3 ms·

Cool thanks will have a look. Id be more interesting on where kubernetes (redundancy) fits into traditional redundancy ? Examples: 1. Do i still need gluster

Cool thanks will have a look. Id be more interesting on where kubernetes (redundancy) fits into traditional redundancy ? Examples: 1. Do i still need gluster/ceph or do i use the longhorn thing ? 2. Db replication? Do i use the usual solutions of master-slaves and clusters or does multiple k-nodes take that over ? 3. Webserver LB with failover ? Do i use LB from hosting vendor, haproxy or does kubernetes have its own thing ?? From what I can tell as a kubernetes-noob the value is: 1) Reproducibility 2) Reliability via redundancy 3) AutoScaling. All of the abkve has to some degree a previous/current solution , so which do i give up/replace with kubernetes-tool ? Sorry yes Im a k-noob

timhaak 4y ago
1 It depends. For larger scale gluster or ceph. But quiet a bit more work. Longhorn you can get up and working quickly. Though if you are on a cloud provider just use their storage system. 2 K8 doesn't magically solve replication unfortunately. Though there are helm charts that will automatically set up a replicated setup for you. I still need to solve backups. Once again if you are on a cloud provide. Just use their Db offering. 3 K8 doesn't have a default out of the box. The repo shows you how to setup traefik to handle this. On cloud providers they have normally integrated it with their lb already. For me the large advantages are reproducibility and no vendor lock in. Also give redundancy and quiet a bit of automation once set up. Auto scaling is always tricky. Lastly if you have the skills it can be far cheaper to run your own in metal. If you don't the the time most likely would be better spent actually coding. Depending where you are in the world and the relevant pay scales.
vbezhenar 4y ago
I'll try to answer, but keep in mind that I'm newbie myself. > 1. Do i still need gluster/ceph or do i use the longhorn thing ? Kubernetes does not care about storage implementation. It contains some abstract ways to request a storage (PersistentVolumeClaim). And then some particular Kubernetes installation will fulfil this request with PersistentVolume. So basically it comes down to your Kubernetes provider. It should have some instructions about volume classes that you can use. If you're installing your Kubernetes on bare metal, you need to think about this aspect yourself, of course. Both ceph and glusterfs are popular options and there're good Kubernetes drivers for it. You can also just use local storage, like docker does, but of course it'll not survive server outage, so it limits your availability. I'm installing my cluster on OpenStack. There's Cinder CSI plugin for Kubernetes, so it provides me storage when I ask for it. My provider uses SAN for one type of volume and Ceph for other type of volume. I think that simplest solution is some kind of NFS server. Kubernetes can consume it as well. > 2. Db replication? Do i use the usual solutions of master-slaves and clusters or does multiple k-nodes take that over ? Basically Kubernetes does not care about your particular configuration. It runs containers and provides those containers with storage, DNS, network, etc. So it's up to you. You can configure database replication with your own tools and scripts if you like. I have had good experience with CloudNativePG. It's so-called Kubernetes Operator. Basically it's a thing that configures database clusters for (postgres) given abstract definition. It can configure master-slaves cluster and it allows for easy backup configuration to S3 storage. There're other operators as well, for almost any popular database. So probably it's better to use those, unless you're very good at database operations. And in the true clouds it might be a good idea to use managed database and not to think about it at all. > 3. Webserver LB with failover ? Do i use LB from hosting vendor, haproxy or does kubernetes have its own thing ?? You need some kind of external load balancer to deliver packets in a high-available way. My hoster provides that as part of its OpenStack package. I guess that every cloud hoster provides it. If you're using bare metal, you need some kind of haproxy and keepalived setup (or some kind of hardware load balancer, I have no idea). This external load balancer have to deliver TCP packets to your worker nodes. Like 1.2.3.4:80 -> 10.1.1.1:30080, 10.11.1.2:30080, 10.11.1.3:30080. And once Kubernetes receives those packets in a high-available way, it routes them as needed. Usually you have ingress controller which provides HTTPS, and then uses HTTP host and path to route request to some pod which serves it in the end. Once request reached Kubernetes, it'll make sure to route it the right way. If your pods deployed in 2+ replicas, it'll be high-available. If your pod deployed in 1 replica and server is died, Kubernetes eventually will reschedule that pod to another server, but there'll be service interruption for a few minutes. So everything high-available should be deployed in 2+ replicas. > From what I can tell as a kubernetes-noob the value is: 1) Reproducibility 2) Reliability via redundancy 3) AutoScaling. Here's my take on Kubernetes value. First is it introduces a language connecting developers and operations. It is important. You don't need developers to hand-waive which ports they need to expose, which services they need to consume and which HTTP routes they need to receive for that particular server. They've got language to express how their service should be used. Second is it provides high-available cluster. And in my testing it's quite stable. My only issue is that it takes too much time to reschedule pods from dead server. I expected it to take few seconds, but it took few minutes. I don't yet know why is that. Third and one I didn't really expect when started to learn it: it provides high-quality solutions for some hard problems. I mentioned database clustering and database backup. I can deploy database with single 50-lines YAML mostly copy&pasted from the example. It'll start master and slave pods and it'll provide continuous backup to S3 using barman. I don't have skills to configure that kind of setup and I expect that I'd need a week at least to come to that setup. Another problem is letsencrypt. Well, it's not that hard, but I've spent many hours debugging some convoluted nginx/caddy/whatever net of docker containers trying to figure out what letsencrypt does not work there. With Kubernetes cert-manager it's just works. All configuration is centralized, all services write ingress and they've got their TLS certificate automagically, whether it's HTTP-01 solver, DNS-01 solver, it's just abstracted away. It has some steep learning curve, that's for sure. Even more so, if you want to deploy it yourself rather than using managed one. I suggest to use managed one if you can. I have some circumstances which prevent me to using managed Kubernetes, but I plan to migrate to managed as soon as I can. It's not that hard, but it takes time and managed Kubernetes is cheap enough. If you can't use managed Kubernetes, try to find a provider with OpenStack API. It'll help with load balancers and storage provisioning. Autoscaling - that part I didn't solve yet. It's not easy if you're not using managed Kubernetes. But if you're using managed Kubernetes, it should be as easy as ticking a checkbox somewhere. > All of the above has to some degree a previous/current solution , so which do i give up/replace with kubernetes-tool ? Well, right now we're using three dedicated servers with docker-composes scattered all over, made with ad-hoc scripts and whatnot, partially working backups, no observability. Chaos. Kubernetes for us looks like a very promising way to throw away that chaos and rebuild operations correctly.
- thunky 4y ago
  > I think that simplest solution is some kind of NFS server. Kubernetes can consume it as well. Isn't this asking for trouble, running MySQL on top of NFS? Also a k-nube.
  - vbezhenar 4y ago
    I don't have experience with that kind of setup, so can't really comment. Kubernetes is not magic, if you're running MySQL in NFS volume, it's just mysql process running in separate cgroup with mounted NFS volume. If you think that's a bad idea, don't do that. When it comes to processes and mounts, Kubernetes just arranges things and rest is done my ordinary Linux APIs. I tried to run Postgres with ceph volume and it works for me. You can configure database to just use local storage. Of course if your server will die, your database will die, so probably you will want some kind of cluster setup or be ready to restore database from backup if that happens. Or you can start with simple setup and migrate to something more complex in the future.
    - thunky 4y ago
      I hadn't heard about k8s "local storage". IIUC this fixes each pod to a single node, so the potential benefits of k8s can't be fully realized. It seems that clustering could still be achieved by running muliple instances across multiple nodes, each with local storage. I might still prefer this to having to manage NFS and any difficulties that might come from using MySQL with remote disks. It just seems like k8s is coercing people to do things that they normally wouldn't do: MySQL on NFS, Postgres on Ceph, etc. In this case, OP just wants to "kubernetify my run of the mil lamp app" and is now being pointed towards NFS, Ceph, etc. I can't help but wonder if the juice is worth the sqeeze. Especially for services like databases that for most people probably don't really require dynamic provisioning or orchestration.

