/10th of all data:


Example 2


Here, a sample of 10% is taken from the second half of the data.



Allows executing JOIN with an array or nested data structure. The intent is similar to the arrayJoin function, but its functionality is broader.

SELECT <expr_list>
FROM <left_subquery>

You can specify only a single ARRAY JOIN clause in a query.

The query execution order is optimized when running ARRAY JOIN. Although ARRAY JOIN must always be specified before the WHERE/PREWHERE clause, it can be performed either before WHERE/PREWHERE (if the result is needed in this clause), or after completing it (to reduce the volume of calculations). The processing order is controlled by the query optimizer.

Supported types of ARRAY JOIN are listed below:

  • ARRAY JOIN - In this case, empty arrays are not included in the result of JOIN.
  • LEFT ARRAY JOIN - The result of JOIN contains rows with empty arrays. The value for an empty array is set to the default value for the array element type (usually 0, empty string or NULL).

The examples below demonstrate the usage of the ARRAY JOIN and LEFT ARRAY JOIN clauses. Let's create a table with an Array type column and insert values into it:

CREATE TABLE arrays_test
    s String,
    arr Array(UInt8)
) ENGINE = Memory;

INSERT INTO arrays_test
VALUES ('Hello', [1,2]), ('World', [3,4,5]), ('Goodbye', []);
│ Hello       │ [1,2]   │
│ World       │ [3,4,5] │
│ Goodbye     │ []      │

The example below uses the ARRAY JOIN clause:

SELECT s, arr
FROM arrays_test
│ Hello │   1 │
│ Hello │   2 │
│ World │   3 │
│ World │   4 │
│ World │   5 │

The next example uses the LEFT ARRAY JOIN clause:

SELECT s, arr
FROM arrays_test
│ Hello       │   1 │
│ Hello       │   2 │
│ World       │   3 │
│ World       │   4 │
│ World       │   5 │
│ Goodbye     │   0 │

Using Aliases

An alias can be specified for an array in the ARRAY JOIN clause. In this case, an array item can be accessed by this alias, but the array itself is accessed by the original name. Example:

SELECT s, arr, a
FROM arrays_test
│ Hello │ [1,2]   │ 1 │
│ Hello │ [1,2]   │ 2 │
│ World │ [3,4,5] │ 3 │
│ World │ [3,4,5] │ 4 │
│ World │ [3,4,5] │ 5 │

Using aliases, you can perform ARRAY JOIN with an external array. For example:

SELECT s, arr_external
FROM arrays_test
ARRAY JOIN [1, 2, 3] AS arr_external;
│ Hello       │            1 │
│ Hello       │            2 │
│ Hello       │            3 │
│ World       │            1 │
│ World       │            2 │
│ World       │            3 │
│ Goodbye     │            1 │
│ Goodbye     │            2 │
│ Goodbye     │            3 │

Multiple arrays can be comma-separated in the ARRAY JOIN clause. In this case, JOIN is performed with them simultaneously (the direct sum, not the cartesian product). Note that all the arrays must have the same size. Example:

SELECT s, arr, a, num, mapped
FROM arrays_test
ARRAY JOIN arr AS a, arrayEnumerate(arr) AS num, arrayMap(x -> x + 1, arr) AS mapped;
│ Hello │ [1,2]   │ 1 │   1 │      2 │
│ Hello │ [1,2]   │ 2 │   2 │      3 │
│ World │ [3,4,5] │ 3 │   1 │      4 │
│ World │ [3,4,5] │ 4 │   2 │      5 │
│ World │ [3,4,5] │ 5 │   3 │      6 │

The example below uses the arrayEnumerate function:

SELECT s, arr, a, num, arrayEnumerate(arr)
FROM arrays_test
ARRAY JOIN arr AS a, arrayEnumerate(arr) AS num;
│ Hello │ [1,2]   │ 1 │   1 │ [1,2]               │
│ Hello │ [1,2]   │ 2 │   2 │ [1,2]               │
│ World │ [3,4,5] │ 3 │   1 │ [1,2,3]             │
│ World │ [3,4,5] │ 4 │   2 │ [1,2,3]             │
│ World │ [3,4,5] │ 5 │   3 │ [1,2,3]             │

ARRAY JOIN With Nested Data Structure

ARRAYJOIN`` also works with nested data structures. Example:

CREATE TABLE nested_test
    s String,
    nest Nested(
    x UInt8,
    y UInt32)
) ENGINE = Memory;

INSERT INTO nested_test
VALUES ('Hello', [1,2], [10,20]), ('World', [3,4,5], [30,40,50]), ('Goodbye', [], []);
│ Hello   │ [1,2]   │ [10,20]    │
│ World   │ [3,4,5] │ [30,40,50] │
│ Goodbye │ []      │ []         │
SELECT s, `nest.x`, `nest.y`
FROM nested_test
│ Hello │      1 │     10 │
│ Hello │      2 │     20 │
│ World │      3 │     30 │
│ World │      4 │     40 │
│ World │      5 │     50 │

When specifying names of nested data structures in ARRAY JOIN, the meaning is the same as ARRAY JOIN with all the array elements that it consists of. Examples are listed below:

