Many programs take the log output of bits of Hadoop, and try and parse it. Some may also put their own back end behind commons-logging, to capture the input without going via Log4J, so as to keep the output more machine-readable.
These programs need log messages that
- are easy to parse by a regexp or other simple string parse (consider quoting values, etc)
- push out the full exception chain rather than stringify() bits of it
- stay stable across versions
- log the things the tools need to analyse: events, data volumes, errors
For these logging tools, ease of parsing, retention of data and stability over time take the edge over readability. In
HADOOP-5073, Jiaqi Tan proposed marking some of the existing log events as evolving towards stability. As someone who regulary patches log messages to improve diagnostics, this creates a conflict of interest. For me, good logs are ones that help people debug their problems without anyone else helping, and if that means improving the text, so be it. Tools like Chukwa have a different need.
What to do? Some options
- Have some messages that are designed purely for other programs to handle
- Have some logs specifically for machines, to which we log alongside the human-centric messages
- Fix many of the common messages, then leave them alone.
- Mark log messages to be left alone (somehow)