Fluentd's output plugin somtimes hit un-recoverable error during chunk flush.

• chunk contains wrong record for output configuration
• output plugin has a bug for specific record
• broken chunk is generated by hardware problem
• wrong setup for destination

Currently, we use retry limit and secondary for handling these chunks but it has several problems.

• bad chunk occupy flush threads until reach retry limit
• non retry limit environment can't rescue bad chunk

So we should care bad chunk for stability and performance.
The idea is if output plugin raises un-recoverable error during chunk flush, such chunks are routed to backup directory.

• The recoverable errors are TypeError, NoMethodError, ArgumentError, etc
• Use UnrecoverableError for plugin specific error

In addition, <system> directive provide backup_dir parameter. The default is /tmp/fluentd.