Aggregation by time window is supported in TDengine. For example, in the case where temperature sensors report the temperature every seconds, the average temperature for every 10 minutes can be retrieved by performing a query with a time window.
Window related clauses are used to divide the data set to be queried into subsets and then aggregation is performed across the subsets. There are three kinds of windows: time window, status window, and session window. There are two kinds of time windows: sliding window and flip time/tumbling window.
TDengine 提供的特色查询包括标签切分查询和窗口切分查询。
## Time Window
## 标签切分查询
The `INTERVAL` clause is used to generate time windows of the same time interval. The `SLIDING` parameter is used to specify the time step for which the time window moves forward. The query is performed on one time window each time, and the time window moves forward with time. When defining a continuous query, both the size of the time window and the step of forward sliding time need to be specified. As shown in the figure blow, [t0s, t0e] ，[t1s , t1e]， [t2s, t2e] are respectively the time ranges of three time windows on which continuous queries are executed. The time step for which time window moves forward is marked by `sliding time`. Query, filter and aggregate operations are executed on each time window respectively. When the time step specified by `SLIDING` is same as the time interval specified by `INTERVAL`, the sliding time window is actually a flip time/tumbling window.
```sql
PARTITIONBYpart_list
```
part_list 可以是任意的标量表达式，包括列、常量、标量函数和它们的组合。
当 PARTITION BY 和标签一起使用时，TDengine 按如下方式处理标签切分子句：
`INTERVAL` and `SLIDING` should be used with aggregate functions and select functions. The SQL statement below is illegal because no aggregate or selection function is used with `INTERVAL`.
- 标签切分子句位于 WHERE 子句之后，且不能和 JOIN 子句一起使用。
- 标签切分子句将超级表数据按指定的标签组合进行切分，每个切分的分片进行指定的计算。计算由之后的子句定义（窗口子句、GROUP BY 子句或 SELECT 子句）。
- 标签切分子句可以和窗口切分子句（或 GROUP BY 子句）一起使用，此时后面的子句作用在每个切分的分片上。例如，将数据按标签 location 进行分组，并对每个组按 10 分钟进行降采样，取其最大值。
The time step specified by `SLIDING` cannot exceed the time interval specified by `INTERVAL`. The SQL statement below is illegal because the time length specified by `SLIDING` exceeds that specified by `INTERVAL`.
FILL 语句指定某一窗口区间数据缺失的情况下的填充模式。填充模式包括以下几种：
1. 不进行填充：NONE（默认填充模式）。
2. VALUE 填充：固定值填充，此时需要指定填充的数值。例如：FILL(VALUE, 1.23)。这里需要注意，最终填充的值受由相应列的类型决定，如 FILL(VALUE, 1.23)，相应列为 INT 类型，则填充值为 1。
3. PREV 填充：使用前一个非 NULL 值填充数据。例如：FILL(PREV)。
4. NULL 填充：使用 NULL 填充数据。例如：FILL(NULL)。
5. LINEAR 填充：根据前后距离最近的非 NULL 值做线性插值填充。例如：FILL(LINEAR)。
6. NEXT 填充：使用下一个非 NULL 值填充数据。例如：FILL(NEXT)。
:::info
1. 使用 FILL 语句的时候可能生成大量的填充输出，务必指定查询的时间区间。针对每次查询，系统可返回不超过 1 千万条具有插值的结果。
2. 在时间维度聚合中，返回的结果中时间序列严格单调递增。
3. 如果查询对象是超级表，则聚合函数会作用于该超级表下满足值过滤条件的所有表的数据。如果查询中没有使用 GROUP BY 语句，则返回的结果按照时间序列严格单调递增；如果查询中使用了 GROUP BY 语句分组，则返回结果中每个 GROUP 内不按照时间序列严格单调递增。
:::
```
SELECT COUNT(*) FROM temp_tb_1 INTERVAL(1m) SLIDING(2m);
```
### 时间窗口
When the time length specified by `SLIDING` is the same as that specified by `INTERVAL`, the sliding window is actually a flip/tumbling window. The minimum time range specified by `INTERVAL` is 10 milliseconds (10a) prior to version 2.1.5.0. Since version 2.1.5.0, the minimum time range by `INTERVAL` can be 1 microsecond (1u). However, if the DB precision is millisecond, the minimum time range is 1 millisecond (1a). Please note that the `timezone` parameter should be configured to be the same value in the `taos.cfg` configuration file on client side and server side.
