Here's a similar recent trace from freebsd/arm:

https://build.golang.org/log/372776e543edae8bfda51d2ec70146a943eb71b6

CL https://golang.org/cl/24550 mentions this issue.

Happened again at https://storage.googleapis.com/go-build-log/44200ef6/freebsd-386-gce101_566a8190.log via https://golang.org/cl/27200 trybot failure

@kardianos, maybe you can take a look at this too? (if you don't have enough on your plate)

But I figure all this stuff is currently more in your head than mine.

@bradfitz I'll look into this shortly. Thanks.

I'm able to reproduce this. By adding the following:

+++ b/src/database/sql/sql_test.go
@@ -1509,6 +1509,7 @@ func TestPendingConnsAfterErr(t *testing.T) {
        unblock := make(chan struct{})
        setHookOpenErr(func() error {
                <-unblock // block until all connections are in flight
+               runtime.Gosched()
                return errOffline
        })

It reliably fails with sql_test.go:148: 1 connections still open after closing DB. However adding runtime.Gosched() to the last line of the func allows the test to pass again (this is all on a std linux amd64 laptop).

My assumption is that we are seeing the failures on slower machines that have slightly different scheduler patterns due to load. I'll look into updating the test soon.

CL https://golang.org/cl/33418 mentions this issue.