SELECT s, `nest.x`, `nest.y`
FROM nested_test
ARRAY JOIN `nest.x`, `nest.y`;
│ Hello │      1 │     10 │
│ Hello │      2 │     20 │
│ World │      3 │     30 │
│ World │      4 │     40 │
│ World │      5 │     50 │

This variation also makes sense:

SELECT s, `nest.x`, `nest.y`
FROM nested_test
ARRAY JOIN `nest.x`;
│ Hello │      1 │ [10,20]    │
│ Hello │      2 │ [10,20]    │
│ World │      3 │ [30,40,50] │
│ World │      4 │ [30,40,50] │
│ World │      5 │ [30,40,50] │

An alias may be used for a nested data structure, in order to select either the JOIN result or the source array. Example:

SELECT s, `n.x`, `n.y`, `nest.x`, `nest.y`
FROM nested_test
│ Hello │   1 │  10 │ [1,2]   │ [10,20]    │
│ Hello │   2 │  20 │ [1,2]   │ [10,20]    │
│ World │   3 │  30 │ [3,4,5] │ [30,40,50] │
│ World │   4 │  40 │ [3,4,5] │ [30,40,50] │
│ World │   5 │  50 │ [3,4,5] │ [30,40,50] │

Example of using the arrayEnumerate function:

SELECT s, `n.x`, `n.y`, `nest.x`, `nest.y`, num
FROM nested_test
ARRAY JOIN nest AS n, arrayEnumerate(`nest.x`) AS num;
│ Hello │   1 │  10 │ [1,2]   │ [10,20]    │   1 │
│ Hello │   2 │  20 │ [1,2]   │ [10,20]    │   2 │
│ World │   3 │  30 │ [3,4,5] │ [30,40,50] │   1 │
│ World │   4 │  40 │ [3,4,5] │ [30,40,50] │   2 │
│ World │   5 │  50 │ [3,4,5] │ [30,40,50] │   3 │

JOIN Clause

Joins the data in the normal SQL JOIN sense.


Not related to ARRAY JOIN.

SELECT <expr_list>
FROM <left_subquery>
(ON <expr_list>)|(USING <column_list>) ...

The table names can be specified instead of <left_subquery> and <right_subquery>. This is equivalent to the SELECT * FROM table subquery, except in a special case when the table has the Join engine – an array prepared for joining.

Supported Types of JOIN

  • CROSS JOIN (or , )

See the standard SQL JOIN description.

Multiple JOIN

Performing queries, ClickHouse rewrites multi-table joins into the sequence of two-table joins. For example, if there are four tables for join ClickHouse joins the first and the second, then joins the result with the third table, and at the last step, it joins the fourth one.

If a query contains the WHERE clause, ClickHouse tries to pushdown filters from this clause through the intermediate join. If it cannot apply the filter to each intermediate join, ClickHouse applies the filters after all joins are completed.

We recommend the JOIN ON or JOIN USING syntax for creating queries. For example:

SELECT * FROM t1 JOIN t2 ON t1.a = t2.a JOIN t3 ON t1.a = t3.a

You can use comma-separated lists of tables in the FROM clause. This works only with the allow_experimental_cross_to_join_conversion = 1 setting. For example:

SELECT * FROM t1, t2, t3 WHERE t1.a = t2.a AND t1.a = t3.a

Don't mix these syntaxes.

ClickHouse doesn't directly support syntax with commas, so we don't recommend using them. The algorithm tries to rewrite the query in terms of CROSS JOIN and INNER JOIN clauses and then proceeds to query processing. When rewriting the query, ClickHouse tries to optimize performance and memory consumption. By default, ClickHouse treats commas as an INNER JOIN clause and converts INNER JOIN to CROSS JOIN when the algorithm cannot guarantee that INNER JOIN returns the required data.


  • ALL — If the right table has several matching rows, ClickHouse creates a Cartesian product from matching rows. This is the standard JOIN behavior in SQL.
  • ANY — If the right table has several matching rows, only the first one found is joined. If the right table has only one matching row, the results of queries with ANY and ALL keywords are the same.
  • ASOF — For joining sequences with a non-exact match. ASOF JOIN usage is described below.


ASOF JOIN is useful when you need to join records that have no exact match.

Tables for ASOF JOIN must have an ordered sequence column. This column cannot be alone in a table, and should be one of the data types: UInt32, UInt64, Float32, Float64, Date, and DateTime.

Syntax ASOF JOIN ... ON:

SELECT expressions_list
FROM table_1
ON equi_cond AND closest_match_cond

You can use any number of equality conditions and exactly one closest match condition. For example, SELECT count() FROM table_1 ASOF LEFT JOIN table_2 ON table_1.a == table_2.b AND table_2.t <= table_1.t.

Conditions supported for the closest match: >, >=, <, <=.


SELECT expressions_list
FROM table_1
ASOF JOIN table_2
USING (equi_column1, ... equi_columnN, asof_column)

ASOF JOIN uses equi_columnX for joining on equality and asof_column for joining on the closest match with the table_1.asof_column >= table_2.asof_column condition. The asof_column column always the last one in the USING clause.

For example, consider the following tables:

     table_1                           table_2

  event   | ev_time | user_id       event   | ev_time | user_id
----------|---------|----------   ----------|---------|----------             
              ...                               ...
event_1_1 |  12:00  |  42         event_2_1 |  11:59  |   42
              ...                 event_2_2 |  12:30  |   42
event_1_2 |  13:00  |  42         event_2_3 |  13:00  |   42
              ...                               ...

