HA jobmanagers redirect to ip address of leader instead of hostname

[Jeroen Steggink | knowsy]

Hi,

I'm having some troubles with Flink jobmanagers in a HA setup within
OpenShift.

I have three jobmanagers, a Zookeeper cluster and a loadbalancer
(Openshift/Kubernetes Route) for the web ui / rest server on the
jobmanagers. Everything works fine, as long as the loadbalancer connects
to the leader. However, when the leader changes and the loadbalancer
connects to a non-leader, the jobmanager redirects to a leader using the
ip address of the host. Since the routing in our network is done using
hostnames, it doesn't know how to find the node using the ip address and
results in a timeout.

So I have a few questions:
1. Why is Flink using the ip addresses instead of the hostname which are
configured in the config? Other times it does use the hostname, like the
info send to Zookeeper.
2. Is there another way of coping with connections to non-leaders
instead of redirects? Maybe proxying through a non-leader to the leader?

Cheers,
Jeroen

[Till Rohrmann]

Hi Jeroen,

this sounds like a bug in Flink that we return sometimes IP addresses instead of hostnames. Could you tell me which Flink version you are using? In the current version, the redirect address and the address retrieved from ZooKeeper should actually be the same.

In the future, we plan to remove the redirect message and simply forward the request to the current leader. This should hopefully avoid these kind of problems.

Cheers,

Till

On Fri, Oct 26, 2018 at 1:40 PM Jeroen Steggink | knowsy <[hidden email]> wrote:

Hi,

I'm having some troubles with Flink jobmanagers in a HA setup within
OpenShift.

I have three jobmanagers, a Zookeeper cluster and a loadbalancer
(Openshift/Kubernetes Route) for the web ui / rest server on the
jobmanagers. Everything works fine, as long as the loadbalancer connects
to the leader. However, when the leader changes and the loadbalancer
connects to a non-leader, the jobmanager redirects to a leader using the
ip address of the host. Since the routing in our network is done using
hostnames, it doesn't know how to find the node using the ip address and
results in a timeout.

So I have a few questions:
1. Why is Flink using the ip addresses instead of the hostname which are
configured in the config? Other times it does use the hostname, like the
info send to Zookeeper.
2. Is there another way of coping with connections to non-leaders
instead of redirects? Maybe proxying through a non-leader to the leader?

Cheers,
Jeroen

[Jeroen Steggink | knowsy]

Hi Till,

Thanks for your reply. We are running version 1.5.4. We can't upgrade to 1.6.x because we are using Apache Beam which doesn't support 1.6.x yet.

I have also made a Jira issue about this: https://issues.apache.org/jira/projects/FLINK/issues/FLINK-10748

Best regards,
Jeroen Steggink

On 08-Nov-18 11:40, Jeroen Steggink | knowsy wrote:

Hi Till,

Thanks for your reply. We are running version 1.5.4. We can't upgrade to 1.6.x because we are using Apache Beam which doesn't support 1.6.x yet.

I have also made a Jira issue about this: https://issues.apache.org/jira/projects/FLINK/issues/FLINK-10748

On 07-Nov-18 16:06, Till Rohrmann wrote:

Hi Jeroen,

this sounds like a bug in Flink that we return sometimes IP addresses instead of hostnames. Could you tell me which Flink version you are using? In the current version, the redirect address and the address retrieved from ZooKeeper should actually be the same.

In the future, we plan to remove the redirect message and simply forward the request to the current leader. This should hopefully avoid these kind of problems.

Cheers,

Till

On Fri, Oct 26, 2018 at 1:40 PM Jeroen Steggink | knowsy <[hidden email]> wrote:

Hi,

I'm having some troubles with Flink jobmanagers in a HA setup within
OpenShift.

I have three jobmanagers, a Zookeeper cluster and a loadbalancer
(Openshift/Kubernetes Route) for the web ui / rest server on the
jobmanagers. Everything works fine, as long as the loadbalancer connects
to the leader. However, when the leader changes and the loadbalancer
connects to a non-leader, the jobmanager redirects to a leader using the
ip address of the host. Since the routing in our network is done using
hostnames, it doesn't know how to find the node using the ip address and
results in a timeout.

So I have a few questions:
1. Why is Flink using the ip addresses instead of the hostname which are
configured in the config? Other times it does use the hostname, like the
info send to Zookeeper.
2. Is there another way of coping with connections to non-leaders
instead of redirects? Maybe proxying through a non-leader to the leader?

Cheers,
Jeroen

[Till Rohrmann]

Thanks for opening this issue Jeroen. I think we need to look into this. Could you maybe share the Flink logs with the community? Ideally you attach them to the JIRA issue.

Cheers,

Till

On Thu, Nov 8, 2018 at 10:51 PM Jeroen Steggink | knowsy <[hidden email]> wrote:

Hi Till,

Thanks for your reply. We are running version 1.5.4. We can't upgrade to 1.6.x because we are using Apache Beam which doesn't support 1.6.x yet.

I have also made a Jira issue about this: https://issues.apache.org/jira/projects/FLINK/issues/FLINK-10748

Best regards,
Jeroen Steggink

On 08-Nov-18 11:40, Jeroen Steggink | knowsy wrote:

Hi Till,

Thanks for your reply. We are running version 1.5.4. We can't upgrade to 1.6.x because we are using Apache Beam which doesn't support 1.6.x yet.

I have also made a Jira issue about this: https://issues.apache.org/jira/projects/FLINK/issues/FLINK-10748

On 07-Nov-18 16:06, Till Rohrmann wrote:

Hi Jeroen,

this sounds like a bug in Flink that we return sometimes IP addresses instead of hostnames. Could you tell me which Flink version you are using? In the current version, the redirect address and the address retrieved from ZooKeeper should actually be the same.

In the future, we plan to remove the redirect message and simply forward the request to the current leader. This should hopefully avoid these kind of problems.

Cheers,

Till

On Fri, Oct 26, 2018 at 1:40 PM Jeroen Steggink | knowsy <[hidden email]> wrote:

Hi,

I'm having some troubles with Flink jobmanagers in a HA setup within
OpenShift.

I have three jobmanagers, a Zookeeper cluster and a loadbalancer
(Openshift/Kubernetes Route) for the web ui / rest server on the
jobmanagers. Everything works fine, as long as the loadbalancer connects
to the leader. However, when the leader changes and the loadbalancer
connects to a non-leader, the jobmanager redirects to a leader using the
ip address of the host. Since the routing in our network is done using
hostnames, it doesn't know how to find the node using the ip address and
results in a timeout.

So I have a few questions:
1. Why is Flink using the ip addresses instead of the hostname which are
configured in the config? Other times it does use the hostname, like the
info send to Zookeeper.
2. Is there another way of coping with connections to non-leaders
instead of redirects? Maybe proxying through a non-leader to the leader?

Cheers,
Jeroen