[SPARK-48441][SQL] Fix StringTrim behaviour for non-UTF8_BINARY collations #46762
Conversation
common/unsafe/src/main/java/org/apache/spark/sql/catalyst/util/CollationSupport.java
Outdated
Show resolved
Hide resolved
uros-db
changed the title
uros-db
changed the title
### What changes were proposed in this pull request? CollationFactory has been updated to no longer mark UNICODE as collation that supportsBinaryCollation. To reflect these changes, various tests have been updated. However, some tests have been (temporarily) removed because StringTrim no longer supports UNICODE collation given the new UNICODE definition in CollationFactory. At this time, StringTrim expression only supports UTF8_BINARY & UTF8_BINARY_LCASE, but not ICU collations. This work is in progress (#46762), so we'll ensure appropriate test coverage with those changes. ### Why are the changes needed? UNICODE collation should not support binary collation. Note: in the future, we may want to consider a collation such as UNICODE_BINARY, which will support binary equality, but also maintain UNICODE ordering. ### Does this PR introduce _any_ user-facing change? Yes, UNICODE is no longer treated as a binary collation. This affects how equality works for UNICODE, and also which codepath is taken for various collation-aware string expression given UNICODE collated string arguments. ### How was this patch tested? Updated existing unit and e2e sql test for UNICODE collation. ### Was this patch authored or co-authored using generative AI tooling? No. Closes #46772 from uros-db/fix-unicode. Authored-by: Uros Bojanic <157381213+uros-db@users.noreply.github.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
uros-db
changed the title
### What changes were proposed in this pull request? Renaming `UTF8_BINARY_LCASE` collation to `UTF8_LCASE`. ### Why are the changes needed? As part of the collation effort in Spark, we've moved away from byte-by-byte logic towards character-by-character logic, so what we used to call `UTF8_BINARY_LCASE` is now more precisely `UTF8_LCASE`. For example, string searching in UTF8_LCASE now works on character-level (rather than on byte-level), which is reflected in this PRs: #46511, #46589, #46682, #46761, #46762. In addition, string comparison also works on character-level now, as per the changes introduced in this PR: #46700. ### Does this PR introduce _any_ user-facing change? Yes, what was previously named `UTF8_BINARY_LCASE` collation, will from now on be named `UTF8_LCASE`. ### How was this patch tested? Existing tests. ### Was this patch authored or co-authored using generative AI tooling? No. Closes #46924 from uros-db/rename-lcase. Authored-by: Uros Bojanic <157381213+uros-db@users.noreply.github.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
| return trimRight(trimLeft(srcString, trimString, collationId), trimString, collationId); | ||
| } | ||
|
|
||
| public static UTF8String lowercaseTrimLeft( |
There was a problem hiding this comment.
Let's add some document about what the method does:
- Create a hash set where we store the lowercase versions of the characters in the trim string.
- Iterate over the characters of the input UTF8 string from left to right and stop when we find a character of the input string that does not belong the trim set.
- Return the appropriate substring of the input string.
| ++searchIndex; | ||
| } | ||
|
|
||
| return srcString.substring(searchIndex, srcString.numChars()); |
There was a problem hiding this comment.
Should we optimize this to return srcString if searchIndex is 0?
| HashSet<Integer> trimChars = new HashSet<>(); | ||
| Iterator<Integer> trimIter = trimString.codePointIterator(); | ||
| while (trimIter.hasNext()) trimChars.add(UCharacter.toLowerCase(trimIter.next())); |
There was a problem hiding this comment.
This approach is a bit problematic because UCharacter.toLowerCase(codepoint) returns a code point which means that İ maps to i. The second problem that we have is that UCharacter.toLowerCase will maps the Greek final sigma to itself, whereas we want it to be mapped to the Greek small sigma.
Here are two simple test cases and the expected results here, that can help verify the correctness of the implementation:
- string:
Iİa, trim string:İ, expected result:a - string
ςσa, trim string:Σ, expected result:a.
These test cases can be natually extended to right trimming and bidirectional trimming of course.
| // Iterate over srcString from the left and find the first character that is not in trimChars. | ||
| String input = srcString.toString(); | ||
| int i = 0; | ||
| while (i < input.length()) { | ||
| String key = CollationFactory.getCollationKey(String.valueOf(input.charAt(i)), collationId); | ||
| if (!trimChars.contains(key)) break; | ||
| ++i; |
There was a problem hiding this comment.
The problem with this approach is that it focuses on matching single code points. Instead I think we need to think of trimming as a prefix-match kind of problem. Essentially if you have input string str for which we have already trimmed k characters, and some trim string trimStr, we want for each UTF8 character in trimStr to check if it is a prefix of substring(str, k+1). If this is the case then we need to trim str, otherwise stop trimming. This is essential if we want to keep Unicode semantics, and specifically character normalization.
Here are two interesting examples/test cases:
- string:
a\u030Aβγδ, trim string:å, expected result:βγδ. - string:
a\u030Aβγδ, trim string:aå, expected result:βγδ.
The second case brings up another interesting issue which having a match against multiple characters in the input string. I think we should follow a greedy approach here and use the match with the longest byte length in the input string.
| while (trimIter.hasNext()) trimChars.add(UCharacter.toLowerCase(trimIter.next())); | ||
|
|
||
| int searchIndex = srcString.numChars(); | ||
| Iterator<Integer> srcIter = srcString.reverseCodePointIterator(); | ||
| while (srcIter.hasNext()) { | ||
| if (!trimChars.contains(UCharacter.toLowerCase(srcIter.next()))) { | ||
| break; | ||
| } | ||
| --searchIndex; | ||
| } |
There was a problem hiding this comment.
Obviously the comments for left trim are applicable to right trim as well.
| } | ||
|
|
||
| public static UTF8String lowercaseTrimLeft( | ||
| public static UTF8String trimLeft( |
There was a problem hiding this comment.
Can we please add documentation for this method as well?
Does this method need to be public? Same for the lowercase version of the algorithm.
| /** | ||
| * Returns the code point starting from the byte at position `index`. | ||
| */ | ||
| public int codePointFrom(int index) { |
There was a problem hiding this comment.
Do we need this (given other comments for the PR)?
| }; | ||
| } | ||
|
|
||
| public int getChar(int index) { |
There was a problem hiding this comment.
Same question here.
Also, whatever remains, we need to add documentation for the methods we add.
| public Iterator<Integer> codePointIterator() { | ||
| return new CodePointIterator(); | ||
| } | ||
| private class CodePointIterator implements Iterator<Integer> { |
There was a problem hiding this comment.
We need an additional comment here that the iterator works correctly only for valid UTF8 strings (because of how it jumps from one character to the next).
There was a problem hiding this comment.
Let's add some interesting test cases along the lines of those posted in the comments.
What changes were proposed in this pull request?
String searching in UTF8_BINARY_LCASE now works on character-level, rather than on byte-level. For example:
ltrim("İ", "i")now returns"İ", because there exist no characters in"İ", starting from the left, such that lowercased version of those characters are equal to"i". Note, however, that there is a byte subsequence of"İ"such that lowercased version of that UTF-8 byte sequence equals to"i"(so the new behaviour is different than the old behaviour).Why are the changes needed?
Fix functions that give unusable results due to one-to-many case mapping when performing string search under UTF8_BINARY_LCASE (see example above).
Does this PR introduce any user-facing change?
Yes, behaviour of
trim*expressions is changed for edge cases with one-to-many case mapping.How was this patch tested?
New unit tests in
CollationSupportSuiteand new e2e sql tests inCollationStringExpressionsSuite.Was this patch authored or co-authored using generative AI tooling?
No.