David Dreyfus created FLINK-10134:
-------------------------------------
Summary: UTF-16 support for TextInputFormat
Key: FLINK-10134
URL: https://issues.apache.org/jira/browse/FLINK-10134
Project: Flink
Issue Type: Bug
Components: Core
Affects Versions: 1.4.2
Reporter: David Dreyfus
It does not appear that Flink supports a charset encoding of "UTF-16". It
particular, it doesn't appear that Flink consumes the Byte Order Mark (BOM) to
establish whether a UTF-16 file is UTF-16LE or UTF-16BE.
TextInputFormat.setCharset("UTF-16") calls DelimitedInputFormat.setCharset(),
which sets TextInputFormat.charsetName and then modifies the previously set
delimiterString to construct the proper byte string encoding of the the
delimiter. This same charsetName is also used in TextInputFormat.readRecord()
to interpret the bytes read from the file.
There are two problems that this implementation would seem to have when using
UTF-16.
# delimiterString.getBytes(getCharset()) in DelimitedInputFormat.java will
return a Big Endian byte sequence including the Byte Order Mark (BOM). The
actual text file will not contain a BOM at each line ending, so the delimiter
will never be read. Moreover, if the actual byte encoding of the file is Little
Endian, the bytes will be interpreted incorrectly.
# TextInputFormat.readRecord() will not see a BOM each time it decodes a byte
sequence with the String(bytes, offset, numBytes, charset) call. Therefore, it
will assume Big Endian, which may not always be correct. [1]
[https://github.com/apache/flink/blob/master/flink-java/src/main/java/org/apache/flink/api/java/io/TextInputFormat.java#L95]
While there are likely many solutions, I would think that all of them would
have to start by reading the BOM from the file when a Split is opened and then
using that BOM to modify the specified encoding to a BOM specific one when the
caller doesn't specify one, and to overwrite the caller's specification if the
BOM is in conflict with the caller's specification. That is, if the BOM
indicates Little Endian and the caller indicates UTF-16BE, Flink should rewrite
the charsetName as UTF-16LE.
I hope this makes sense and that I haven't been testing incorrectly or
misreading the code.
I've verified the problem on version 1.4.2. I believe the problem exists on all
versions.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
