We should add more fault injection checkpoint tests. The tests should:

• Test a (small) corpus of input, pipeline, and expected output. Test with mongodb and kafka source and sink.
• Inject faults in key places in operator,source,and sink code.
  • operator-- send data msg
  • sink-- failure in middle of batch
  • source-- failure in between reading doc and updating offsets
• Validate the output and stats is correct even when faults occur.

This is similar to what some existing tests do in checkpoint.js.