In case of using integer, bool, or string to represent the status of a device at any given moment, continuous rows with the same status belong to a status window. Once the status changes, the status window closes. As shown in the following figure, there are two status windows according to status, [2019-04-28 14:22:07，2019-04-28 14:22:10] and [2019-04-28 14:22:11，2019-04-28 14:22:12]. Status window is not applicable to STable for now.
INTERVAL 和 SLIDING 子句需要配合聚合和选择函数来使用。以下 SQL 语句非法：
`STATE_WINDOW` is used to specify the column on which the status window will be based. For example：
```
SELECT * FROM temp_tb_1 INTERVAL(1m);
SELECT COUNT(*), FIRST(ts), status FROM temp_tb_1 STATE_WINDOW(status);
```
SLIDING 的向前滑动的时间不能超过一个窗口的时间范围。以下语句非法：
## Session Window
```
SELECT COUNT(*) FROM temp_tb_1 INTERVAL(1m) SLIDING(2m);
The primary key, i.e. timestamp, is used to determine which session window a row belongs to. If the time interval between two adjacent rows is within the time range specified by `tol_val`, they belong to the same session window; otherwise they belong to two different session windows. As shown in the figure below, if the limit of time interval for the session window is specified as 12 seconds, then the 6 rows in the figure constitutes 2 time windows, [2019-04-28 14:22:10，2019-04-28 14:22:30] and [2019-04-28 14:23:10，2019-04-28 14:23:30], because the time difference between 2019-04-28 14:22:30 and 2019-04-28 14:23:10 is 40 seconds, which exceeds the time interval limit of 12 seconds.
If the time interval between two continuous rows are within the time interval specified by `tol_value` they belong to the same session window; otherwise a new session window is started automatically. Session window is not supported on STable for now.
- Aggregate functions and select functions can be used in `function_list`, with each function having only one output. For example COUNT, AVG, SUM, STDDEV, LEASTSQUARES, PERCENTILE, MIN, MAX, FIRST, LAST. Functions having multiple outputs, such as DIFF or arithmetic operations can't be used.
-`LAST_ROW` can't be used together with window aggregate.
- Scalar functions, like CEIL/FLOOR, can't be used with window aggregate.
-`WHERE` clause can be used to specify the starting and ending time and other filter conditions
-`FILL` clause is used to specify how to fill when there is data missing in any window, including:
1. NONE: No fill (the default fill mode)
2. VALUE：Fill with a fixed value, which should be specified together, for example `FILL(VALUE, 1.23)`
3. PREV：Fill with the previous non-NULL value, `FILL(PREV)`
4. NULL：Fill with NULL, `FILL(NULL)`
5. LINEAR：Fill with the closest non-NULL value, `FILL(LINEAR)`
6. NEXT：Fill with the next non-NULL value, `FILL(NEXT)`
1. A huge volume of interpolation output may be returned using `FILL`, so it's recommended to specify the time range when using `FILL`. The maximum number of interpolation values that can be returned in a single query is 10,000,000.
2. The result set is in ascending order of timestamp when you aggregate by time window.
3. If aggregate by window is used on STable, the aggregate function is performed on all the rows matching the filter conditions. If `GROUP BY` is not used in the query, the result set will be returned in ascending order of timestamp; otherwise the result set is not exactly in the order of ascending timestamp in each group.
```
:::
SELECT COUNT(*), FIRST(ts) FROM temp_tb_1 SESSION(ts, tol_val);
```
Aggregate by time window is also used in continuous query, please refer to [Continuous Query](/develop/continuous-query).
### 示例
## Examples
智能电表的建表语句如下：
A table of intelligent meters can be created by the SQL statement below：
The average current, maximum current and median of current in every 10 minutes for the past 24 hours can be calculated using the SQL statement below, with missing values filled with the previous non-NULL values.
```
SELECT AVG(current), MAX(current), APERCENTILE(current, 50) FROM meters
