Support arbitrary network dialers

[fumin]

One such use case is to use the library within the constraints of Google
App Engine
https://developers.google.com/appengine/docs/go/sockets/reference

[johto]

This line doesn't appear to work in Go 1.0.

[johto]

I've added some comments about the implementation, but in general I consider the idea workable.

[fumin]

Thanks for the comments, the part on the DialTimeout case is especially important. I have also ensured backwards compatibility down to go1.0.2 .

[johto]

My apologies, @fumin, I wasn't aware you had updated this PR. On a quick look it looks like you've addressed all my comments, so I'll have another look this week. Sorry for the delay.

[fumin]

@johto no worries, I have resolved the conflicts due to changes in the master repo in the meantime. However, I see that the build failed for the combination of go1.2 and pg9.3 with a connection timeout error. I ran this combination a couple of times locally and every time all tests passed. Could the failure be due to just a temporary glitch in the build cluster?

[johto]

However, I see that the build failed for the combination of go1.2 and pg9.3 with a connection timeout error. I ran this combination a couple of times locally and every time all tests passed. Could the failure be due to just a temporary glitch in the build cluster?

I suppose that's possible. But it is a bit suspicious that after changing how dialing works connecting suddenly times out.. I don't see anything obvious in the code, though, but I'll have a closer look later.

In the meanwhile, is there a reason to require a pointer receiver in defaultDialer? It seems completely unnecessary.

[fumin]

Right, it might be clearer for defaultDialer to be a struct instead of a pointer, I've updated this change. However, this time the build failed on a different combination go1.3 + pg8.4 on a different test case TestListenerFailedQuery. I ran this combination on my computer and scrutinized the code a few more times but couldn't find anything decisive except noticing that errRecover has been replaced. Not being able to reproduce the error on my machine makes it hard for me to investigate, I wonder are there other ways to reproduce the problem?

[johto]

Right, it might be clearer for defaultDialer to be a struct instead of a pointer, I've updated this change. However, this time the build failed on a different combination go1.3 + pg8.4 on a different test case TestListenerFailedQuery. I ran this combination on my computer and scrutinized the code a few more times but couldn't find anything decisive except noticing that errRecover has been replaced. Not being able to reproduce the error on my machine makes it hard for me to investigate, I wonder are there other ways to reproduce the problem?

I can't see anything wrong with the code, nor can I reproduce the problem in any of the environments. I've committed this in 5ff334f.

Thanks!