ASOF JOIN can take the timestamp of a user event from table_1 and find an event in table_2 where the timestamp is closest to the timestamp of the event from table_1 corresponding to the closest match condition. Equal timestamp values are the closest if available. Here, the user_id column can be used for joining on equality and the ev_time column can be used for joining on the closest match. In our example, event_1_1 can be joined with event_2_1 and event_1_2 can be joined with event_2_3, but event_2_2 can't be joined.


ASOF join is not supported in the Join table engine.

To set the default strictness value, use the session configuration parameter join_default_strictness.


When using a normal JOIN, the query is sent to remote servers. Subqueries are run on each of them in order to make the right table, and the join is performed with this table. In other words, the right table is formed on each server separately.

When using GLOBAL ... JOIN, first the requestor server runs a subquery to calculate the right table. This temporary table is passed to each remote server, and queries are run on them using the temporary data that was transmitted.

Be careful when using GLOBAL. For more information, see the section Distributed subqueries.

Usage Recommendations

When running a JOIN, there is no optimization of the order of execution in relation to other stages of the query. The join (a search in the right table) is run before filtering in WHERE and before aggregation. In order to explicitly set the processing order, we recommend running a JOIN subquery with a subquery.


        count() AS hits
    FROM test.hits
    GROUP BY CounterID
        sum(Sign) AS visits
    FROM test.visits
    GROUP BY CounterID
) USING CounterID
│   1143050 │ 523264 │  13665 │
│    731962 │ 475698 │ 102716 │
│    722545 │ 337212 │ 108187 │
│    722889 │ 252197 │  10547 │
│   2237260 │ 196036 │   9522 │
│  23057320 │ 147211 │   7689 │
│    722818 │  90109 │  17847 │
│     48221 │  85379 │   4652 │
│  19762435 │  77807 │   7026 │
│    722884 │  77492 │  11056 │

Subqueries don't allow you to set names or use them for referencing a column from a specific subquery. The columns specified in USING must have the same names in both subqueries, and the other columns must be named differently. You can use aliases to change the names of columns in subqueries (the example uses the aliases hits and visits).

The USING clause specifies one or more columns to join, which establishes the equality of these columns. The list of columns is set without brackets. More complex join conditions are not supported.

The right table (the subquery result) resides in RAM. If there isn't enough memory, you can't run a JOIN.

Each time a query is run with the same JOIN, the subquery is run again because the result is not cached. To avoid this, use the special Join table engine, which is a prepared array for joining that is always in RAM.

In some cases, it is more efficient to use IN instead of JOIN. Among the various types of JOIN, the most efficient is ANY LEFT JOIN, then ANY INNER JOIN. The least efficient are ALL LEFT JOIN and ALL INNER JOIN.

If you need a JOIN for joining with dimension tables (these are relatively small tables that contain dimension properties, such as names for advertising campaigns), a JOIN might not be very convenient due to the fact that the right table is re-accessed for every query. For such cases, there is an "external dictionaries" feature that you should use instead of JOIN. For more information, see the section External dictionaries.

Memory Limitations

ClickHouse uses the hash join algorithm. ClickHouse takes the <right_subquery> and creates a hash table for it in RAM. If you need to restrict join operation memory consumption use the following settings:

When any of these limits is reached, ClickHouse acts as the join_overflow_mode setting instructs.

Processing of Empty or NULL Cells

While joining tables, the empty cells may appear. The setting join_use_nulls define how ClickHouse fills these cells.

If the JOIN keys are Nullable fields, the rows where at least one of the keys has the value NULL are not joined.

Syntax Limitations

For multiple JOIN clauses in a single SELECT query:

  • Taking all the columns via * is available only if tables are joined, not subqueries.
  • The PREWHERE clause is not available.

For ON, WHERE, and GROUP BY clauses:

  • Arbitrary expressions cannot be used in ON, WHERE, and GROUP BY clauses, but you can define an expression in a SELECT clause and then use it in these clauses via an alias.

WHERE Clause

If there is a WHERE clause, it must contain an expression with the UInt8 type. This is usually an expression with comparison and logical operators. This expression will be used for filtering data before all other transformations.

If indexes are supported by the database table engine, the expression is evaluated on the ability to use indexes.


This clause has the same meaning as the WHERE clause. The difference is in which data is read from the table. When using PREWHERE, first only the columns necessary for executing PREWHERE are read. Then the other columns are read that are needed for running the query, but only those blocks where the PREWHERE expression is true.

It makes sense to use PREWHERE if there are filtration conditions that are used by a minority of the columns in the query, but that provide strong data filtration. This reduces the volume of data to read.

For example, it is useful to write PREWHERE for queries that extract a large number of columns, but that only have filtration for a few columns.

PREWHERE is only supported by tables from the *MergeTree family.

A query may simultaneously specify PREWHERE and WHERE. In this case, PREWHERE precedes WHERE.

If the 'optimize_move_to_prewhere' setting is set to 1 and PREWHERE is omitted, the system uses heuristics to automatically move parts of expressions from WHERE to PREWHERE.


This is one of the most important parts of a column-oriented DBMS.

