Conversation
|
🙈 The PR is closed and the preview is expired. |
memodi
force-pushed
the
dnsname-tracking
branch
from
2145af3 to
5cf49b7
Compare
memodi
force-pushed
the
dnsname-tracking
branch
from
7e1ad4b to
d0d4cbf
Compare
memodi
force-pushed
the
dnsname-tracking
branch
from
d0d4cbf to
605685a
Compare
jpinsonneau
left a comment
jpinsonneau
left a comment
There was a problem hiding this comment.
Looks good, thanks @memodi !
Co-authored-by: Amogh Rameshappa Devapura <aramesha@redhat.com>
| options ndots:5 | ||
| ``` | ||
|
|
||
| Short DNS names for cluster services causes high load on the cluster DNS service |
There was a problem hiding this comment.
typo
| Short DNS names for cluster services causes high load on the cluster DNS service | |
| Short DNS names for cluster services cause high load on the cluster DNS service |
| resulting in higher latencies, negative caching and increased dns traffic. This | ||
| negative impact can be prevented by using Fully Qualified Domain Name (FQDN) in | ||
| the requests. After updating the hostname to `nginx.server.svc.cluster.local.` | ||
| in the curl requests, we are not seeing any NXDOMAINS and reduced unnecessary | ||
| dns traffic in our cluster. You can imagine the performance impact if such | ||
| configuration issue propagated to hundreds of services in your cluster. | ||
|
|
There was a problem hiding this comment.
"DNS" is sometimes written in capitals sometimes not, I would suggest to stick to just one form (across the full post)
| issues, it comes with some caveats to favor performance. This feature isn't | ||
| supported with Prometheus as datastore since storing DNS names as metric values | ||
| could cause high cardinality. That means, if you're looking to use this feature |
There was a problem hiding this comment.
On cardinality, maybe we can tell that we're currently evaluating the impact - see my comment here - I think eventually we can add that to the metrics
|
/lgtm |
|
@memodi you can merge when you want |
stleerh
left a comment
stleerh
left a comment
There was a problem hiding this comment.
Maybe it's obvious, but you could mention that this shows only internal DNS names in your cluster. That is, it is not tracking outbound DNS requests, so you won't see how many requests are for youtube.com. 😏
| you must use Loki as your datasource. We're actively working to measure the | ||
| performance impact and expose DNS names as Prometheus metrics, though. | ||
|
|
||
| Captured DNS names will be truncated at 32 bytes to balance the |
There was a problem hiding this comment.
30 bytes
(Note: Although the code allocates 32 bytes, one is for the nil character, and another used to store the length of the first DNS label.)
That's not true, it could track outbound DNS requests as well, it should just get initiated inside the cluster. |
Oh, I probably didn't see it on my cluster because I had sampling turned on. |
@memodi @stleerh |
@jotak yes, I had observed that during testing, see here netobserv/netobserv-ebpf-agent#820 (comment) where @jpinsonneau mentions reasons other reasons why it could be empty. That's why I always need to add a UI filter |
6 participants
DNS name tracking blog