[bind10-dev] Losing startup logs

[Michal 'vorner' Vaner]

Hello

On Fri, Mar 30, 2012 at 10:09:47AM -0700, JINMEI Tatuya / 神明達哉 wrote:
> I'd also note that the dump overhead at the highest log level may
> be an issue.  On startup it may not matter much, but if, e.g., it
> doubles startup time we probably don't want to do that.

I don't think it is much problem. We need to store just few messages at the
startup. We need to store them until we load the configuration and that is
pretty soon. All the expensive actions (like loading zones, etc.) happen after
that and the time for startup before is more bounded by communication
roundtrips. The full log of startup with -v is 30kB/138 lines here over all the
component, which contains all the post-config logs as well. I think this is
safe.

Anyway, I want to filter the log messages when dumping them by the new logging
level.

> This also leads to the usage of emergency-case.  Since we cannot
> predict when (or even whether) that happens, we'd need to keep logging
> (while discarding older ones) in the buffer all the time at the
> highest log level.  I suspect it's too heavy for busier servers.  By
> "if the buffer is enabled" do you mean it's off by default?  But then
> the feature would be less useful because it would be most helpful for
> things like unpredictable crash.

Yes, I believe logging all the messages by default would make it very slow. I
still believe it would be order of magnitude faster than writing it to disk and
you would not have to get through all the messages in the file. So it would be
nice feature for testing it in lab or when someone with a small personal domain
keeps a DNS server somewhere. I don't think this would be usable for busy
servers, though.

> Regardless of when to enable it, one possibility of mitigating
> overhead would be to just log message IDs, instead of all textual
> messages including textual representation of parameter objects.  Of
> course it will lose a substantial amount of information, but it will
> also be much more efficient, and at least give a more detailed hint
> (than, e.g, a stack trace) that what program did before the bad thing
> happened.

This could be one way, but I think we'd still want to be able to do the full log
way. At least for the startup or for debugging problems.

With regards

> Regarding whether the emergency-case logging is worthwhile, I'm not
> yet so sure.  I see the potential of its usefulness, but I have many
> questions about the feasibility as stated above, too.

Yes, something that wasn't done before or we don't know about it. It's hard to
guess sometimes if it would work. That's one reason why I proposed to have it
off by default.

With regards

-- 
Anyone who goes to a psychiatrist ought to have his head examined.
		-- Samuel Goldwyn

Michal 'vorner' Vaner
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: Digital signature
URL: <https://lists.isc.org/pipermail/bind10-dev/attachments/20120402/5227fe92/attachment.bin>