If there is a GROUP BY clause, it must contain a list of expressions. Each expression will be referred to here as a "key". All the expressions in the SELECT, HAVING, and ORDER BY clauses must be calculated from keys or from aggregate functions. In other words, each column selected from the table must be used either in keys or inside aggregate functions.

If a query contains only table columns inside aggregate functions, the GROUP BY clause can be omitted, and aggregation by an empty set of keys is assumed.


    median(FetchTiming > 60 ? 60 : FetchTiming),
    count() - sum(Refresh)
FROM hits

However, in contrast to standard SQL, if the table doesn't have any rows (either there aren't any at all, or there aren't any after using WHERE to filter), an empty result is returned, and not the result from one of the rows containing the initial values of aggregate functions.

As opposed to MySQL (and conforming to standard SQL), you can't get some value of some column that is not in a key or aggregate function (except constant expressions). To work around this, you can use the 'any' aggregate function (get the first encountered value) or 'min/max'.


    domainWithoutWWW(URL) AS domain,
    any(Title) AS title -- getting the first occurred page header for each domain.
FROM hits
GROUP BY domain

For every different key value encountered, GROUP BY calculates a set of aggregate function values.

GROUP BY is not supported for array columns.

A constant can't be specified as arguments for aggregate functions. Example: sum(1). Instead of this, you can get rid of the constant. Example: count().

NULL processing

For grouping, ClickHouse interprets NULL as a value, and NULL=NULL.

Here's an example to show what this means.

Assume you have this table:

│ 1 │    2 │
│ 2 │ ᴺᵁᴸᴸ │
│ 3 │    2 │
│ 3 │    3 │
│ 3 │ ᴺᵁᴸᴸ │

The query SELECT sum(x), y FROM t_null_big GROUP BY y results in:

│      4 │    2 │
│      3 │    3 │
│      5 │ ᴺᵁᴸᴸ │

You can see that GROUP BY for y = NULL summed up x, as if NULL is this value.

If you pass several keys to GROUP BY, the result will give you all the combinations of the selection, as if NULL were a specific value.


If the WITH TOTALS modifier is specified, another row will be calculated. This row will have key columns containing default values (zeros or empty lines), and columns of aggregate functions with the values calculated across all the rows (the "total" values).

This extra row is output in JSON*, TabSeparated*, and Pretty* formats, separately from the other rows. In the other formats, this row is not output.

In JSON* formats, this row is output as a separate 'totals' field. In TabSeparated* formats, the row comes after the main result, preceded by an empty row (after the other data). In Pretty* formats, the row is output as a separate table after the main result.

WITH TOTALS can be run in different ways when HAVING is present. The behavior depends on the 'totals_mode' setting. By default, totals_mode = 'before_having'. In this case, 'totals' is calculated across all rows, including the ones that don't pass through HAVING and 'max_rows_to_group_by'.

The other alternatives include only the rows that pass through HAVING in 'totals', and behave differently with the setting max_rows_to_group_by and group_by_overflow_mode = 'any'.

after_having_exclusive – Don't include rows that didn't pass through max_rows_to_group_by. In other words, 'totals' will have less than or the same number of rows as it would if max_rows_to_group_by were omitted.

after_having_inclusive – Include all the rows that didn't pass through 'max_rows_to_group_by' in 'totals'. In other words, 'totals' will have more than or the same number of rows as it would if max_rows_to_group_by were omitted.

after_having_auto – Count the number of rows that passed through HAVING. If it is more than a certain amount (by default, 50%), include all the rows that didn't pass through 'max_rows_to_group_by' in 'totals'. Otherwise, do not include them.

totals_auto_threshold – By default, 0.5. The coefficient for after_having_auto.

If max_rows_to_group_by and group_by_overflow_mode = 'any' are not used, all variations of after_having are the same, and you can use any of them (for example, after_having_auto).

You can use WITH TOTALS in subqueries, including subqueries in the JOIN clause (in this case, the respective total values are combined).

GROUP BY in External Memory

You can enable dumping temporary data to the disk to restrict memory usage during GROUP BY. The max_bytes_before_external_group_by setting determines the threshold RAM consumption for dumping GROUP BY temporary data to the file system. If set to 0 (the default), it is disabled.

When using max_bytes_before_external_group_by, we recommend that you set max_memory_usage about twice as high. This is necessary because there are two stages to aggregation: reading the date and forming intermediate data (1) and merging the intermediate data (2). Dumping data to the file system can only occur during stage 1. If the temporary data wasn't dumped, then stage 2 might require up to the same amount of memory as in stage 1.

For example, if max_memory_usage was set to 10000000000 and you want to use external aggregation, it makes sense to set max_bytes_before_external_group_by to 10000000000, and max_memory_usage to 20000000000. When external aggregation is triggered (if there was at least one dump of temporary data), maximum consumption of RAM is only slightly more than max_bytes_before_external_group_by.

With distributed query processing, external aggregation is performed on remote servers. In order for the requester server to use only a small amount of RAM, set distributed_aggregation_memory_efficient to 1.

When merging data flushed to the disk, as well as when merging results from remote servers when the distributed_aggregation_memory_efficient setting is enabled, consumes up to 1/256 * the_number_of_threads from the total amount of RAM.

