GROUP BY Modifiers
Starting from v7.4.0, the GROUP BY clause of TiDB supports the WITH ROLLUP modifier.
In the GROUP BY clause, you can specify one or more columns as a group list and append the WITH ROLLUP modifier after the list. Then, TiDB will conduct multidimensional descending grouping based on the columns in the group list and provide you with summary results for each group in the output.
Grouping method:
- The first grouping dimension includes all columns in the group list.
- Subsequent grouping dimensions start from the right end of the grouping list and exclude one more column at a time to form new groups.
Aggregation summaries: for each dimension, the query performs aggregation operations, and then aggregates the results of this dimension with the results of all previous dimensions. This means that you can get aggregated data at different dimensions, from detailed to overall.
With this grouping method, if there are N columns in the group list, TiDB aggregates the query results on N+1 groups.
For example:：
SELECT count(1) FROM t GROUP BY a,b,c WITH ROLLUP;
In this example, TiDB will aggregate the calculation results of count(1) on 4 groups (that is, {a, b, c}, {a, b}, {a}, and {}) and output the summary results for each group.
Use cases
Aggregating and summarizing data from multiple columns is commonly used in OLAP (Online Analytical Processing) scenarios. By using the WITH ROLLUP modifier, you can get additional rows that display super summary information from other high-level dimensions in your aggregated results. Then, you can use the super summary information for advanced data analysis and report generation.
Prerequisites
Currently, TiDB supports generating valid execution plans for the WITH ROLLUP syntax only in TiFlash MPP mode. Therefore, make sure that your TiDB cluster has been deployed with TiFlash nodes and that target fact tables are configured with TiFlash replicas properly.
Examples
Suppose you have a profit table named bank with the year, month, day, and profit columns.
CREATE TABLE bank
(
year INT,
month VARCHAR(32),
day INT,
profit DECIMAL(13, 7)
);
ALTER TABLE bank SET TIFLASH REPLICA 1; -- Add a TiFlash replica for the table
INSERT INTO bank VALUES(2000, "Jan", 1, 10.3),(2001, "Feb", 2, 22.4),(2000,"Mar", 3, 31.6)
To get the profit for the bank per year, you can use a simple GROUP BY clause as follows:
SELECT year, SUM(profit) AS profit FROM bank GROUP BY year;
+------+--------------------+
| year | profit |
+------+--------------------+
| 2001 | 22.399999618530273 |
| 2000 | 41.90000057220459 |
+------+--------------------+
2 rows in set (0.15 sec)
In addition to yearly profits, bank reports usually also need to include the overall profit for all years or monthly divided profits for detailed profit analysis. Before v7.4.0, you have to use different GROUP BY clauses in multiple queries and join the results using UNION to obtain aggregated summaries. Starting from v7.4.0, you can simply achieve the desired results in a single query by appending the WITH ROLLUP modifier to the GROUP BY clause.
SELECT year, month, SUM(profit) AS profit from bank GROUP BY year, month WITH ROLLUP ORDER BY year desc, month desc;
+------+-------+--------------------+
| year | month | profit |
+------+-------+--------------------+
| 2001 | Feb | 22.399999618530273 |
| 2001 | NULL | 22.399999618530273 |
| 2000 | Mar | 31.600000381469727 |
| 2000 | Jan | 10.300000190734863 |
| 2000 | NULL | 41.90000057220459 |
| NULL | NULL | 64.30000019073486 |
+------+-------+--------------------+
6 rows in set (0.025 sec)
The preceding results include aggregated data at different dimensions: by both year and month, by year, and overall. In the results, a row without NULL values indicates that the profit in that row is calculated by grouping both year and month. A row with a NULL value in the month column indicates that profit in that row is calculated by aggregating all months in a year, while a row with a NULL value in the year column indicates that profit in that row is calculated by aggregating all years.
Specifically:
- The
profitvalue in the first row comes from the 2-dimensional group{year, month}, representing the aggregation result for the fine-grained{2000, "Jan"}group. - The
profitvalue in the second row comes from the 1-dimensional group{year}, representing the aggregation result for the mid-level{2001}group. - The
profitvalue in the last row comes from the 0-dimensional grouping{}, representing the overall aggregation result.
NULL values in the WITH ROLLUP results are generated just before the Aggregate operator is applied. Therefore, you can use NULL values in SELECT, HAVING, and ORDER BY clauses to further filter the aggregated results.