timhaak 4y ago

1 It depends. For larger scale gluster or ceph. But quiet a bit more work. Longhorn you can get up and working quickly. Though if you are on a cloud provider just use their storage system. 2 K8 doesn't magically solve replication unfortunately. Though there are helm charts that will automatically set up a replicated setup for you. I still need to solve backups. Once again if you are on a cloud provide. Just use their Db offering. 3 K8 doesn't have a default out of the box. The repo shows you how to setup traefik to handle this. On cloud providers they have normally integrated it with their lb already. For me the large advantages are reproducibility and no vendor lock in. Also give redundancy and quiet a bit of automation once set up. Auto scaling is always tricky. Lastly if you have the skills it can be far cheaper to run your own in metal. If you don't the the time most likely would be better spent actually coding. Depending where you are in the world and the relevant pay scales.

vbezhenar 4y ago

I'll try to answer, but keep in mind that I'm newbie myself. > 1. Do i still need gluster/ceph or do i use the longhorn thing ? Kubernetes does not care about storage implementation. It contains some abstract ways to request a storage (PersistentVolumeClaim). And then some particular Kubernetes installation will fulfil this request with PersistentVolume. So basically it comes down to your Kubernetes provider. It should have some instructions about volume classes that you can use. If you're installing your Kubernetes on bare metal, you need to think about this aspect yourself, of course. Both ceph and glusterfs are popular options and there're good Kubernetes drivers for it. You can also just use local storage, like docker does, but of course it'll not survive server outage, so it limits your availability. I'm installing my cluster on OpenStack. There's Cinder CSI plugin for Kubernetes, so it provides me storage when I ask for it. My provider uses SAN for one type of volume and Ceph for other type of volume. I think that simplest solution is some kind of NFS server. Kubernetes can consume it as well. > 2. Db replication? Do i use the usual solutions of master-slaves and clusters or does multiple k-nodes take that over ? Basically Kubernetes does not care about your particular configuration. It runs containers and provides those containers with storage, DNS, network, etc. So it's up to you. You can configure database replication with your own tools and scripts if you like. I have had good experience with CloudNativePG. It's so-called Kubernetes Operator. Basically it's a thing that configures database clusters for (postgres) given abstract definition. It can configure master-slaves cluster and it allows for easy backup configuration to S3 storage. There're other operators as well, for almost any popular database. So probably it's better to use those, unless you're very good at database operations. And in the true clouds it might be a good idea to use managed database and not to think about it at all. > 3. Webserver LB with failover ? Do i use LB from hosting vendor, haproxy or does kubernetes have its own thing ?? You need some kind of external load balancer to deliver packets in a high-available way. My hoster provides that as part of its OpenStack package. I guess that every cloud hoster provides it. If you're using bare metal, you need some kind of haproxy and keepalived setup (or some kind of hardware load balancer, I have no idea). This external load balancer have to deliver TCP packets to your worker nodes. Like 1.2.3.4:80 -> 10.1.1.1:30080, 10.11.1.2:30080, 10.11.1.3:30080. And once Kubernetes receives those packets in a high-available way, it routes them as needed. Usually you have ingress controller which provides HTTPS, and then uses HTTP host and path to route request to some pod which serves it in the end. Once request reached Kubernetes, it'll make sure to route it the right way. If your pods deployed in 2+ replicas, it'll be high-available. If your pod deployed in 1 replica and server is died, Kubernetes eventually will reschedule that pod to another server, but there'll be service interruption for a few minutes. So everything high-available should be deployed in 2+ replicas. > From what I can tell as a kubernetes-noob the value is: 1) Reproducibility 2) Reliability via redundancy 3) AutoScaling. Here's my take on Kubernetes value. First is it introduces a language connecting developers and operations. It is important. You don't need developers to hand-waive which ports they need to expose, which services they need to consume and which HTTP routes they need to receive for that particular server. They've got language to express how their service should be used. Second is it provides high-available cluster. And in my testing it's quite stable. My only issue is that it takes too much time to reschedule pods from dead server. I expected it to take few seconds, but it took few minutes. I don't yet know why is that. Third and one I didn't really expect when started to learn it: it provides high-quality solutions for some hard problems. I mentioned database clustering and database backup. I can deploy database with single 50-lines YAML mostly copy&pasted from the example. It'll start master and slave pods and it'll provide continuous backup to S3 using barman. I don't have skills to configure that kind of setup and I expect that I'd need a week at least to come to that setup. Another problem is letsencrypt. Well, it's not that hard, but I've spent many hours debugging some convoluted nginx/caddy/whatever net of docker containers trying to figure out what letsencrypt does not work there. With Kubernetes cert-manager it's just works. All configuration is centralized, all services write ingress and they've got their TLS certificate automagically, whether it's HTTP-01 solver, DNS-01 solver, it's just abstracted away. It has some steep learning curve, that's for sure. Even more so, if you want to deploy it yourself rather than using managed one. I suggest to use managed one if you can. I have some circumstances which prevent me to using managed Kubernetes, but I plan to migrate to managed as soon as I can. It's not that hard, but it takes time and managed Kubernetes is cheap enough. If you can't use managed Kubernetes, try to find a provider with OpenStack API. It'll help with load balancers and storage provisioning. Autoscaling - that part I didn't solve yet. It's not easy if you're not using managed Kubernetes. But if you're using managed Kubernetes, it should be as easy as ticking a checkbox somewhere. > All of the above has to some degree a previous/current solution , so which do i give up/replace with kubernetes-tool ? Well, right now we're using three dedicated servers with docker-composes scattered all over, made with ad-hoc scripts and whatnot, partially working backups, no observability. Chaos. Kubernetes for us looks like a very promising way to throw away that chaos and rebuild operations correctly.

thunky 4y ago
> I think that simplest solution is some kind of NFS server. Kubernetes can consume it as well. Isn't this asking for trouble, running MySQL on top of NFS? Also a k-nube.
- vbezhenar 4y ago
  I don't have experience with that kind of setup, so can't really comment. Kubernetes is not magic, if you're running MySQL in NFS volume, it's just mysql process running in separate cgroup with mounted NFS volume. If you think that's a bad idea, don't do that. When it comes to processes and mounts, Kubernetes just arranges things and rest is done my ordinary Linux APIs. I tried to run Postgres with ceph volume and it works for me. You can configure database to just use local storage. Of course if your server will die, your database will die, so probably you will want some kind of cluster setup or be ready to restore database from backup if that happens. Or you can start with simple setup and migrate to something more complex in the future.
  - thunky 4y ago
    I hadn't heard about k8s "local storage". IIUC this fixes each pod to a single node, so the potential benefits of k8s can't be fully realized. It seems that clustering could still be achieved by running muliple instances across multiple nodes, each with local storage. I might still prefer this to having to manage NFS and any difficulties that might come from using MySQL with remote disks. It just seems like k8s is coercing people to do things that they normally wouldn't do: MySQL on NFS, Postgres on Ceph, etc. In this case, OP just wants to "kubernetify my run of the mil lamp app" and is now being pointed towards NFS, Ceph, etc. I can't help but wonder if the juice is worth the sqeeze. Especially for services like databases that for most people probably don't really require dynamic provisioning or orchestration.