When external aggregation is enabled, if there was less than max_bytes_before_external_group_by of data (i.e. data was not flushed), the query runs just as fast as without external aggregation. If any temporary data was flushed, the run time will be several times longer (approximately three times).

If you have an ORDER BY with a LIMIT after GROUP BY, then the amount of used RAM depends on the amount of data in LIMIT, not in the whole table. But if the ORDER BY doesn't have LIMIT, don't forget to enable external sorting (max_bytes_before_external_sort).


A query with the LIMIT n BY expressions clause selects the first n rows for each distinct value of expressions. The key for LIMIT BY can contain any number of expressions.

ClickHouse supports the following syntax:

  • LIMIT [offset_value, ]n BY expressions
  • LIMIT n OFFSET offset_value BY expressions

During query processing, ClickHouse selects data ordered by sorting key. The sorting key is set explicitly using an ORDER BY clause or implicitly as a property of the table engine. Then ClickHouse applies LIMIT n BY expressions and returns the first n rows for each distinct combination of expressions. If OFFSET is specified, then for each data block that belongs to a distinct combination of expressions, ClickHouse skips offset_value number of rows from the beginning of the block and returns a maximum of n rows as a result. If offset_value is bigger than the number of rows in the data block, ClickHouse returns zero rows from the block.

LIMIT BY is not related to LIMIT. They can both be used in the same query.


Sample table:

CREATE TABLE limit_by(id Int, val Int) ENGINE = Memory;
INSERT INTO limit_by values(1, 10), (1, 11), (1, 12), (2, 20), (2, 21);


SELECT * FROM limit_by ORDER BY id, val LIMIT 2 BY id
│  1 │  10 │
│  1 │  11 │
│  2 │  20 │
│  2 │  21 │
SELECT * FROM limit_by ORDER BY id, val LIMIT 1, 2 BY id
│  1 │  11 │
│  1 │  12 │
│  2 │  21 │

The SELECT * FROM limit_by ORDER BY id, val LIMIT 2 OFFSET 1 BY id query returns the same result.

The following query returns the top 5 referrers for each domain, device_type pair with a maximum of 100 rows in total (LIMIT n BY + LIMIT).

    domainWithoutWWW(URL) AS domain,
    domainWithoutWWW(REFERRER_URL) AS referrer,
    count() cnt
FROM hits
GROUP BY domain, referrer, device_type
LIMIT 5 BY domain, device_type


Allows filtering the result received after GROUP BY, similar to the WHERE clause. WHERE and HAVING differ in that WHERE is performed before aggregation (GROUP BY), while HAVING is performed after it. If aggregation is not performed, HAVING can't be used.


The ORDER BY clause contains a list of expressions, which can each be assigned DESC or ASC (the sorting direction). If the direction is not specified, ASC is assumed. ASC is sorted in ascending order, and DESC in descending order. The sorting direction applies to a single expression, not to the entire list. Example: ORDER BY Visits DESC, SearchPhrase

For sorting by String values, you can specify collation (comparison). Example: ORDER BY SearchPhrase COLLATE 'tr' - for sorting by keyword in ascending order, using the Turkish alphabet, case insensitive, assuming that strings are UTF-8 encoded. COLLATE can be specified or not for each expression in ORDER BY independently. If ASC or DESC is specified, COLLATE is specified after it. When using COLLATE, sorting is always case-insensitive.

We only recommend using COLLATE for final sorting of a small number of rows, since sorting with COLLATE is less efficient than normal sorting by bytes.

Rows that have identical values for the list of sorting expressions are output in an arbitrary order, which can also be nondeterministic (different each time). If the ORDER BY clause is omitted, the order of the rows is also undefined, and may be nondeterministic as well.

NaN and NULL sorting order:

  • With the modifier NULLS FIRST — First NULL, then NaN, then other values.
  • With the modifier NULLS LAST — First the values, then NaN, then NULL.
  • Default — The same as with the NULLS LAST modifier.


For the table

│ 1 │ ᴺᵁᴸᴸ │
│ 2 │    2 │
│ 1 │  nan │
│ 2 │    2 │
│ 3 │    4 │
│ 5 │    6 │
│ 6 │  nan │
│ 7 │ ᴺᵁᴸᴸ │
│ 6 │    7 │
│ 8 │    9 │

Run the query SELECT * FROM t_null_nan ORDER BY y NULLS FIRST to get:

│ 1 │ ᴺᵁᴸᴸ │
│ 7 │ ᴺᵁᴸᴸ │
│ 1 │  nan │
│ 6 │  nan │
│ 2 │    2 │
│ 2 │    2 │
│ 3 │    4 │
│ 5 │    6 │
│ 6 │    7 │
│ 8 │    9 │

When floating point numbers are sorted, NaNs are separate from the other values. Regardless of the sorting order, NaNs come at the end. In other words, for ascending sorting they are placed as if they are larger than all the other numbers, while for descending sorting they are placed as if they are smaller than the rest.

Less RAM is used if a small enough LIMIT is specified in addition to ORDER BY. Otherwise, the amount of memory spent is proportional to the volume of data for sorting. For distributed query processing, if GROUP BY is omitted, sorting is partially done on remote servers, and the results are merged on the requestor server. This means that for distributed sorting, the volume of data to sort can be greater than the amount of memory on a single server.

