Fix fast leader query back-off

[larry-cdn77]

Fix #3690: Metadata leader query timer backoff never applied

[larry-cdn77]

Test checks to see that after 1 second 6 triggers occur of an exponentially backed off 10ms timer (you expect triggers at 10 ms, 30, 70, 150, 310 and 630)

[larry-cdn77]

I don't immediately see why the rdkafka_timer symbols failed to link in the CI build

[edenhill]

Use in code (src/rdtimer.c) unittests to test the private implementation, see for example unittest_broker.

Also make sure the test runs quickly, preferably in under a handful of seconds.

You can run a single unittest by: TESTS=0000 RD_UT_TEST=backoff make in the tests/ directory.

[larry-cdn77]

Thank you. I had neglected the unit test framework. It is a unit test now reduced in duration to one second.

[larry-cdn77]

In order for my test to run reliably I had to round up the microsecond-millisecond conversion of sleep time in rd_kafka_timers_run. Otherwise the value is rounded down and I get a 'dribbling' effect of waiting a bit more, a bit more and a tiny bit more still...

                        sleeptime = rd_kafka_timers_next(rkts, timeout_us,
                                                         0 /*no-lock*/);

                        if (sleeptime > 0) {
                                cnd_timedwait_ms(&rkts->rkts_cond,
                                                 &rkts->rkts_lock,
                                                 (int)(sleeptime / 1000) + 1);
                        }

[cla-assistant[bot]]

All committers have signed the CLA.