[KYLIN-4810] TrieDictionary is not correctly build

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: v2.3.2
• Fix Version/s: v3.1.2
• Component/s: Job Engine
• Labels: Dictionary
• Assignee: Shengjun Zheng
• Reporter: Shengjun Zheng
• Created: 07/Nov/20 16:12
• Updated: 27/Apr/21 02:34
• Resolved: 02/Mar/21 05:59

Description

Hi, recently, I've met a problem in our product environment: Segments failed to merge because TrieDictionaryForest was disordered

java.lang.IllegalStateException: Invalid input data. Unordered data cannot be split into multi trees
    at org.apache.kylin.dict.TrieDictionaryForestBuilder.addValue(TrieDictionaryForestBuilder.java:92)
    at org.apache.kylin.dict.TrieDictionaryForestBuilder.addValue(TrieDictionaryForestBuilder.java:78)
    at org.apache.kylin.dict.DictionaryGenerator$StringTrieDictForestBuilder.addValue(DictionaryGenerator.java:214)
    at org.apache.kylin.dict.DictionaryGenerator.buildDictionary(DictionaryGenerator.java:81)
    at org.apache.kylin.dict.DictionaryGenerator.buildDictionary(DictionaryGenerator.java:65)
    at org.apache.kylin.dict.DictionaryGenerator.mergeDictionaries(DictionaryGenerator.java:106)

After some analysis, we found out when there is large values in a dict-encoded column, iterating over a single TrieDictionaryTree will get unordered data.

 
Digging into the source code,  the root cause is as described: 

1. Kylin will split a TrieTree Node into two parts when a single nodes's value length is more than 255 bytes
2. Then, these tow parts of value will be added to build the TrieTree. In fact the splitted two parts should not be used as new values to add to the TrieTree.
3. Step 2 will cause the TrieDictionaryTree build more leave nodes，and the extra leaf nodes will be 'end-value' of dictionary tree;
4. It has no impact to the correctness of the dict tree itself, except for adding some additional nodes .
5. But If you spit a UTF-8 word, you will get unordered data when iterating over the tree ( Something todo with Java UTF-8  String Serialize/Deserialize implementations. Please Refer to JDK sun.nio.cs.UTF_8.class)

How to re-produce ? Run test code :

TrieDictionaryForestBuilder builder = new TrieDictionaryForestBuilder(new StringBytesConverter());
String longUrl = "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx你好~~~";
builder.addValue(longUrl);
TrieDictionaryForest<String> dict = builder.build();

TrieDictionaryForestBuilder mergeBuild = new TrieDictionaryForestBuilder(new StringBytesConverter());
for (int i = dict.getMinId(); i <= dict.getMaxId(); i++) {
    String str = dict.getValueFromId(i);
    System.out.println("add value into merge tree");
    mergeBuild.addValue(str);
}

The log output of this test code is:
add value into merge tree
add value into merge tree
16:59:36 [main] INFO org.apache.kylin.dict.TrieDictionaryForestBuilder.addValue(TrieDictionaryForestBuilder.java:127) values not in ascending order, previous 'xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\xEF\xBF\xBD', current 'xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\xE4\xBD\xA0\xE5\xA5\xBD~~~'

We can see from the test code's output：

1. We only add 1 value but the tire dictionary tree turn out to have 2 end vlaues
2. Iterating over the TrieDictionary Tree got unordered data

We address this problem by

1. classify values which is a whole column value, which is splitted value,
2. not mark splitted value as end-value of a TrieTree Node.

I wonder if there is something wrong, thanks.