If there is not enough RAM, it is possible to perform sorting in external memory (creating temporary files on a disk). Use the setting max_bytes_before_external_sort for this purpose. If it is set to 0 (the default), external sorting is disabled. If it is enabled, when the volume of data to sort reaches the specified number of bytes, the collected data is sorted and dumped into a temporary file. After all data is read, all the sorted files are merged and the results are output. Files are written to the /var/lib/clickhouse/tmp/ directory in the config (by default, but you can use the 'tmp_path' parameter to change this setting).

Running a query may use more memory than 'max_bytes_before_external_sort'. For this reason, this setting must have a value significantly smaller than 'max_memory_usage'. As an example, if your server has 128 GB of RAM and you need to run a single query, set 'max_memory_usage' to 100 GB, and 'max_bytes_before_external_sort' to 80 GB.

External sorting works much less effectively than sorting in RAM.


Expressions specified in the SELECT clause are calculated after all the operations in the clauses described above are finished. These expressions work as if they apply to separate rows in the result. If expressions in the SELECT clause contain aggregate functions, then ClickHouse processes aggregate functions and expressions used as their arguments during the GROUP BY aggregation.

If you want to include all columns in the result, use the asterisk (*) symbol. For example, SELECT * FROM ....

To match some columns in the result with a re2 regular expression, you can use the COLUMNS expression.


For example, consider the table:

CREATE TABLE default.col_names (aa Int8, ab Int8, bc Int8) ENGINE = TinyLog

The following query selects data from all the columns containing the a symbol in their name.

SELECT COLUMNS('a') FROM col_names
│  1 │  1 │

The selected columns are returned not in the alphabetical order.

You can use multiple COLUMNS expressions in a query and apply functions to them.

For example:

SELECT COLUMNS('a'), COLUMNS('c'), toTypeName(COLUMNS('c')) FROM col_names
│  1 │  1 │  1 │ Int8           │

Each column returned by the COLUMNS expression is passed to the function as a separate argument. Also you can pass other arguments to the function if it supports them. Be careful when using functions. If a function doesn't support the number of arguments you have passed to it, ClickHouse throws an exception.

For example:

SELECT COLUMNS('a') + COLUMNS('c') FROM col_names
Received exception from server (version 19.14.1):
Code: 42. DB::Exception: Received from localhost:9000. DB::Exception: Number of arguments for function plus doesn't match: passed 3, should be 2. 

In this example, COLUMNS('a') returns two columns: aa and ab. COLUMNS('c') returns the bc column. The + operator can't apply to 3 arguments, so ClickHouse throws an exception with the relevant message.

Columns that matched the COLUMNS expression can have different data types. If COLUMNS doesn't match any columns and is the only expression in SELECT, ClickHouse throws an exception.


If DISTINCT is specified, only a single row will remain out of all the sets of fully matching rows in the result. The result will be the same as if GROUP BY were specified across all the fields specified in SELECT without aggregate functions. But there are several differences from GROUP BY:

  • DISTINCT can be applied together with GROUP BY.
  • When ORDER BY is omitted and LIMIT is defined, the query stops running immediately after the required number of different rows has been read.
  • Data blocks are output as they are processed, without waiting for the entire query to finish running.

DISTINCT is not supported if SELECT has at least one array column.

DISTINCT works with NULL as if NULL were a specific value, and NULL=NULL. In other words, in the DISTINCT results, different combinations with NULL only occur once.

ClickHouse supports using the DISTINCT and ORDER BY clauses for different columns in one query. The DISTINCT clause is executed before the ORDER BY clause.

Example table:

│ 2 │ 1 │
│ 1 │ 2 │
│ 3 │ 3 │
│ 2 │ 4 │

When selecting data with the SELECT DISTINCT a FROM t1 ORDER BY b ASC query, we get the following result:

│ 2 │
│ 1 │
│ 3 │

If we change the sorting direction SELECT DISTINCT a FROM t1 ORDER BY b DESC, we get the following result:

│ 3 │
│ 1 │
│ 2 │

Row 2, 4 was cut before sorting.

Take this implementation specificity into account when programming queries.

LIMIT Clause

LIMIT m allows you to select the first m rows from the result.

LIMIT n, m allows you to select the first m rows from the result after skipping the first n rows. The LIMIT m OFFSET n syntax is also supported.

n and m must be non-negative integers.

If there isn't an ORDER BY clause that explicitly sorts results, the result may be arbitrary and nondeterministic.


You can use UNION ALL to combine any number of queries. Example:

SELECT CounterID, 1 AS table, toInt64(count()) AS c
    FROM test.hits
    GROUP BY CounterID


SELECT CounterID, 2 AS table, sum(Sign) AS c
    FROM test.visits
    GROUP BY CounterID
    HAVING c > 0

Only UNION ALL is supported. The regular UNION (UNION DISTINCT) is not supported. If you need UNION DISTINCT, you can write SELECT DISTINCT from a subquery containing UNION ALL.

Queries that are parts of UNION ALL can be run simultaneously, and their results can be mixed together.

The structure of results (the number and type of columns) must match for the queries. But the column names can differ. In this case, the column names for the final result will be taken from the first query. Type casting is performed for unions. For example, if two queries being combined have the same field with non-Nullable and Nullable types from a compatible type, the resulting UNION ALL has a Nullable type field.

