[
https://issues.apache.org/jira/browse/IO-279?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15241662#comment-15241662
]
Randall Theobald commented on IO-279:
-------------------------------------
I am also encountering this problem with 2.4. Using it on RHEL 5.11 (yes, very
old). It seems that there is a window that gets hit a few times per day where
the OS file system updates the file modification date slightly before the file
size is updated, hence Tailer thinks it needs to re-read from the very
beginning.
To workaround, I have removed all checks to file modification time from the run
method. In my case, detecting changes by file size alone is enough. It would be
nice for this to be configurable.
> Tailer erroneously considers file as new
> ----------------------------------------
>
> Key: IO-279
> URL: https://issues.apache.org/jira/browse/IO-279
> Project: Commons IO
> Issue Type: Bug
> Affects Versions: 2.0.1, 2.4
> Reporter: Sergio Bossa
> Attachments: IO-279.patch, disable_resetting.patch, fix-tailer.patch,
> modify-test-fixed.patch, modify-test.patch
>
>
> Tailer sometimes erroneously considers the tailed file as new, forcing a
> repositioning at the start of the file: I'm still unable to reproduce this in
> a test case, because it only happens to me with huge log files during Apache
> Tomcat startup.
> This is the piece of code causing the problem:
> {code}
> // See if the file needs to be read again
> if (length > position) {
> // The file has more content than it did last time
> last = System.currentTimeMillis();
> position = readLines(reader);
> } else if (FileUtils.isFileNewer(file, last)) {
> /* This can happen if the file is truncated or overwritten
> * with the exact same length of information. In cases like
> * this, the file position needs to be reset
> */
> position = 0;
> reader.seek(position); // cannot be null here
> // Now we can read new lines
> last = System.currentTimeMillis();
> position = readLines(reader);
> }
> {code}
> What probably happens is that the new file content is about to be written on
> disk, the date is already updated but content is still not flushed, so actual
> length is untouched and there you go.
> In other words, I think there should be some better method to verify the
> condition above, rather than relying only on dates: keeping and comparing the
> hash code of the latest line may be a solution, but may hurt performances ...
> other ideas?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
