This topic describes how to use the APPROX_COUNT_DISTINCT function to improve the performance of your deployments.
Limits
This function is supported only in Realtime Compute for Apache Flink that uses Ververica Runtime (VVR) 3.0.0 or later.
Background information
When you optimize the COUNT DISTINCT function, distinct key information must be saved in the state data of the aggregate node. If a large number of distinct keys exist, the read/write overhead of state data is high. This causes a bottleneck in the performance optimization of deployments. In many cases, accurate computation is not necessary. If you want to achieve high deployment performance at the expense of a small portion of accuracy, you can use the APPROX_COUNT_DISTINCT function. APPROX_COUNT_DISTINCT supports miniBatch and local-global optimization on the aggregate node. When you use this function, make sure that the following requirements are met:
The input data does not contain retract messages.
A large number of distinct keys, such as unique visitors (UVs), exist. The APPROX_COUNT_DISTINCT function cannot bring obvious benefits if only a small number of distinct keys exist.
Syntax
APPROX_COUNT_DISTINCT(col [, accuracy])Input parameters
Parameter | Data type | Description |
col | All data types | The name of the field. |
accuracy | FLOAT | The computation accuracy. This parameter is optional. A larger value indicates a higher accuracy, which leads to a higher state overhead. A higher state overhead weakens the performance of the APPROX_COUNT_DISTINCT function. Valid values: (0.0, 1.0). Default value: 0.99. |
Example
Test data
Table 1. T1 a (VARCHAR)
b (BIGINT)
Hi
1
Hi
2
Hi
3
Hi
4
Hi
5
Hi
6
Hello
1
Hello
2
Hello
3
Hello
4
Test statements
SELECT a, APPROX_COUNT_DISTINCT(b) as b, APPROX_COUNT_DISTINCT(b, 0.9) as c FROM T1 GROUP BY a;Test results
a (VARCHAR)
b (BIGINT)
c (BIGINT)
Hi
6
6
Hello
4
4