Queries that are parts of UNION ALL can't be enclosed in brackets. ORDER BY and LIMIT are applied to separate queries, not to the final result. If you need to apply a conversion to the final result, you can put all the queries with UNION ALL in a subquery in the FROM clause.


Add the INTO OUTFILE filename clause (where filename is a string literal) to redirect query output to the specified file. In contrast to MySQL, the file is created on the client side. The query will fail if a file with the same filename already exists. This functionality is available in the command-line client and clickhouse-local (a query sent via HTTP interface will fail).

The default output format is TabSeparated (the same as in the command-line client batch mode).


Specify 'FORMAT format' to get data in any specified format. You can use this for convenience, or for creating dumps. For more information, see the section "Formats". If the FORMAT clause is omitted, the default format is used, which depends on both the settings and the interface used for accessing the DB. For the HTTP interface and the command-line client in batch mode, the default format is TabSeparated. For the command-line client in interactive mode, the default format is PrettyCompact (it has attractive and compact tables).

When using the command-line client, data is passed to the client in an internal efficient format. The client independently interprets the FORMAT clause of the query and formats the data itself (thus relieving the network and the server from the load).

IN Operators

The IN, NOT IN, GLOBAL IN, and GLOBAL NOT IN operators are covered separately, since their functionality is quite rich.

The left side of the operator is either a single column or a tuple.


SELECT UserID IN (123, 456) FROM ...
SELECT (CounterID, UserID) IN ((34, 123), (101500, 456)) FROM ...

If the left side is a single column that is in the index, and the right side is a set of constants, the system uses the index for processing the query.

Don't list too many values explicitly (i.e. millions). If a data set is large, put it in a temporary table (for example, see the section "External data for query processing"), then use a subquery.

The right side of the operator can be a set of constant expressions, a set of tuples with constant expressions (shown in the examples above), or the name of a database table or SELECT subquery in brackets.

If the right side of the operator is the name of a table (for example, UserID IN users), this is equivalent to the subquery UserID IN (SELECT * FROM users). Use this when working with external data that is sent along with the query. For example, the query can be sent together with a set of user IDs loaded to the 'users' temporary table, which should be filtered.

If the right side of the operator is a table name that has the Set engine (a prepared data set that is always in RAM), the data set will not be created over again for each query.

The subquery may specify more than one column for filtering tuples. Example:

SELECT (CounterID, UserID) IN (SELECT CounterID, UserID FROM ...) FROM ...

The columns to the left and right of the IN operator should have the same type.

The IN operator and subquery may occur in any part of the query, including in aggregate functions and lambda functions. Example:

    avg(UserID IN
        SELECT UserID
        FROM test.hits
        WHERE EventDate = toDate('2014-03-17')
    )) AS ratio
FROM test.hits
GROUP BY EventDate
│ 2014-03-17 │        1 │
│ 2014-03-18 │ 0.807696 │
│ 2014-03-19 │ 0.755406 │
│ 2014-03-20 │ 0.723218 │
│ 2014-03-21 │ 0.697021 │
│ 2014-03-22 │ 0.647851 │
│ 2014-03-23 │ 0.648416 │

For each day after March 17th, count the percentage of pageviews made by users who visited the site on March 17th. A subquery in the IN clause is always run just one time on a single server. There are no dependent subqueries.

NULL processing

During request processing, the IN operator assumes that the result of an operation with NULL is always equal to 0, regardless of whether NULL is on the right or left side of the operator. NULL values are not included in any dataset, do not correspond to each other and cannot be compared.

Here is an example with the t_null table:

│ 1 │ ᴺᵁᴸᴸ │
│ 2 │    3 │

Running the query SELECT x FROM t_null WHERE y IN (NULL,3) gives you the following result:

│ 2 │

You can see that the row in which y = NULL is thrown out of the query results. This is because ClickHouse can't decide whether NULL is included in the (NULL,3) set, returns 0 as the result of the operation, and SELECT excludes this row from the final output.

FROM t_null
┌─in(y, tuple(NULL, 3))─┐
│                     0 │
│                     1 │

Distributed Subqueries

There are two options for IN-s with subqueries (similar to JOINs): normal IN / JOIN and GLOBAL IN / GLOBAL JOIN. They differ in how they are run for distributed query processing.


Remember that the algorithms described below may work differently depending on the settings distributed_product_mode setting.

When using the regular IN, the query is sent to remote servers, and each of them runs the subqueries in the IN or JOIN clause.

When using GLOBAL IN / GLOBAL JOINs, first all the subqueries are run for GLOBAL IN / GLOBAL JOINs, and the results are collected in temporary tables. Then the temporary tables are sent to each remote server, where the queries are run using this temporary data.

For a non-distributed query, use the regular IN / JOIN.

Be careful when using subqueries in the IN / JOIN clauses for distributed query processing.

Let's look at some examples. Assume that each server in the cluster has a normal local_table. Each server also has a distributed_table table with the Distributed type, which looks at all the servers in the cluster.

For a query to the distributed_table, the query will be sent to all the remote servers and run on them using the local_table.

For example, the query

SELECT uniq(UserID) FROM distributed_table

will be sent to all remote servers as

SELECT uniq(UserID) FROM local_table

and run on each of them in parallel, until it reaches the stage where intermediate results can be combined. Then the intermediate results will be returned to the requestor server and merged on it, and the final result will be sent to the client.

