Hi, Thanks for reply! Yes, we running Kafka on 32bit machine, but ‘log.index.size.max.bytes’ defaults to 10MB according http://kafka.apache.org/documentation.html#persistence and it is not redefined in our config.
$ find kafka-data-dir -name ‘*.index’ | wc -l 425 $ find kafka-data-dir -size +10M -name ‘*.index’ | wc -l 0 Pavel. 24 авг. 2014 г., в 19:19, Jay Kreps <jay.kr...@gmail.com> написал(а): > Is it possible that this is a 32 bit machine and you have more than 2gb of > index? > > -jay > > On Sunday, August 24, 2014, pavel.zalu...@sematext.com < > pavel.zalu...@sematext.com> wrote: > >> Hi, >> >> Sometimes we get: >> >> Caused by: java.lang.OutOfMemoryError: Map failed >> at sun.nio.ch.FileChannelImpl.map0(Native Method) >> at sun.nio.ch.FileChannelImpl.map(FileChannelImpl.java:885) >> ... 25 more >> >> (excuse me, we lost full stack trace, it starts from java.io.IOException >> which is caused by OOM) >> >> As far as I understand it happens because of failed ‘mmap’ call during >> index mapping: >> OffsetIndex.scala: >> >> val idx = raf.getChannel.map(FileChannel.MapMode.READ_WRITE, 0, >> len) >> >> Probably it happens because of memory limit. >> >> But after restart (and installing more memory, adjusting ulimits, etc.) we >> end with long startup time. This IOException is not handled, active log >> files are never flushed, some .index files maybe corrupted. It leads to >> startup times (starting from 40min for us). >> >> Is it possible to handle this exception, flush active log files, indexes >> and exit properly? In fact it can take ‘infinite’ time to recover all >> things in case of big number of topics/partitions. >> >> Thanks, >> Pavel.
