nsqd: bump go-svc to properly handle SIGTERM

[judwhite]

Fixes #756

Relevant change: judwhite/go-svc@be413dc#diff-2ac5da282c671a5f874701b939d24f13L16

Travis: https://travis-ci.org/judwhite/go-svc/builds/127990343

/cc @mpe

[mreiferson]

thanks for the quick fix @judwhite

[mreiferson]

will issue a release, since this is a critical fix.

[mreiferson]

cc @jehiah

[jehiah]

👍

[ploxiln]

I did some experimentation (in preparation of rolling out a new build which is now running fine in production) and found that this issue is not as serious as it sounds (to someone less familiar with nsqd internals, like me).

nsqd 0.3.7 saves its metadata whenever a new channel is created, so even if it fails to save on exit, the topics/channels metadata is pretty much always up-to-date anyway, and new channels persist after an ungraceful restart.

[mreiferson]

@ploxiln the problem is that it also won't persist messages in memory to disk

[ploxiln]

yeah, that's bad, ignore me 👅

[mreiferson]

for anyone following along at home, please read:

https://groups.google.com/forum/#!topic/nsq-users/yOckPLUWi4Q

[mreiferson]

@judwhite thinking about this more, particularly the genesis of the regression, it comes down to the fact that this signal handling code shouldn't be part of go-svc.

Applications depending on go-svc should pass in a channel that can trigger "exit". I even pointed this out on the original PR, but I neglected to confirm that the code moved over without modification 😦 .

[judwhite]

My mistake.

[mreiferson]

It's all good, shit happens, I missed it too. Would it be possible to make those changes?

[judwhite]

Yeah, I'll open a PR at go-svc and a corresponding one here to pass the signals. I'll tag you when it's ready.