This preference list matches is used to pick prefered field from k8s
node object. It was introduced in metrics-server 0.3 and changed default
behaviour to use DNS instead of IP addresses. It was merged into k8s
1.12 and caused breaking change by introducing dependency on DNS
configuration.
[stackdriver addon] Bump prometheus-to-sd to v0.5.0 to pick up security fixes.
[fluentd-gcp addon] Bump fluentd-gcp-scaler to v0.5.1 to pick up security fixes.
[fluentd-gcp addon] Bump event-exporter to v0.2.4 to pick up security fixes.
[fluentd-gcp addon] Bump prometheus-to-sd to v0.5.0 to pick up security fixes.
[metatada-proxy addon] Bump prometheus-to-sd v0.5.0 to pick up security fixes.
Fix three issues with the fluentd-gcp liveness probe:
h1. STUCK_THRESHOLD_SECONDS was overridden by LIVENESS_THRESHOLD_SECONDS
if defined
Probably a copy/paste issue introduced in edf1ffc074
h1. `[[` is [a bashism](https://stackoverflow.com/a/47576482), and will always failed when called with `/bin/sh`
Introduced by a844523c20
Given that we call the liveness probe with `/bin/sh`, we cannot use the
double-bracketed `[[` syntax for test, as it is not POSIX-compliant and
will throw an error.
Annoyingly, even through it prints an error, `sh` returns with exit code 0
in this case:
```bash
root@fluentd-7mprs:/# sh liveness.sh
liveness.sh: 8: liveness.sh: [[: not found
liveness.sh: 15: liveness.sh: [[: not found
root@fluentd-7mprs:/# echo $?
0
```
Which means the liveness probe is considered successful by Kubernetes,
despite failing to test things as it was intended. This is also
probably the reason why this bug wasn't reported sooner :)
Thankfully, the test in this case can just as easily be written as
POSIX-compliant as it doesn't use any bash-specific features within the
`[[` block.
h1. Buffers are transient and cannot be relied upon for monitoring
Finally, after fixing the above issue, we started seeing the fluentd
containers being restarted very often, and found an issue with the
underlying logic of the liveness probe.
The probe checks that the pod is still alive by running the following
command:
`find /var/log/fluentd-buffers -type f -newer /tmp/marker-stuck -print -quit`
This checks if any _regular_ file exists under `/var/log/fluentd-buffers`
that is more recent than a predetermined time, and will return an empty
string otherwise.
The issue is that these buffers are temporary and volatile, they get created and
deleted constantly. Here is an example of running that check every second on a
running fluentd:
```
root@fluentd-eks-playground-jdc8m:/# LIVENESS_THRESHOLD_SECONDS=${LIVENESS_THRESHOLD_SECONDS:-300};
root@fluentd-eks-playground-jdc8m:/# STUCK_THRESHOLD_SECONDS=${LIVENESS_THRESHOLD_SECONDS:-900};
root@fluentd-eks-playground-jdc8m:/# touch -d "${STUCK_THRESHOLD_SECONDS} seconds ago" /tmp/marker-stuck;
root@fluentd-eks-playground-jdc8m:/# touch -d "${LIVENESS_THRESHOLD_SECONDS} seconds ago" /tmp/marker-liveness;
root@fluentd-eks-playground-jdc8m:/# while true; do date ; find /var/log/fluentd-buffers -type f -newer /tmp/marker-stuck -print -quit ; sleep 1 ; done
Fri Feb 22 10:52:57 UTC 2019
Fri Feb 22 10:52:58 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827964ccf4c7004103c3fa7c8533f85.log
Fri Feb 22 10:52:59 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827964ccf4c7004103c3fa7c8533f85.log
Fri Feb 22 10:53:00 UTC 2019
Fri Feb 22 10:53:01 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827964fb8b2eedcccd2763ea7775cc2.log
Fri Feb 22 10:53:02 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827964fb8b2eedcccd2763ea7775cc2.log
Fri Feb 22 10:53:03 UTC 2019
Fri Feb 22 10:53:04 UTC 2019
Fri Feb 22 10:53:05 UTC 2019
Fri Feb 22 10:53:06 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827965564883997b673d703af54848b.log
Fri Feb 22 10:53:07 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827965564883997b673d703af54848b.log
Fri Feb 22 10:53:08 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer/buffer.b5827965564883997b673d703af54848b.log
Fri Feb 22 10:53:09 UTC 2019
Fri Feb 22 10:53:10 UTC 2019
Fri Feb 22 10:53:11 UTC 2019
Fri Feb 22 10:53:12 UTC 2019
Fri Feb 22 10:53:13 UTC 2019
Fri Feb 22 10:53:14 UTC 2019
Fri Feb 22 10:53:15 UTC 2019
Fri Feb 22 10:53:16 UTC 2019
```
We can see buffers being created, then disappearing. The LivenessProbe running
under these conditions has a ~50% chance of failing, despite fluentd being
perfectly happy.
