[
https://issues.apache.org/jira/browse/KAFKA-3323?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15178187#comment-15178187
]
Michael Schiff edited comment on KAFKA-3323 at 3/3/16 5:56 PM:
---------------------------------------------------------------
Regarding solutions:
One alternative is to group segments to clean, not according to segment size in
bytes, but by limiting the difference between the segment base offset and the
max offset that goes into it.
I see two issues with this. The first is that we could wind up creating
segments that are actually pretty small, size wise. The second issue is that,
for people like me, who already have production topics with log segments with
(offsets - baseOffset) > Int.MAX_VALUE, changing the grouping does not fix our
issue. If you already have such log segments, they cannot be properly indexed
as is.
The second solution is to make indexes use long valued keys. This would mean
changing the index format, forcing you to rebuild all of your existing indices.
However, you can then continue grouping segments to clean the same way,
ensuring good sized segments. It also means that people with segments that are
already affected are not forced to drop their data.
was (Author: michael.schiff):
Regarding solutions:
One alternative is to group segments to clean, not according to segment size in
bytes, but by limiting the difference between the segment base offset and the
max offset that goes into it.
I see two issues with this. The first is that we could wind up creating
segments that are actually pretty small, size wise. The second issue is that,
for people like me, who already have production topics with log segments with
(offsets - baseOffset) > Int.MAX_VALUE, changing the grouping does not fix our
issue. If you already have such log segments, they cannot be properly indexed
as is.
The second solution is to make indexes use long valued keys. This would mean
changing the index format, forcing you to rebuild all of your existing indices.
However, you can then continue grouping segments to clean the same way,
ensuring good sized segments. It also means that people with segments that are
already incorrect are not forced to drop their data.
> Negative offsets in Log Segment Index files due to Integer overflow when
> compaction is enabled
> -----------------------------------------------------------------------------------------------
>
> Key: KAFKA-3323
> URL: https://issues.apache.org/jira/browse/KAFKA-3323
> Project: Kafka
> Issue Type: Bug
> Components: log
> Affects Versions: 0.8.1.1
> Reporter: Michael Schiff
> Assignee: Jay Kreps
> Attachments: index_dump.txt, log_dump.txt
>
>
> {code}
> /**
> * Append an entry for the given offset/location pair to the index. This
> entry must have a larger offset than all subsequent entries.
> */
> def append(offset: Long, position: Int) {
> inLock(lock) {
> require(!isFull, "Attempt to append to a full index (size = " + size +
> ").")
> if (size.get == 0 || offset > lastOffset) {
> debug("Adding index entry %d => %d to %s.".format(offset, position,
> file.getName))
> this.mmap.putInt((offset - baseOffset).toInt)
> this.mmap.putInt(position)
> this.size.incrementAndGet()
> this.lastOffset = offset
> require(entries * 8 == mmap.position, entries + " entries but file
> position in index is " + mmap.position + ".")
> } else {
> throw new InvalidOffsetException("Attempt to append an offset (%d) to
> position %d no larger than the last offset appended (%d) to %s."
> .format(offset, entries, lastOffset, file.getAbsolutePath))
> }
> }
> }
> {code}
> OffsetIndex.append assumes that (offset - baseOffset) can be represented as
> an integer without overflow. If the LogSegment is from a compacted topic,
> this assumption may not be valid. The result is a quiet integer overflow,
> which stores a negative value into the index. This breaks the binary search
> used to lookup offset positions -> large intervals of offsets are skipped by
> consumers who are bootstrapping themselves on the topic.
> I believe that the issue is caused by the LogCleaner. Specifically, by the
> groupings produced by
> {code}
> /**
> * Group the segments in a log into groups totaling less than a given size.
> the size is enforced separately for the log data and the index data.
> * We collect a group of such segments together into a single
> * destination segment. This prevents segment sizes from shrinking too much.
> *
> * @param segments The log segments to group
> * @param maxSize the maximum size in bytes for the total of all log data
> in a group
> * @param maxIndexSize the maximum size in bytes for the total of all index
> data in a group
> *
> * @return A list of grouped segments
> */
> private[log] def groupSegmentsBySize(segments: Iterable[LogSegment],
> maxSize: Int, maxIndexSize: Int): List[Seq[LogSegment]]
> {code}
> Since this method is only concerned with grouping by size, without taking
> baseOffset and groupMaxOffset into account, it will produce groups that when
> cleaned into a single segment, have offsets that overflow. This is more
> likely for topics with low key cardinality, but high update volume, as you
> could wind up with very few cleaned records, but with very high offsets.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
