GH-149: Improve error messages on write thread failure

[C0urante]

Addresses #149.

The KCBQThreadPoolExecutor is modified to only track one write thread exception at a time, which allows us to include a complete stack trace of that exception when failing the task.

The "Attempted to reduce batch size below 1." error is rewritten to include more information on a potential root cause and make the attached cause of the exception clearer.

Some cleanup of unnecessary exception wrapping when instantiating various AbstractConfig subclasses is also made.

[ddasarathan]

From what I understand, all threads can throw exceptions but only the first gets logged and that bubbles up and kills the task. Since the exception will be thrown, we do not want to reset the encounteredError atomic ref?

[ddasarathan]

LGTM

[C0urante]

From what I understand, all threads can throw exceptions but only the first gets logged and that bubbles up and kills the task. Since the exception will be thrown, we do not want to reset the encounteredError atomic ref?

We do not reset the reference because the exception may be used multiple times. We throw exceptions encountered from write threads in both BigQuerySinkTask::flush (as part of the call to KCBQThreadPoolExecutor::awaitCurrentTasks) and at the very beginning of BigQuerySinkTask::put; the former prevents us from committing offsets for data that we weren't able to write to BigQuery (see #68), and the latter causes the task to fail (since throwing an error from SinkTask::flush doesn't actually cause a task to fail).

[C0urante]

@ddasarathan thanks for taking a look. Given the LGTM I'll merge but if you are unsatisfied by my explanation of the way we use KCBQThreadPoolExecutor.encounteredError please let me know; happy to revert if something isn't right here or file a follow-up if there's room for improvement.