[ https://issues.apache.org/jira/browse/KAFKA-631?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13558971#comment-13558971 ]
Jay Kreps commented on KAFKA-631: --------------------------------- I did some testing on the I/O throttling and verified that this does indeed maintain the expected I/O rate. Two gotchas in this, first you can't look at iostat because the OS will batch up writes and then asynchronously flush them out at a rate greater than what we requested. Second since the limit is on read and write combined a limit of 5MB/sec will lead to the offset map building happening at exactly 5MB/sec but the cleaning will be closer to 2.5MB/sec because cleaning involves first reading in messages then writing them back out so 1MB of cleaning does 2MB of I/O (assuming 100% retention). > Implement log compaction > ------------------------ > > Key: KAFKA-631 > URL: https://issues.apache.org/jira/browse/KAFKA-631 > Project: Kafka > Issue Type: New Feature > Components: core > Affects Versions: 0.8.1 > Reporter: Jay Kreps > Assignee: Jay Kreps > Attachments: KAFKA-631-v1.patch, KAFKA-631-v2.patch > > > Currently Kafka has only one way to bound the space of the log, namely by > deleting old segments. The policy that controls which segments are deleted > can be configured based either on the number of bytes to retain or the age of > the messages. This makes sense for event or log data which has no notion of > primary key. However lots of data has a primary key and consists of updates > by primary key. For this data it would be nice to be able to ensure that the > log contained at least the last version of every key. > As an example, say that the Kafka topic contains a sequence of User Account > messages, each capturing the current state of a given user account. Rather > than simply discarding old segments, since the set of user accounts is > finite, it might make more sense to delete individual records that have been > made obsolete by a more recent update for the same key. This would ensure > that the topic contained at least the current state of each record. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira