AWS has soft support limit for 40 attached EBS devices. Assuming there is just
one root device, use the rest for persistent volumes.
The devices will have name /dev/xvdba - /dev/xvdcm, leaving /dev/sda - /dev/sdz
to the system.
Also, add better error handling and propagate error
"Too many EBS volumes attached to node XYZ" to a pod.
We have previously tried building a full cloudprovider in e2e for AWS;
this wasn't the best idea, because e2e runs on a different machine than
normal operations, and often doesn't even run in AWS. In turn, this
meant that the cloudprovider had to do extra work and have extra code,
which we would like to get rid of. Indeed, I got rid of some code which
tolerated not running in AWS, and this broke e2e.
There are known issues with the attached-volume state cache that just aren't
possible to fix with the current interface.
Replace it with a map of the active attach jobs (that was the original
requirement, to avoid a nasty race condition).
This costs us an extra DescribeInstance call on attach/detach, but that
seems worth it if it ends this class of bugs.
Fix#15073
Either ELB is slow to delete (in which case the bumped timeout will
help), or the security groups are otherwise blocked (in which case
logging them will help us track this down).
Fix#17626
We know the ELB call will fail, so we error out early rather than
hitting the API. Preserves rate limit quota, and also allows us to give
a more self-evident message.
Fix#21993
Now that we can't build an awsInstance from metadata, because of the
PrivateDnsName issue, we might as well simplify the arguments.
Create a 'placeholder' method though - newAWSInstanceFromMetadata - that
documents the desire to use metadata, shows how we would get it, but
links to the bug which explains why we can't use it.
Had to move other things around too to avoid a weird api ->
cloudprovider dependency.
Also adding fixes per code reviews.
(This is a squash of the previously approved commits)
This has two main advantages:
* The use of the mock package to verify API calls against the aws SDK
* Nicer error messages for asserts without having to use if statements
We return an error if the user specifies a non 0.0.0.0/0 load balancer
source restriction on OpenStack, where we can't enforce the restriction
(currently).
This refactors #21431 to pull a lot of the code into cloudprovider so it
can be reused by AWS.
It also changes the name of the annotation to be non-GCE specific:
service.beta.kubernetes.io/load-balancer-source-ranges
Fix#21651
for Instance.List and Routes.List which we will definitely have
more than 500 of when supporting 1000 nodes.
Add TODOs for other GCE List API calls to do similar fixes.
Add more logging to GCE's routecontroller.go when creating or deleting routes.
Fix the AWS subnet lookup that checks if a subnet is public, which was
missing a few cases:
- Subnets without explicit routing tables, which use the main VPC
routing table.
- Routing tables not tagged with KubernetesCluster. The filter for this
is now removed.
Like everything else AWS, we differentiate between k8s-owned security
groups and k8s-not-owned security groups using tags.
When we are setting up the ingress rule for ELBs, pick the security
group that is tagged over any others.
We continue to tolerate a single security group being untagged, but
having multiple security groups without tagging is now an error, as it
leads to undefined behaviour.
We also log at startup if the cluster tag is not defined.
Fix#21986
Follow up from #20731. I have no way of testing this.
There's an additional group of functions (Get|Delete|Reserve)GlobalStaticIP that can create an IP without the
service description, but those are not called anywhere in the Kubernetes codebase and are probably for the
Ingress project. I'm leaving those alone for now.
Add aws cloud config:
[global]
disableSecurityGroupIngress = true
The aws provider creates an inbound rule per load balancer on the node
security group. However, this can quickly run into the AWS security
group rule limit of 50.
This disables the automatic ingress creation. It requires that the user
has setup a rule that allows inbound traffic on kubelet ports from the
local VPC subnet (so load balancers can access it). E.g. `10.82.0.0/16
30000-32000`.
Limits: http://docs.aws.amazon.com/AmazonVPC/latest/UserGuide/VPC_Appendix_Limits.html#vpc-limits-security-groups
Authors: @jsravn, @balooo
When finding instance by node name in AWS, only retrieve running
instances. Otherwise terminated, old nodes can show up with the same
tag when rebuilding nodes in the cluster.
Another improvement made is to filter instances by the node names
provided, rather than selecting all instances and filtering in code.
Authors: @jsravn, @chbatey, @balooo
This applies a cross-request time delay when we observe
RequestLimitExceeded errors, unlike the default library behaviour which
only applies a *per-request* backoff.
Issue #12121
In the AWS API (generally) we tag things we create, and then we filter
to find them. However, creation & tagging are typically two separate
calls. So there is a chance that we will create an object, but fail to
tag it.
We fix this (done here in the case of security groups, but we can do
this more generally) by retrieving the resource without a tag filter.
If the retrieved resource has the correct tags, great. If it has the
tags for another cluster, that's a problem, and we raise an error. If
it has no tags at all, we add the tags.
This only works where the resource is uniquely named (or we can
otherwise retrieve it uniquely). For security groups, the SG name comes
from the service UUID, so that's unique.
Fixes#11324
Volume names have now format <cluster-name>-dynamic-<pv-name>.
pv-name is guaranteed to be unique in Kubernetes cluster, adding
<cluster-name> ensures we don't conflict with any running cluster
in the cloud project (kube-controller-manager --cluster-name=XXX).
'kubernetes' is the default cluster name.
AWS doesn't support type=LoadBalancer with UDP services. For now, we
simply skip over the test with type=LoadBalancer on AWS for the UDP
service.
Fix#20911
This commit allows the AWS cloud provider plugin to work on EC2 instances
that do not have a public IP. The EC2 metadata service returns a 404 for the
'public-ipv4' endpoint for private instances, and the plugin was bubbling this
up as a fatal error.
We are (sadly) using a copy-and-paste of the GCE PD code for AWS EBS.
This code hasn't been updated in a while, and it seems that the GCE code
has some code to make volume mounting more robust that we should copy.
GCE disks don't have tags, we must encode the tags into Description field.
It's encoded as JSON, which is both human and machine readable:
description: '{"kubernetes.io/created-for/pv/name":"pv-gce-oxwts","kubernetes.io/created-for/pvc/name":"myclaim","kubernetes.io/created-for/pvc/namespace":"default"}'
The ip permission method now checks for containment, not equality, so
order of parameters matter. This change fixes
`removeSecurityGroupIngress` to pass in the removal permission first to
compare against the existing permission.
Change isEqualIPPermission to consider the entire list of security group
ids on when checking if a security group id has already been added.
This is used for example when adding and removing ingress rules to the
cluster nodes from an elastic load balancer. Without this, once there
are multiple load balancers, the method as it stands incorrectly returns
false even if the security group id is in the list of group ids. This
causes a few problems: dangling security groups which fill up an
account's limit since they don't get removed, and inability to recreate
load balancers in certain situations (receiving an
InvalidPermission.Duplicate from AWS when adding the same security
group).
According to AWS, the ELB healthy threshold is "Number of consecutive health check successes before declaring an EC2 instance healthy." It has an unusual interaction with Kubernetes, since all nodes will enter either an unhealthy state or a healthy state together depending on the service's healthiness as a whole.
We have observed that if our service goes down for the unhealthy threshold (which is 2 checks at 30 second intervals = 60 seconds), then the ELB will stop serving traffic to all nodes in the cluster, and will wait for the healthy threshold (currently 10 * 30 = 300 seconds) AFTER the service is restored to add back the cluster nodes, meaning it remains unreachable for an extra 300 seconds.
With the new settings, the ELB will continue to timeout dead nodes after 60 seconds, but will restore healthy nodes after 20 seconds. The minimum value for healthyThreshold is 2, and the minimum value for interval is 5 seconds. I went for 10 seconds instead of the minimum sort of arbitrarily because I was not sure how much this value may affect the scalability of clusters in EC2, as it does put some extra load on the kube-proxy.
We don't cope well if a PD is in multiple zones, but this is actually
fairly easy to detect. This is probably justified purely on the basis
that we never want to delete the wrong volume (DeleteDisk), but also
because this means that we now warn on creation if a disk is in multiple
zones (with the labeling admission controller).
This also means that with the scheduling predicate in place, that many
of our volume problems "go away" in practice: you still can't create or
delete a volume when it is ambiguous, but thereafter the volume will be
labeled with the zone, that will match it only to nodes with the same
zone, and then we query for the volume in that zone when we
attach/detach it.
This removes a panic I mistakenly introduced when an instance is not
found, and also restores the exact prior behaviour for
getInstanceByName, where it returns cloudprovider.InstanceNotFound when
the instance is not found.
We adapt the existing code to work across all zones in a region.
We require a feature-flag to enable Ubernetes-Lite
Reasons:
* There are some behavioural changes if users create volumes with
the same name in two zones.
* We don't want to make one API call per zone if we're not running
Ubernetes-Lite.
* Ubernetes-Lite is still experimental.
There isn't a parallel flag implemented for AWS, because at the moment
there would be no behaviour changes from this.
findInstancesByNodeNames was a simple loop around
findInstanceByNodeName, which made an EC2 API call for each call.
We've had trouble with this sort of behaviour hitting EC2 rate limits on
bigger clusters (e.g. #11979).
Instead, change this method to fetch _all_ the tagged EC2 instances, and
then loop through the local results. This is one API call (modulo
paging).
We are currently only using findInstancesByNodeNames for the load
balancer, where we attach every node, so we were fetching all but one of
the instances anyway.
Issue #11979
If a route already exists but is invalid (e.g. from a crash), we
automatically delete it before trying to create a route that would
otherwise conflict.
For AWS EBS, a volume can only be attached to a node in the same AZ.
The scheduler must therefore detect if a volume is being attached to a
pod, and ensure that the pod is scheduled on a node in the same AZ as
the volume.
So that the scheduler need not query the cloud provider every time, and
to support decoupled operation (e.g. bare metal) we tag the volume with
our placement labels. This is done automatically by means of an
admission controller on AWS when a PersistentVolume is created backed by
an EBS volume.
Support for tagging GCE PVs will follow.
Pods that specify a volume directly (i.e. without using a
PersistentVolumeClaim) will not currently be scheduled correctly (i.e.
they will be scheduled without zone-awareness).
This synchronizes Cinder with AWS EBS code, where we already tag volumes with
claim.Namespace and claim.Name (and pv.Name, as suggested in separate PR).
General purpose SSD ('gp2') volume type is just slighly more expensive than
Magnetic ('standard' / default in AWS), while the performance gain is pretty
significant.
So far, the volumes were created only during testing, where the extra cost
won't make any difference. In future, we plan to introduce QoS classes, where
users could choose SSD/Magnetic depending on their use cases.
'gp2' is just the default volume type for (hopefuly) short period before these
QoS classes are implemented.
From some reason, MiBs were used for public functions and AWS cloud provider
recalculated them to GiB. Let's expose what AWS really supports and don't hide
real allocation units.
k8s-merge-robot