For example, you can use NULL in the HAVING clause to filter and view the aggregated results of 2-dimensional groups only:
SELECT year, month, SUM(profit) AS profit FROM bank GROUP BY year, month WITH ROLLUP HAVING year IS NOT null AND month IS NOT null;
+------+-------+--------------------+
| year | month | profit |
+------+-------+--------------------+
| 2000 | Mar | 31.600000381469727 |
| 2000 | Jan | 10.300000190734863 |
| 2001 | Feb | 22.399999618530273 |
+------+-------+--------------------+
3 rows in set (0.02 sec)
Note that if a column in the GROUP BY list contains native NULL values, the aggregation results of WITH ROLLUP might mislead the query results. To address this issue, you can use the GROUPING() function to distinguish native NULL values from NULL values generated by WITH ROLLUP. This function takes a grouping expression as a parameter and returns 0 or 1 to indicate whether the grouping expression is aggregated in the current result. 1 represents aggregated, and 0 represents not aggregated.
The following example shows how to use the GROUPING() function:
SELECT year, month, SUM(profit) AS profit, grouping(year) as grp_year, grouping(month) as grp_month FROM bank GROUP BY year, month WITH ROLLUP ORDER BY year DESC, month DESC;
+------+-------+--------------------+----------+-----------+
| year | month | profit | grp_year | grp_month |
+------+-------+--------------------+----------+-----------+
| 2001 | Feb | 22.399999618530273 | 0 | 0 |
| 2001 | NULL | 22.399999618530273 | 0 | 1 |
| 2000 | Mar | 31.600000381469727 | 0 | 0 |
| 2000 | Jan | 10.300000190734863 | 0 | 0 |
| 2000 | NULL | 41.90000057220459 | 0 | 1 |
| NULL | NULL | 64.30000019073486 | 1 | 1 |
+------+-------+--------------------+----------+-----------+
6 rows in set (0.028 sec)
From this output, you can get an understanding of the aggregation dimension of a row directly from the results of grp_year and grp_month, which prevents interference from native NULL values in the year and month grouping expressions.
The GROUPING() function can accept up to 64 grouping expressions as parameters. In the output of multiple parameters, each parameter generates a result of 0 or 1, and these parameters collectively form a 64-bit UNSIGNED LONGLONG with each bit as 0 or 1. You can use the following formula to get the bit position of each parameter as follows:
GROUPING(day, month, year):
result for GROUPING(year)
+ result for GROUPING(month) << 1
+ result for GROUPING(day) << 2
By using multiple parameters in the GROUPING() function, you can efficiently filter aggregate results at any high dimension. For example, you can quickly filter the aggregate results for each year and all years by using GROUPING(year, month).
SELECT year, month, SUM(profit) AS profit, grouping(year) as grp_year, grouping(month) as grp_month FROM bank GROUP BY year, month WITH ROLLUP HAVING GROUPING(year, month) <> 0 ORDER BY year DESC, month DESC;
+------+-------+--------------------+----------+-----------+
| year | month | profit | grp_year | grp_month |
+------+-------+--------------------+----------+-----------+
| 2001 | NULL | 22.399999618530273 | 0 | 1 |
| 2000 | NULL | 41.90000057220459 | 0 | 1 |
| NULL | NULL | 64.30000019073486 | 1 | 1 |
+------+-------+--------------------+----------+-----------+
3 rows in set (0.023 sec)
How to interpret the ROLLUP execution plan
To meet the requirements of multidimensional grouping, multidimensional data aggregation uses the Expand operator to replicate data. Each replica corresponds to a group at a specific dimension. With the data shuffling capability of MPP, the Expand operator can rapidly reorganize and calculate a large volume of data between multiple TiFlash nodes, fully utilizing the computational power of each node.
The implementation of the Expand operator is similar to that of the Projection operator. The difference is that Expand is a multi-level Projection, which contains multiple levels of projection operation expressions. For each row of the raw data, the Projection operator generates only one row in results, whereas the Expand operator generates multiple rows in results (the number of rows is equal to the number of levels in projection operation expressions).
