In PySpark, how can I log to log4j from inside a transformation

Question

I want to log to the standard logger inside an executor during transformation with log levels and formatting respected. Unfortunately I can\'t get access to the log4j logger obj

Answer 1

After a few hours of digging into the spark repository, it seems this is impossible to achieve currently. The executor doesn't actually have a jvm instance it's attached to, the data is just streamed over the socket without a jvm native binding to utilize.

Here's the worker creation code that streams the error messages to stderr:

private def createSimpleWorker(): Socket = {
  ...
  val worker = pb.start()

  // Redirect worker stdout and stderr
  redirectStreamsToStderr(worker.getInputStream, worker.getErrorStream)

  ...
}

/**
 * Redirect the given streams to our stderr in separate threads.
 */
private def redirectStreamsToStderr(stdout: InputStream, stderr: InputStream) {
  try {
    new RedirectThread(stdout, System.err, "stdout reader for " + pythonExec).start()
    new RedirectThread(stderr, System.err, "stderr reader for " + pythonExec).start()
  } catch {
    case e: Exception =>
      logError("Exception in redirecting streams", e)
  }
}

And here's the worker.py code for communicating the job processing. There's no place to emit log messages or message type which indicates a log event.

try:
    ...
    command = pickleSer._read_with_length(infile)
    if isinstance(command, Broadcast):
        command = pickleSer.loads(command.value)
    func, profiler, deserializer, serializer = command
    init_time = time.time()

    def process():
        iterator = deserializer.load_stream(infile)
        serializer.dump_stream(func(split_index, iterator), outfile)

    if profiler:
        profiler.profile(process)
    else:
        process()
except Exception:
    try:
        write_int(SpecialLengths.PYTHON_EXCEPTION_THROWN, outfile)
        write_with_length(traceback.format_exc().encode("utf-8"), outfile)
    except IOError:
        # JVM close the socket
        pass
    except Exception:
        # Write the error to stderr if it happened while serializing
        print("PySpark worker failed with exception:", file=sys.stderr)
        print(traceback.format_exc(), file=sys.stderr)
    exit(-1)
finish_time = time.time()
report_times(outfile, boot_time, init_time, finish_time)
write_long(shuffle.MemoryBytesSpilled, outfile)
write_long(shuffle.DiskBytesSpilled, outfile)

# Mark the beginning of the accumulators section of the output
write_int(SpecialLengths.END_OF_DATA_SECTION, outfile)
write_int(len(_accumulatorRegistry), outfile)
for (aid, accum) in _accumulatorRegistry.items():
    pickleSer._write_with_length((aid, accum._value), outfile)
...

And finally the message types available:

class SpecialLengths(object):
    END_OF_DATA_SECTION = -1
    PYTHON_EXCEPTION_THROWN = -2
    TIMING_DATA = -3
    END_OF_STREAM = -4
    NULL = -5

Answer 2

Have a look at this question

Similar situation

You can get your map function to return you an object that can contain a stack trace string or a real object, and a bool flag stating if there was an error. This can be useful to debug a task that has side effects, or if you have specific data conditions that cause failures.