We found some problems during the test.
Data: 200Million docs, 1 shard, 0 replica
hits | avg | sum | value_count |
----------- | ------- | ------- | ----------- |
20,000 | .038s | .033s | .063s |
200,000 | .127s | .125s | .334s |
2,000,000 | .789s | .729s | 3.176s |
20,000,000 | 4.200s | 3.239s | 22.787s |
200,000,000 | 21.000s | 22.000s | 154.917s |
The performance of `avg`, `sum` and other is very close when performing
statistics, but the performance of `value_count` has always been poor,
even not on an order of magnitude. Based on some common-sense knowledge,
we think that `value_count` and sum are similar operations, and the time
consumed should be the same. Therefore, we have discussed the agg
of `value_count`.
The principle of counting in es is to traverse the field of each
document. If the field is an ordinary value, the count value is
increased by 1. If it is an array type, the count value is increased
by n. However, the problem lies in traversing each document and taking
out the field, which changes from disk to an object in the Java
language. We summarize its current problems with Elasticsearch as:
- Number cast to string overhead, and GC problems caused by a large
number of strings
- After the number type is converted to string, sorting and other
unnecessary operations are performed
Here is the proof of type conversion overhead.
```
// Java long to string source code, getChars is very time-consuming.
public static String toString(long i) {
int size = stringSize(i);
if (COMPACT_STRINGS) {
byte[] buf = new byte[size];
getChars(i, size, buf);
return new String(buf, LATIN1);
} else {
byte[] buf = new byte[size * 2];
StringUTF16.getChars(i, size, buf);
return new String(buf, UTF16);
}
}
```
test type | average | min | max | sum
------------ | ------- | ---- | ----------- | -------
double->long | 32.2ns | 28ns | 0.024ms | 3.22s
long->double | 31.9ns | 28ns | 0.036ms | 3.19s
long->String | 163.8ns | 93ns | 1921 ms | 16.3s
particularly serious.
Our optimization code is actually very simple. It is to manage different
types separately, instead of uniformly converting to string unified
processing. We added type identification in ValueCountAggregator, and
made special treatment for number and geopoint types to cancel their
type conversion. Because the string type is reduced and the string
constant is reduced, the improvement effect is very obvious.
hits | avg | sum | value_count | value_count | value_count | value_count | value_count | value_count |
| | | double | double | keyword | keyword | geo_point | geo_point |
| | | before | after | before | after | before | after |
----------- | ------- | ------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- |
20,000 | 38s | .033s | .063s | .026s | .030s | .030s | .038s | .015s |
200,000 | 127s | .125s | .334s | .078s | .116s | .099s | .278s | .031s |
2,000,000 | 789s | .729s | 3.176s | .439s | .348s | .386s | 3.365s | .178s |
20,000,000 | 4.200s | 3.239s | 22.787s | 2.700s | 2.500s | 2.600s | 25.192s | 1.278s |
200,000,000 | 21.000s | 22.000s | 154.917s | 18.990s | 19.000s | 20.000s | 168.971s | 9.093s |
- The results are more in line with common sense. `value_count` is about
the same as `avg`, `sum`, etc., or even lower than these. Previously,
`value_count` was much larger than avg and sum, and it was not even an
order of magnitude when the amount of data was large.
- When calculating numeric types such as `double` and `long`, the
performance is improved by about 8 to 9 times; when calculating the
`geo_point` type, the performance is improved by 18 to 20 times.
b99a50bcb9