I believe that check is probably ok for fluentd installs using large
amounts of buffers, in which case the liveness probe will be correct more
often than not, but fluentd installs that use buffering less intensively
will be negatively impacted by this.
My solution to fix this is to check the last updated time of buffering
_folders_ within `/var/log/fluentd_buffers`. These _do_ get updated when
buffers are created, and do not get deleted as buffers are emptied,
making them the perfect candidate for our use.
Here's an example with the `-d` flag for directories:
```
root@fluentd-eks-playground-jdc8m:/# while true; do date ; find /var/log/fluentd-buffers -type d -newer /tmp/marker-stuck -print -quit ; sleep 1 ; done
Fri Feb 22 10:57:51 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:52 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:53 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:54 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:55 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:56 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:57 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:58 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:57:59 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:58:00 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:58:01 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:58:02 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
Fri Feb 22 10:58:03 UTC 2019
/var/log/fluentd-buffers/kubernetes.system.buffer
```
And example of the directory being updated as new buffers come in:
```
root@fluentd-eks-playground-jdc8m:/# ls -lah /var/log/fluentd-buffers/kubernetes.system.buffer
total 0
drwxr-xr-x 2 root root 6 Feb 22 11:17 .
drwxr-xr-x 3 root root 38 Feb 22 11:14 ..
root@fluentd-eks-playground-jdc8m:/# ls -lah /var/log/fluentd-buffers/kubernetes.system.buffer
total 16K
drwxr-xr-x 2 root root 224 Feb 22 11:18 .
drwxr-xr-x 3 root root 38 Feb 22 11:14 ..
-rw-r--r-- 1 root root 1.8K Feb 22 11:18 buffer.b58279be6e21e8b29fc333a7d50096ed0.log
-rw-r--r-- 1 root root 215 Feb 22 11:18 buffer.b58279be6e21e8b29fc333a7d50096ed0.log.meta
-rw-r--r-- 1 root root 429 Feb 22 11:18 buffer.b58279be6f09bdfe047a96486a525ece2.log
-rw-r--r-- 1 root root 195 Feb 22 11:18 buffer.b58279be6f09bdfe047a96486a525ece2.log.meta
root@fluentd-eks-playground-jdc8m:/# ls -lah /var/log/fluentd-buffers/kubernetes.system.buffer
total 0
drwxr-xr-x 2 root root 6 Feb 22 11:18 .
drwxr-xr-x 3 root root 38 Feb 22 11:14 ..
```
Similar to `--no-negcache` on dnsmasq, this prevents issues which poll DNS for orchestration such as operators with StatefulSets. It can also be very confusing for users when negative caching results in a change they just made seeming to be broken until the cache expires. This assumes that 5 seconds is reasonable and will still catch repeated AAAA negative responses. We could also set the denial cache size to zero which should effectively fully disable it like dnsmasq in kube-dns but testing shows this approach seems to work well in our (albeit small) test clusters.
These DaemonSets supports only Linux today, so this change updates the
specs to reflect this limitation. The labels have recently been promoted
to GA. Using the beta labels for now until node-master version skew
problem no longer exists.
The node_exporter CPU use is bursty, as it needs a bit of CPU at scrape time. Don't set a CPU limit to avoid collection stalls.
Set the request to 100m to more closely match the typical max core needs.
setting CPU_LIMITS to '1' fixes the following log appearing every 60 seconds:
Running: kubectl set resources -n kube-system ds fluentd-gcp-v3.1.0 -c fluentd-gcp --requests=cpu=100m,memory=200Mi --limits=cpu=1000m,memory=500Mi
error: info: {extensions v1beta1 daemonsets} "fluentd-gcp-v3.1.0" was not changed
this PR does not change scaler's behaviour, pods are scaled correctly despite error in the logs
