An update on this. It appears that the phenomenon I'm seeing is that disk space is freed on restart, but it's not due files getting deleted on restart, but instead files are getting truncated on restart. It appears that log files get pre-allocated to a larger size than is used right away. Upon restart, they get truncated to the size of the file that actually contains data. Does this make sense?
Before restart, I see a large number of log files size 2.1Gb. Upon restart, the disk space reclaimed drops to almost half that, on average.
Jason On Thu, May 23, 2013 at 8:55 PM, Jun Rao <[EMAIL PROTECTED]> wrote:
As Jun says, we haven't seen that issue and no one else has reported that but it sounds like a bug of some kind.
In 0.7 we don't do any preallocation of anything. The only time files shrink is during recovery--we re-checksum all messages that may not have been flushed and if any invalid messages are found we truncate them off. This would only occur after an unclean shutdown (kill -9) and you would get a bunch of errors in log warning you this was happening.
On Sun, Jul 14, 2013 at 1:36 AM, Jason Rosenberg <[EMAIL PROTECTED]> wrote:
Apparently it's not hard-allocated space, and would be released under resource pressure....seems we may need to update how we monitor disk space usage, etc....
But, it seems that the default log file size of 1.1Gb, causes it to jump to preallocate an extra Gb. So, in theory, if I set a strategic log file size to be just under the threshold that forces it to exponentially double the size from 1Gb to 2Gb, I should be able to mostly solve this issue. E.g. use 950Mb instead of 1.1Gb max log file size.
If I change the max log file size for a broker, and restart it, will it respect the new size going forward?
Jason On Sun, Jul 14, 2013 at 9:44 AM, Jay Kreps <[EMAIL PROTECTED]> wrote:
It looks like xfs will reclaim the preallocated space for a file, after it is closed.
Does kafka close a file after it has reached it's max size and started writing to the next log file in sequence? Or does it keep all open until they are deleted, or the server quits (that's what it seems like).
I could imagine that it might need to keep log files open, in order to allow consumers access to them. But does it keep them open indefinitely, after there is no longer any data to be written to them, and no consumers are currently attempting to read from them?
Jason On Tue, Jul 16, 2013 at 4:32 PM, Jay Kreps <[EMAIL PROTECTED]> wrote:
Yes, Kafka keeps all log files open indefinitely. There is no inherent reason this needs to be the case, though, it would be possible to LRU out old file descriptors and close them if they are not accessed for a few hours and then reopen on the first access. We just haven't implemented anything like that.
It would be good to understand this a little better. Does xfs pre-allocate space for all open files? Perhaps just closing the file on log role and opening it read-only would solve the issue? Is this at all related to the use of sparse files for the indexes (i.e. RandomAccessFile.setLength(10MB) when we create the index)? Does this effect other filesystems or just xfs?
-Jay On Fri, Jul 26, 2013 at 12:42 AM, Jason Rosenberg <[EMAIL PROTECTED]> wrote:
My only experience so far with this is using XFS. It appears the XFS behavior is evolving, and in fact, we see somewhat different behavior from 2 of our CentOS kernel versions in use. I've been trying to ask questions about all this on the XFS.org mailing list, but so far, having not much luck understanding the xfs versioning correlated to CentOS versions.
Anyway, yes, I think it would definitely be worth trying the solution you suggest, which would be to close the file on rotation, and re-open read-only. Or to close files after a few hours of not being accessed. If a patch for one of these approaches can be cobbled together, I'd love to test it out on our staging environment. I'd be willing to experiment with such a patch myself, although I'm not 100% of all the places to look (but might dive in).
Xfs appears to the option of using dynamic, speculative preallocation, in which case it progressively doubles the amount of space reserved for a file, as the file grows. It does do this for all open files. If the file is closed, it will then release the preallocated space not in use. It's not clear whether this releasing of space happens immediately on close, and whether re-opening the file read-only immediately, will keep it from releasing space (still trying to gather more info on that).
I haven't looked too much at the index files, but those too appear to have this behavior (e.g. preallocated size is always on the order of double the actual size, until the app is restarted).
Jason On Fri, Jul 26, 2013 at 12:46 PM, Jay Kreps <[EMAIL PROTECTED]> wrote:
Cool, good to know. On Fri, Jul 26, 2013 at 2:00 PM, Jason Rosenberg <[EMAIL PROTECTED]> wrote:
NEW: Monitor These Apps!
Apache Lucene, Apache Solr and all other Apache Software Foundation project and their respective logos are trademarks of the Apache Software Foundation.
Elasticsearch, Kibana, Logstash, and Beats are trademarks of Elasticsearch BV, registered in the U.S. and in other countries. This site and Sematext Group is in no way affiliated with Elasticsearch BV.
Service operated by Sematext