Is the code free of data races (did you try building/running with -race)?
Did you by chance test it with the go1.8beta2? Does it happen with go1.8?

Thanks for the report. Additional question:

• Is there any unsafe or cgo involved?

There are no any cgo and unsafe imports.
I have not tested with -race, so I do it soon, thank you for it.

I cannot reproduce bug on my machine with go1.7. But I do not want to use go1.8 on production.

-race shows only one place - read/write boolean variable. I have rewrite this part of code, and will test it on production soon...

When you say "I cannot reproduce bug on my machine with go1.7" did you mean to say go1.8?

CC @aclements

I cannot reproduce on my machine, because I cannot load it as production server and cannot wait for weeks...

Currently I suggest to suspend this discussion, because compiler used was 1.5 (I open binary and find go1.5 string in it). debian package system did not use 1.7. So I completely recompile my program and put on production. Currently I will wait for 1 month to reproduce bug. If it is not reproduced, then bug should be closed as invalid.

If you were using 1.5 you could be seeing #12932 (?)

It is very possible, because this bug is reason why I try to use 1.7 which is not present in debian 7, so my installation was incorrect.

Based on the line numbers in the traceback, I can confirm that this binary was actually built with Go 1.5.1. Since we're no longer supporting Go 1.5, I'm going to close this bug.

It's likely the @ALTree is right and what you're seeing is #12932. That was fixed in Go 1.6 and backported for Go 1.5.2, so upgrading should fix the problem.