This page contains useful miscellaneous information that may make your life easier while developing Impala.
Capturing Crash Dumps
To automatically save crash dump files you must first set the max size for core dumps as unlimited. When a crash occurs, the core dump will will be placed in your current directory and will be named 'core'. You can then open this file and inspect it with gdb. For example:
ulimit -c unlimited be/build/debug/service/runquery -query="some query that causes crash" gdb be/build/debug/service/runquery ./core # view call stack > (gdb) backtrace
Logging
We use the Google Logging Library. See http://google-glog.googlecode.com/svn/trunk/doc/glog.html
The library defines logging levels of ERROR, INFO and WARNING. We also use verbose logging which can be turned on with with environment variables. e.g:
export GLOG_v=1
Or for specific modules:
export GLOG_vmodule="hdfs*=2,run=1"
You can enable logging in general and turn if off for specific modules that may be too chatty at that level
export GLOG_v=3 export GLOG_vmodule="state-store=0"
We have defined the levels 1,2 and 3 to represent information that is displayed per query/connection, file or row:
#define VLOG_CONNECTION VLOG(1) #define VLOG_QUERY VLOG(1) #define VLOG_FILE VLOG(2) #define VLOG_ROW VLOG(3)
Logging generally goes to files in /logs. The naming is:
/logs/cluster/<program_name>.<host_name>.log.[ERROR|INFO|WARNING].<timestamp>.<pid>
The last instance of a particular program log will be linked to:
/logs/cluster/<program_name>.[ERROR|INFO|WARNING]
VLOG output goes into the INFO file. You can send it to standard error by setting the environment:
export GLOG_logtostderr=1
Existing Hadoop Installations
If you can avoid it, don't install or run another Hadoop on your system. This is an easy source of problems when developing as out-of-date binaries, headers and configuration files can get silently picked up.