FIX
1. remove float and double literal toString and getStringValue introduced by
PR #23504 and PR #23271
These functions lead to wrong cast result of double and float literal
2. fix compute signature for datetimev2 always produce scale 6
3. fix stats calculator failed when generate node stats with two same column
4. constant fold on fe failed when cast double to integral
TODO
after fix the first problem, some mv matching not work well, fix them later
- test_dup_mv_div
- test_dup_mv_json
- test_tcu
materialized view def is as following:
> select l_linenumber, o_custkey
> from orders
> left join lineitem on lineitem.L_ORDERKEY = orders.O_ORDERKEY
> where o_custkey = 1;
when query is as following, it can be rewritten by mv above
it requires that query has reject null filters on the join right input,
current supported filter are "=", "<", "<=", ">", ">=", "<=>"
> select IFNULL(orders.O_CUSTKEY, 0) as custkey_not_null,
> case when l_linenumber in (1,2,3) then l_linenumber else o_custkey end as case_when
> from orders
> inner join lineitem on orders.O_ORDERKEY = lineitem.L_ORDERKEY
> where o_custkey = 1 and l_linenumber > 0;
Only write editlog for manual analyze task, don't need to do so for auto tasks to reduce writing editlog.
Add error message to job info while task failed.
Support to create partition materialized view using nodata table
Such as the table def as following:
> CREATE TABLE `test_no_data` (
> `user_id` LARGEINT NOT NULL COMMENT '"用户id"',
> `date` DATE NOT NULL COMMENT '"数据灌入日期时间"',
> `num` SMALLINT NOT NULL COMMENT '"数量"'
> ) ENGINE=OLAP
> DUPLICATE KEY(`user_id`, `date`, `num`)
> COMMENT 'OLAP'
> PARTITION BY RANGE(`date`)
> (PARTITION p201701_1000 VALUES [('0000-01-01'), ('2017-02-01')),
> PARTITION p201702_2000 VALUES [('2017-02-01'), ('2017-03-01')),
> PARTITION p201703_all VALUES [('2017-03-01'), ('2017-04-01')))
> DISTRIBUTED BY HASH(`user_id`) BUCKETS 2
> PROPERTIES ('replication_num' = '1') ;
when table test_no_data has no data, it also support to create partition materialized view as following:
> CREATE MATERIALIZED VIEW no_data_partition_mv
> BUILD IMMEDIATE REFRESH AUTO ON MANUAL
> partition by(`date`)
> DISTRIBUTED BY RANDOM BUCKETS 2
> PROPERTIES ('replication_num' = '1')
> AS
> SELECT * FROM test_no_data where date > '2017-05-01';
>
Query rewrite by mv support bitmap_union and bitmap_union_count roll up, aggregate functions which supports roll up is listed as following:
| 查询中函数 | 物化视图中函数 | 函数上卷后 |
|------------------|--------------|--------------------|
| max | max | max |
| min | min | min |
| sum | sum | sum |
| count | count | sum |
| count(distinct ) | bitmap_union | bitmap_union_count |
| bitmap_union | bitmap_union | bitmap_union|
| bitmap_union_count | bitmap_union | bitmap_union_count |
this depends on https://github.com/apache/doris/pull/29256
test_unique_table.groovy and test_unique_table_like.groovy both use database test_unique_db.
If they run at the same time, we may got the following errors:
java.sql.SQLException: errCode = 2, detailMessage = There are still some transactions in the COMMITTED state waiting to be completed. The database [default_cluster:test_unique_db] cannot be dropped. If you want to forcibly drop(cannot be recovered), please use "DROP database FORCE".
Both Master & Branch2.0 have this problem.
An error occurred when starting BE with JDK17
```java
Exception: java.lang.StackOverflowError thrown from the UncaughtExceptionHandler in thread "process reaper"
```
This error occurs when BE's java code calls Runtime.exec() to fork the child process.
It turned out that Doris was calling the `glog` library in the C++ layer to cause this problem.
The solution comes from: https://github.com/google/glog/issues/975
"operator_id" should be invisible, but the local shuffle is a planned operator in the BE (Backend), without a plan node ID. We use it in profiles and other places, and there might be duplicates. Therefore, we switch it to a negative number here to distinguish it as a plan node ID.
The current logic for SQL dialect conversion is all in the `fe-core` module, which may lead to the following issues:
- Changes to the dialect conversion logic may occur frequently, requiring users to upgrade the Doris version frequently within the fe-core module, leading to a longer change cycle.
- The cost of customized development is high, requiring users to replace the fe-core JAR package.
Turning it into a plugin can address the above issues properly.
Problem:
fe ut failed cause of null pointer error
Cause:
fe ut getting statement context from connection context failed
Resolved:
add null pointer judgement
Force to use zonemap for collecting string type min max.
String type is not using zonemap for min max, because zonemap value at BE side is truncated at 512 bytes which may cause the value not accurate. But it's OK for statisitcs min max, and this could also avoid scan whole table while sampling.
