Async futures issue

[mlanzuisi]

Hello,

while trying to implement the async version of command with redis-plus-plus library, I tried making something like the following:
1 - each command sent has a future return, as following

std::future *myfuture;
myfuture = new std::future;
*myfuture = m_async_redis_cluster->set(key_sw, value_sw, std::chrono::seconds(seconds));

2 - each future is added to the last position of a vector, this way

m_pending_bool_futures.push_back(std::move(*new_future));

3 - after a number of operations, results are read this way:

bool ret = m_pending_bool_futures.front().get();
m_pending_bool_futures.erase(m_pending_bool_futures.begin());

The problem is: when I start calling this last part (the removal of the futures), during the "get()" I have the following exceptions:

Caught exception "connection is closing" (only 1, the first error)
Caught exception "No associated state" (many subsequent messages)

Sometimes I receive the
Caught exception "failed to connect to Redis (anritsu-redis-server-0.svc-redis:6379): Timeout"

as first message instead of the "connection is closing" one.

What can this be? Some libuv connection timeout or anything else?

[sewenew]

Sorry, but I didi some tests, and cannot reproduce the problem.

Can you use the latest code on master branch to have a test? If you still have the problem, please give a minimum compilable code snippet that can reproduce the problem. So that I can debug it. Thanks!

Regards

[mlanzuisi]

Hi,

try with attached file.
main_futures.zip

[sewenew]

• "failed to connect to Redis (anritsu-redis-server-0.svc-redis:6379): Timeout" means your set command is timed out.
• "connection is closing" means AsyncRedisCluster is destroyed (destructor is called) before the queued commands are sent to Redis.

I did some research on your code, and it's buggy.

bool ret = m_pending_bool_futures.front().get();       // ---------> (1)
std::printf("Ret set \"%u\"\n", ret);
m_pending_bool_futures.erase(m_pending_bool_futures.begin());   // -----> (2)

If (1) throws, (2) won't be executed, i.e. the front element won't be erased but already been call with future::get. The loop continues, (1) execute again with the original front element. However, since it has already been called with future::get, the behavior is undefined.

As you described above that your code catches exception, either timeout or connection is closing, you'll run into the problem, i.e. call future::get twice on the same future object. That's why you see No associated state.

In order to fix the problem, you can try the following pseudo code:

for (int i = 0; i < key_no; i++) {
  try {
     m_pending_bool_futures.push_back(m_async_redis_cluster->set(key, value));
     if (m_pending_bool_futures.size() > FUTURE_LIMIT / 2) {
        try {
          bool ret = m_pending_bool_futures.front().get();
        } catch (const Error &) {
            // error handling
        }
        // always erase it, no matter whether we get exception or not.
        m_pending_bool_futures.erase(m_pending_bool_futures.begin());
     }
  } catch (const Error &e) {
     // error handling
  }
}
// ensure all future have been called with `get`
for (auto &fut : m_pending_bool_futures) {
  try {
     bool ret = fut.get();
  } catch (const Error &e) {
  }
}

Sorry for the late reply...

Regards

[mlanzuisi]

Thank you for your reply.
After changing what you described, I always have the same errors, either "failed to connect to Redis" or "connection is closing", but "No associated state" disappeared, thank you.

Anyway, it seems that if I start using set() and get() functions too fast, there are these errors.
They disappear completely if I use a "sleep(1)" function between the AsyncRedisCluster constructor and the loops of get/set,

[sewenew]

You should ensure all future::get have been called before AsyncRedisCluster is destroyed. Otherwise, you'll get "connection is closing", since the destructor of AsyncRedisCluster will close the underlying connection.

Regards

[mlanzuisi]

Thank you, issue can be considered closed