Read/write out-of-order bug? about memjs HOT 7 CLOSED

Interesting. OK. This is solvable by using the opaque field in the header. I'll get to it.

from memjs.

Aha! The problem is not with memcached responding out of order, but rather with memjs requesting out of order.

What's going on is that since the there is no connection yet (it's established lazily), the first operation (after enqueing the response and error) blocks on the connection to be established. Meanwhile, example.js continues on it's merry way and performs the second operation. Now the connection has been set, so it immediately proceeds to write the request to the socket. Only now do we switch back to the first operation. So the requests are sent out of order.

Nesting both operations within a higher level operation solves this:

client.get('dummy', function() {
  client.get('AAA');
  client.set('BBB', 'ZZZ');
  client.close(); // close really needs to be nested to ensure ordering, but at least on localhost this happens to work
});

In principal, opaque would solve this, but it turns out to be kind of a messy solution that I'd like to avoid if we can rely on ordering. I think this is more easily fixable by mucking with the connection code a bit (specifically sticking requests in a common buffer that's flushed upon the first request that sees a connected socket).

Thanks for finding this! It's a kind of subtle one, but very bad and important to fix.

from memjs.

Ah yes, I guess memcached physically can't respond out of order, so it can't be a problem with out of order responses. And whilst memjs might use multiple memcached servers, each of them will have a separate queue, so any queue-based fix should work for them all. Thanks for looking at this. Look forward to the fix!

from memjs.

@alevy Is this being worked on? Do you have any sort of idea how long it might take?

from memjs.

@ithinkihaveacat thanks for the nudge. My original (to actually write to the socket lazily) turned out to be complicated, but I did something dumber, but that I'm pretty sure works and is fairly clean. Check out the lazyconn branch or pull request #22.

If you can give this a test before i merge that would be awesome!

from memjs.

Thanks! I had a bit of a quick test tonight, and it seems okay so far.

from memjs.

Looks good to me! Thanks!

from memjs.