cpp-client loses spans on short-lived processes; Close() not flushing buffers?

[jaegertracing]

cpp-client loses spans on short-lived processes; Close() not flushing buffers?

[ringerc]

Related #53

[ringerc]

RemoteReport::close() makes no effort to flush by the looks:

void RemoteReporter::close()
{
    {
        std::unique_lock<std::mutex> lock(_mutex);
        if (!_running) {
            return;
        }
        _running = false;
        lock.unlock();
        _cv.notify_one();
    }
    _thread.join();
}

[ringerc]

It probably makes sense to log dropped buffers on close instead, with a hint that an explicit flush before close is advised (using #53 for flush). But that'd really want the opentracing API to adopt the Flush method.

[rnburn]

The Close method is meant to be optional; so if you don't want to wait on the tracer to flush all it's spans you can just not call it.

I think the right solution is for jaeger to move the current code in close to its destructor and change close to flush the spans, but require it to be called explicitly.

[ringerc]

@rnburn Makes sense. I'll amend my push that adds an optional Tracer.flush() to Jaeger's tracer so that it also calls it on explicit Close(), but from the dtor.

@isaachier I've written a small test program to reproduce this bug. It reliably loses spans. ringerc@2a36dd5 (on master) or ringerc@ff654d9 (on my fix/feature integration branch, but not using flush).

jaegertracing::Tracer::flush() does not seem to be working/sufficient. If you uncomment

        //static_cast<jaegertracing::Tracer*>(tracer.get())->flush();

in src/jaegertracing/shortlived.cpp, it still often loses spans. Wireshark will show that the span was never sent.

[ringerc]

Additional test runs show that without flush() it loses spans pretty much 100% of the time.

If I uncomment flush() on my integration test branch it often takes several seconds to exit so I'm not waking a thread properly or something.

Anyway, current behaviour clearly broken.

[ringerc]

I pushed an improved test case (ringerc@e5a6234). It clearly shows that we lose spans unless:

• We explicitly flush() with ./src/jaegertracing/shortlived "" 1 and the flush actually works (per below)
• We sleep > 10 seconds with ./src/jaegertracing/shortlived "" 0 11

The flush must actually take effect. Sometimes it seems to get missed;

Test with opname "shortlived--flush-sleep0s, flush 1
flushing...  flushed (flush took 0ms)
not sleeping before exit 

in which case the span is also lost. So there's something wrong with my flush method. Locking issues?

Lowering the configured bufferFlushInterval reduces the necessary sleep, it's obviously an issue with not waking the flush thread properly.

@isaachier I could use advice at this point, my C++ and threading is ... limited.

I also have no idea how to integrate this into the regression tests. Crossdock somehow?

[isaachier]

In general there might be a problem with blocking IO and threading. I'm about to call it a night, but will try to look at this tomorrow. Worst case, will be able to look over the weekend.

[isaachier]

I think I have a simpler approach for this specific issue. About the flush method being exposed overall, that can be a new PR. See #59 for first fix.