Now let's examine a query with IN:

SELECT uniq(UserID) FROM distributed_table WHERE CounterID = 101500 AND UserID IN (SELECT UserID FROM local_table WHERE CounterID = 34)
  • Calculation of the intersection of audiences of two sites.

This query will be sent to all remote servers as

SELECT uniq(UserID) FROM local_table WHERE CounterID = 101500 AND UserID IN (SELECT UserID FROM local_table WHERE CounterID = 34)

In other words, the data set in the IN clause will be collected on each server independently, only across the data that is stored locally on each of the servers.

This will work correctly and optimally if you are prepared for this case and have spread data across the cluster servers such that the data for a single UserID resides entirely on a single server. In this case, all the necessary data will be available locally on each server. Otherwise, the result will be inaccurate. We refer to this variation of the query as "local IN".

To correct how the query works when data is spread randomly across the cluster servers, you could specify distributed_table inside a subquery. The query would look like this:

SELECT uniq(UserID) FROM distributed_table WHERE CounterID = 101500 AND UserID IN (SELECT UserID FROM distributed_table WHERE CounterID = 34)

This query will be sent to all remote servers as

SELECT uniq(UserID) FROM local_table WHERE CounterID = 101500 AND UserID IN (SELECT UserID FROM distributed_table WHERE CounterID = 34)

The subquery will begin running on each remote server. Since the subquery uses a distributed table, the subquery that is on each remote server will be resent to every remote server as

SELECT UserID FROM local_table WHERE CounterID = 34

For example, if you have a cluster of 100 servers, executing the entire query will require 10,000 elementary requests, which is generally considered unacceptable.

In such cases, you should always use GLOBAL IN instead of IN. Let's look at how it works for the query

SELECT uniq(UserID) FROM distributed_table WHERE CounterID = 101500 AND UserID GLOBAL IN (SELECT UserID FROM distributed_table WHERE CounterID = 34)

The requestor server will run the subquery

SELECT UserID FROM distributed_table WHERE CounterID = 34

and the result will be put in a temporary table in RAM. Then the request will be sent to each remote server as

SELECT uniq(UserID) FROM local_table WHERE CounterID = 101500 AND UserID GLOBAL IN _data1

and the temporary table _data1 will be sent to every remote server with the query (the name of the temporary table is implementation-defined).

This is more optimal than using the normal IN. However, keep the following points in mind:

  1. When creating a temporary table, data is not made unique. To reduce the volume of data transmitted over the network, specify DISTINCT in the subquery. (You don't need to do this for a normal IN.)
  2. The temporary table will be sent to all the remote servers. Transmission does not account for network topology. For example, if 10 remote servers reside in a datacenter that is very remote in relation to the requestor server, the data will be sent 10 times over the channel to the remote datacenter. Try to avoid large data sets when using GLOBAL IN.
  3. When transmitting data to remote servers, restrictions on network bandwidth are not configurable. You might overload the network.
  4. Try to distribute data across servers so that you don't need to use GLOBAL IN on a regular basis.
  5. If you need to use GLOBAL IN often, plan the location of the ClickHouse cluster so that a single group of replicas resides in no more than one data center with a fast network between them, so that a query can be processed entirely within a single data center.

It also makes sense to specify a local table in the GLOBAL IN clause, in case this local table is only available on the requestor server and you want to use data from it on remote servers.

Extreme Values

In addition to results, you can also get minimum and maximum values for the results columns. To do this, set the extremes setting to 1. Minimums and maximums are calculated for numeric types, dates, and dates with times. For other columns, the default values are output.

An extra two rows are calculated – the minimums and maximums, respectively. These extra two rows are output in JSON*, TabSeparated*, and Pretty* formats, separate from the other rows. They are not output for other formats.

In JSON* formats, the extreme values are output in a separate 'extremes' field. In TabSeparated* formats, the row comes after the main result, and after 'totals' if present. It is preceded by an empty row (after the other data). In Pretty* formats, the row is output as a separate table after the main result, and after totals if present.

Extreme values are calculated for rows before LIMIT, but after LIMIT BY. However, when using LIMIT offset, size, the rows before offset are included in extremes. In stream requests, the result may also include a small number of rows that passed through LIMIT.


The GROUP BY and ORDER BY clauses do not support positional arguments. This contradicts MySQL, but conforms to standard SQL. For example, GROUP BY 1, 2 will be interpreted as grouping by constants (i.e. aggregation of all rows into one).

You can use synonyms (AS aliases) in any part of a query.

You can put an asterisk in any part of a query instead of an expression. When the query is analyzed, the asterisk is expanded to a list of all table columns (excluding the MATERIALIZED and ALIAS columns). There are only a few cases when using an asterisk is justified:

  • When creating a table dump.
  • For tables containing just a few columns, such as system tables.
  • For getting information about what columns are in a table. In this case, set LIMIT 1. But it is better to use the DESC TABLE query.
  • When there is strong filtration on a small number of columns using PREWHERE.
  • In subqueries (since columns that aren't needed for the external query are excluded from subqueries).

In all other cases, we don't recommend using the asterisk, since it only gives you the drawbacks of a columnar DBMS instead of the advantages. In other words using the asterisk is not recommended.