Wrong result when searching on json attributes

[Nick-S-2018]

Wrong result when searching on json attributes

[sanikolaev]

@Nick-S-2018 As discussed on today's dev call, please check if type casting works in this case and how applicable it is to be used in the clients.

[sanikolaev]

double(f.a) doesn't work:

mysql> drop table if exists t; create table t(f json); insert into t(id,f) values(1,'{"a":2.5}'); select * from t where double(f.a) <= 2;
--------------
drop table if exists t
--------------

Query OK, 0 rows affected (0.00 sec)

--------------
create table t(f json)
--------------

Query OK, 0 rows affected (0.00 sec)

--------------
insert into t(id,f) values(1,'{"a":2.5}')
--------------

Query OK, 1 row affected (0.00 sec)

--------------
select * from t where double(f.a) <= 2
--------------

+------+----------------+
| id   | f              |
+------+----------------+
|    1 | {"a":2.500000} |
+------+----------------+
1 row in set (0.00 sec)
--- 1 out of 1 results in 0ms ---

[tomatolog]

seems the right form that works now is the

select * from t where f.a <= 2.0;
Empty set (0.00 sec)
--- 0 out of 0 results in 0ms ---

[Nick-S-2018]

It looks like the correct float comparison can be guaranteed for int numbers with less than 7 digits; the limit is 2097152 (the first 22-bit number in binary format). For the numbers less than that conversion to float should work fine, like in the example above; for the larger numbers precision loss appears.

[sanikolaev]

It looks like the correct float comparison can be guaranteed for int numbers with less than 7 digits

Is this what you mean by the general theory of float comparison or you mean our specific case?

[sanikolaev]

mysql > select * from t where f.a <= 2;

Nick has an idea of converting N to N.0 in this case when N is in a specific range. @Nick-S-2018 pls provide more details.

[Nick-S-2018]

Unfortunately, this idea guarantees only the precision for values with a certain number of digits. E.g., '+-2097152' range would fix the issue only for values with one digit after decimal point, like in the original example. The more digits after the point a stored value has, the lesser such range would became. So, this approach does not look effective.

[Nick-S-2018]

Elastic always converts the compared value to the data type of the compared field. In case of a JSON object attribute, its type is defined by the detected type of the first inserted value, unless it has been set explicitly on the index creation . By default, Elastic detects numbers with floating points as 'floats', so the compared numeric value will be converted to 'float'. However, if 'double' has been set as the attribute data type, the compared value will be treated as 'double' too.

[sanikolaev]

As discussed, we'll implement support for where double(f.a) <= 2.