fix(sql): handle commit/rollback error when tracing transcation

[kagaya85]

Now, sql plugin will close a span when commit/rollback returns error, it may call span.End() repeatly and cause a panic, we want return the database/sql errors to the upper instead of get a panic.

Add defer function to handle the return error and will not call span.End() when error is not nil

fix #38

[kagaya85]

How about add the error message to the span and close it even the commit failed, to avoid unexpected panic, this may need add a method like span.IsEnd() to check if the transcation span is closed. If we don't call span.End() when commit failed, the previous related span in this transcation period may not be reported.

[wu-sheng]

Now, sql plugin will close a span when commit/rollback returns error, it may call span.End() repeatly and cause a panic, we want return the database/sql errors to the upper instead of get a panic.
Add defer function to handle the return error and will not call span.End() when error is not nil

I think we should consider separating the span, because as you mentioned in the codes, if there is connection leak(commit error, but no rollback called) or rollback fails, the span wouldn't be closed as expected, which is unexpected.

The tracing system is designed to tell the truth, rather than not working as well as original codes.

You could check the SkyWalking Java agent's JDBC plugin, commit and rollback have there own spans, then no panic and always presents the complete trace.

[wu-sheng]

LGTM, wait for @arugal

[kagaya85]

Please hold this PR, we consider using separate Begin/Commit/Rollback span instead of the Tx span, in addition, there are two considerations may need to discuss:

• After using the separate span, the trace will not show the entire execution process for each transcation directly, for example, understanding one transcation cost time will become difficult
• Span validity check is also needed in this PR, even in septarate tx span, we need to ensure the parent span is still valid when start a new span to avoid panic, I can improve this in another PR

[wu-sheng]

After using the separate span, the trace will not show the entire execution process for each transcation directly, for example, understanding one transcation cost time will become difficult

You could calculate the whole time by using tag or log in the span, such as commit or rollback. Also, the backend provides the logic endpoint concept(logic name + duration in tags) for you to measure things like this.

Span validity check is also needed in this PR, even in septarate tx span, we need to ensure the parent span is still valid when start a new span to avoid panic, I can improve this in another PR

You could make the choice whether in this PR or through another one.