The following is an example of an execution plan:
explain SELECT year, month, grouping(year), grouping(month), SUM(profit) AS profit FROM bank GROUP BY year, month WITH ROLLUP;
+----------------------------------------+---------+--------------+---------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| id | estRows | task | access object | operator info |
+----------------------------------------+---------+--------------+---------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| TableReader_44 | 2.40 | root | | MppVersion: 2, data:ExchangeSender_43 |
| └─ExchangeSender_43 | 2.40 | mpp[tiflash] | | ExchangeType: PassThrough |
| └─Projection_8 | 2.40 | mpp[tiflash] | | Column#6->Column#12, Column#7->Column#13, grouping(gid)->Column#14, grouping(gid)->Column#15, Column#9->Column#16 |
| └─Projection_38 | 2.40 | mpp[tiflash] | | Column#9, Column#6, Column#7, gid |
| └─HashAgg_36 | 2.40 | mpp[tiflash] | | group by:Column#6, Column#7, gid, funcs:sum(test.bank.profit)->Column#9, funcs:firstrow(Column#6)->Column#6, funcs:firstrow(Column#7)->Column#7, funcs:firstrow(gid)->gid, stream_count: 8 |
| └─ExchangeReceiver_22 | 3.00 | mpp[tiflash] | | stream_count: 8 |
| └─ExchangeSender_21 | 3.00 | mpp[tiflash] | | ExchangeType: HashPartition, Compression: FAST, Hash Cols: [name: Column#6, collate: binary], [name: Column#7, collate: utf8mb4_bin], [name: gid, collate: binary], stream_count: 8 |
| └─Expand_20 | 3.00 | mpp[tiflash] | | level-projection:[test.bank.profit, <nil>->Column#6, <nil>->Column#7, 0->gid],[test.bank.profit, Column#6, <nil>->Column#7, 1->gid],[test.bank.profit, Column#6, Column#7, 3->gid]; schema: [test.bank.profit,Column#6,Column#7,gid] |
| └─Projection_16 | 3.00 | mpp[tiflash] | | test.bank.profit, test.bank.year->Column#6, test.bank.month->Column#7 |
| └─TableFullScan_17 | 3.00 | mpp[tiflash] | table:bank | keep order:false, stats:pseudo |
+----------------------------------------+---------+--------------+---------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
10 rows in set (0.05 sec)
In this example execution plan, you can view the multiple-level expression of the Expand operator in the operator info column of the Expand_20 row. It consists of 2-dimensional expressions, and you can view the schema information of the Expand operator at the end of the row, which is schema: [test.bank.profit, Column#6, Column#7, gid].
In the schema information of the Expand operator, GID is generated as an additional column. Its value is calculated by the Expand operator based on the grouping logic of different dimensions, and the value reflects the relationship between the current data replica and the grouping set. In most cases, the Expand operator uses a Bit-And operation, which can represent 63 combinations of grouping items for ROLLUP, corresponding to 64 dimensions of grouping. In this mode, TiDB generates the GID value depending on whether the grouping set of the required dimension contains grouping expressions when the current data replica is replicated, and it fills a 64-bit UINT64 value in the order of columns to be grouped.
In the preceding example, the order of columns in the grouping list is [year, month], and the dimension groups generated by the ROLLUP syntax are {year, month}, {year}, and {}. For the dimension group {year, month}, both year and month are required columns, so TiDB fills the bit positions for them with 1 and 1 correspondingly. This forms a UINT64 of 11...0, which is 3 in decimal. Therefore, the projection expression is [test.bank.profit, Column#6, Column#7, 3->gid] (where column#6 corresponds to year, and column#7 corresponds to month).
The following is an example row of the raw data:
+------+-------+------+------------+
| year | month | day | profit |
+------+-------+------+------------+
| 2000 | Jan | 1 | 10.3000000 |
+------+-------+------+------------+
After the Expand operator is applied, you can get the following three rows of results:
+------------+------+-------+-----+
| profit | year | month | gid |
+------------+------+-------+-----+
| 10.3000000 | 2000 | Jan | 3 |
+------------+------+-------+-----+
| 10.3000000 | 2000 | NULL | 1 |
+------------+------+-------+-----+
| 10.3000000 | NULL | NULL | 0 |
+------------+------+-------+-----+
Note that the SELECT clause in the query uses the GROUPING function. When the GROUPING function is used in the SELECT, HAVING, or ORDER BY clauses, TiDB rewrites it during the logical optimization phase, transforms the relationship between the GROUPING function and the GROUP BY items into a GID related to the logic of dimension group (also known as grouping set), and fills this GID as metadata into the new GROUPING function.