On 07/19/2013 12:15 PM, Jakub Hrozek wrote:
>On Mon, Jul 08, 2013 at 09:43:42PM +0200, Michal Židek wrote:
>>On 07/04/2013 07:45 PM, Simo Sorce wrote:
>>>On Thu, 2013-07-04 at 16:23 +0200, Michal Židek wrote:
>>>>https://fedorahosted.org/sssd/ticket/1966 (SSSD failover doesn't work
if
>>>>the first DNS server in resolv.conf is unavailable).
>>>>
>>>>The problem here is, that if first nameserver in resolv.conf is down,
>>>>the resolution is too slow and SSSD will not wait for the result of
>>>>ares_search and go offline. In my case the resolution was sometimes more
>>>>than a minute, because all search domains in resolv.conf were searched
>>>>inside ares_search() call using the first (not working) nameserver in
>>>>the first place and then with the working nameserver (and before that,
>>>>SSSD tried to figure out the domain name from my incorrectly set
>>>>hostname, which added more unnecessary DNS lookups).
>>>>
>>>>To avoid this problem, the option dns_discovery_domain must be set
>>>>properly, so that only the correct domain is searched, but even that is
>>>>not enough, because the default timeout for dns resolver operation in
>>>>sssd is too low. This patch rises the default value to 15 seconds
>>>>(instead of 5 seconds).
>>>>
>>>>Another option might be to lower the amount of time ares waits for a
>>>>nameserver to respond (currently it is 5 seconds, that is why 5 second
>>>>for the entire dns resolution is not sufficient), but I do not want to
>>>>do this.
>>>
>>>Why not ?
>>>
>>>5 seconds these days is quite awfully high all in all.
>>>I think 1 second should be plenty in most cases, so maybe a default of 2
>>>seconds should be considered.
>>>
>>
>>I also think that 5 seconds is high value, but in general, I have bad
>>feeling when lowering not configurable timeouts (theoretically it could
>>cause some DNS servers that are reachable now to become unreachable).
>>
>
>Right, but in practice I can't think of a valid network configuration
>where the name server would take so long to reply. I mean, if the
>network was so slow that just resolving name would take 5 seconds, then
>chances are we wouldn't be able to download the entry from LDAP at all.
>And if the nameserver is slow, then let's rather skip it and fail over
>to another one.
>
>>Another possibility would be to make this value configurable. Then I
>>would have no problem lowering the default value. But, should this be
>>configurable? We already have one option to configure timeout for the
>>entire DNS resolution, adding another timeout for subtaks of this
>>operation does not sound nice to me. But I do not have strong opinion
>>on this.
>>
>
>Here is what we discussed with Michal today:
> * make the per-nameserver timeout lower a little. Now it's 5
> seconds, which is too long especially considering the failover
> timeout is 5 seconds as well. We could change it to 2 or 3
> seconds.
> * the glibc allows 3 nameservers. The failover request timeut could
> then default to 3*nameserver_timeout to give the resolver a chance
> to iterate over all the configured name servers.
>
>>Btw, once the correct data is returned from DNS servers, we no longer
>>require any DNS operation, so it only happens first time the service
>>is required (for example the first 'getent' call will be slow, but
>>after that the service is fast). This is not so bad IMO.
>>
>>>>These patches also change man pages, so probably master only (string
>>>>freeze)? Even if this is a really small change.
>
>We accept changes again now that we've released 1.10.1.
>
>But this change would be OK even for a string-frozen release, because it
>only changes the number, not translatable strings.
>
>>>>
>>>>I was also thinking, would it make sense to write a warning to the logs
>>>>if the dns_discovery_domain option is not set? It seem to be important
>>>>to set it properly for cases like this one.
>>>
>>>I wonder, should we use the machine name to automatically set the
>>>discovery domain, and only if that fails fall back to the ones defined
>>>in resolv.conf ?
>>
>>Actually, we do this. Wireshark shows that the first srv lookup was
>>constructed from my hostname. But the whole process is unclear to me,
>>for example I do not know why we call the resolv_gethostbyname_send()
>>inside of resolv_get_domain_send(). It adds another DNS lookups
>>for A records (which causes another slow-down in the case of bad
>>nameserver searched first). In wireshark I also see searches for A/AAA
>>records for names like:
>>mzidek.example.com
>>mzidek.example.com.searchdom1.com
>>mzidek.example.com.searchdom2.com
>>mzidek.example.com.searchdom3.com
>
>IIRC we try to canonicalize the host name before chopping of the first
>part and using the rest as domain name.
>
>Would using gethostname() be safe in real world environments?
>I know that host names are often set to complete garbage and can't be
>relied on..but is it OK to just say "fix your environment" in that case?
>
>Personally I would argue the current canonicalization doesn't add much,
>because if the host name is set to complete garbage then the canonicalization
>wouldn't magically fix it..
>
>>
>>where
mzidek.example.com is my hostname and searchdom1/2/3 are search
>>domains specified in resolv.conf. Why do we even try to resolve these
>>names? It seems like waste of time to me. If the dns_discovery_domain is
>>properly set, these lookups do not appear.
>>
>>Thanks
>>Michal
>>
>>PS: Patches are the same as before.
>>
>>>
>>>Simo.
>>>
>>
>
New patches attached. As discussed above, I changed the timeout to
contact DNS server from 5s to 2s, and timeout for the entire DNS resolution
from 5 to 6 seconds (instead of 15s as proposed originally)
(so it has enough time to contact 3 nameservers).
I tested it with one and two unreachable nameservers.
Thanks
Michal
Works for me, too, so I'll ACK the patches. But I'll leave them on the
list for a day longer in case somebody else would like to voice their
concern about changing the default timeout value. I personally think
it's OK.