[
https://issues.apache.org/jira/browse/CSV-196?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15866217#comment-15866217
]
Matt Sun commented on CSV-196:
------------------------------
[~britter] I want to clarify that I never asked to have CSVParser store the
byte information. I was saying that if CSV Token could store some information
about the raw data, downstream users can use those info to do some computation,
FOR EXAMPLE counting bytes. Regarding performance, I don't think storing raw
data will incur *significant* cost. Timing wise, Lexer is already reading input
file char by char, storing the characters read will not increase time
complexity. It's still same order of time complexity. You may argue that
appending to StringBuffer is a cost, I agree. However, I wouldn't say it's
*significant*. Memory wise, given the fact a CSV token is fairly small, I also
don't think it will increase the burden of memory.
But your suggestion of "opt-in" sounds fine and reasonable. Another suggestion
I have is to only store the number of characters read by the Lexer in Token.
That saves a little time and memory space.
[~b.eckenfels] Do you mean offset from the beginning of the file? In splitting
case, it will be more useful to store the offset of the *END* of each record.
While hadoop is processing the split, it wants to make sure it doesn't go
across split boundary. After reading a CSV record, the program could figure out
the current position by retrieve the information of offset of the *END* of each
record. If only beginning offset is given and hadoop knows the beginning is
within the boundary, the end of the record may still go beyond the boundary.
I'm not sure how *easy* it is doing this day? Could you briefly point out how
to achieve this? What is the performance and memory impact?
> Store the information of raw data read by lexer
> -----------------------------------------------
>
> Key: CSV-196
> URL: https://issues.apache.org/jira/browse/CSV-196
> Project: Commons CSV
> Issue Type: Improvement
> Components: Parser
> Affects Versions: 1.4
> Reporter: Matt Sun
> Labels: patch
> Fix For: Patch Needed
>
> Original Estimate: 48h
> Remaining Estimate: 48h
>
> It will be good to have CSVParser class to store the info of whether a field
> was enclosed by quotes in the original source file.
> For example, for this data sample:
> A, B, C
> a1, "b1", c1
> CSVParser gives us record a1, b1, c1, which is helpful because it parsed
> double quotes, but we also lost the information of original data at the same
> time. We can't tell from the CSVRecord returned whether the original data is
> enclosed by double quotes or not.
> In our use case, we are integrating Apache Hadoop APIs with Commons CSV. CSV
> is one kind of input of Hadoop Jobs, which should support splitting input
> data. To accurately split a CSV file into pieces, we need to count the bytes
> of data CSVParser actually read. CSVParser doesn't have accurate information
> of whether a field was enclosed by quotes, neither does it store raw data of
> the original source. Downstream users of commons CSVParser is not able to get
> those info.
> To suggest a fix: Extend the token/CSVRecord to have a boolean field
> indicating whether the column was enclosed by quotes. While Lexer is doing
> getNextToken, set the flag if a field is encapsulated and successfully parsed.
> I find another issue reported with similar request, but it was marked as
> resolved: [CSV91]
> https://issues.apache.org/jira/browse/CSV-91?jql=project%20%3D%20CSV%20AND%20text%20~%20%22with%20quotes%22
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
