Wrap consumer.poll() for KafkaConsumer iteration

[dpkp]

Now that heartbeats have been moved to a background thread, the pythonic iterator can just wrap consumer.poll() and no longer needs to worry about breaking from iteration to support sending background requests. With this approach, some consumers may need to tweak max_poll_records and/or max_poll_interval_ms, such that consumer's average per-record processing time * max_poll_records < max_poll_interval_ms. Defaults remain at 500 max_poll_records and 5mins for max_poll_interval. This should be sufficient for the vast majority of consumers.

My local testing suggests that this change fixes consumer performance issues reported in #1888 .

Given the reduction in overall complexity, this change may also improve #1672 (though won't completely fix).

Because this is a relatively significant change to internals, I have added a configuration option to allow reverting back to the <= 1.4.6 iterator logic:

consumer = KafkaConsumer(bootstrap_servers=[...], legacy_iterator=True, ...)

---

This change is 

[jeffwidman]

Source code change looks relatively straightforward to me.

I question whether adding a feature flag is worth the trouble unless you truly think there's a significant chance it will need to be reverted. Folks can always revert to 1.4.6 if needed, although I suppose 1.4.7 will include a number of other changes so this makes it simpler to isolate the cause of problems.

Looks like Travis is failing, may be related to #1891 and https://github.com/dpkp/kafka-python/pull/1886/files#r316860737.

[dpkp]

I question whether adding a feature flag is worth the trouble unless you truly think there's a significant chance it will need to be reverted. Folks can always revert to 1.4.6 if needed, although I suppose 1.4.7 will include a number of other changes so this makes it simpler to isolate the cause of problems.

Ok, I'm convinced -- I updated to make the new poll()-based iterator the default. If there is a problem, users can revert to prior implementation by setting legacy_iterator=True in KafkaConsumer.

Looks like Travis is failing, may be related to #1891 and https://github.com/dpkp/kafka-python/pull/1886/files#r316860737.

I think this was related to deciding whether to update offsets in the fetcher (we need to do this for standard poll(), but not when wrapping with the iterator). I've pushed a fix and will see whether the next test run is clean.

Thanks for the review!