Jan, I agree with you philosophically. I think one practical challenge has
to do with data formats. Many people use untyped events, so there is simply
no guarantee on the form of the input. E.g. many companies use JSON without
any kind of schema so it becomes very hard to assert anything about the
input which makes these programs very fragile to the "one accidental
message publication that creates an unsolvable problem.

For that reason I do wonder if limiting to just serialization actually gets
you a useful solution. For JSON it will help with the problem of
non-parseable JSON, but sounds like it won't help in the case where the
JSON is well-formed but does not have any of the fields you expect and
depend on for your processing. I expect the reason for limiting the scope
is it is pretty hard to reason about correctness for anything that stops in
the middle of processing an operator DAG?


On Fri, Jun 2, 2017 at 4:50 AM, Jan Filipiak <[EMAIL PROTECTED]>