Re: HAproxy and detect split-brain (network failures)
Hi, This is not HAProxy's role, this is the tool you use to ensure high availability to do that. I could see a way where HAProxy can report one interface failing, maybe this could help you to detect if you're in a split brain situation. cheers On Thu, Nov 29, 2012 at 11:51 AM, Hermes Flying flyingher...@yahoo.com wrote: Hi, I am looking into using HAProxy as our load balancer. I see that you are using a primary/backup approach. I was wondering how does HAProxy (if it does) address split-brain situation? Do you have a mechanism to detect and avoid it? Do you have some standard recommendation to all those using your solution? Thanks
Re: HAproxy and detect split-brain (network failures)
Hi, We use Keepalived http://www.keepalived.org/index.html to manage the Virtual IP address management between our two physical HAproxy servers. It maintains heartbeat between the servers, and in the event of failure passes ensures that the VIPs are migrated and the service is brought up. Also handles migration back after a restart of our primary, so that if available, that is the server that owns the IPs. We use Mercurial to manage the configuration files between the two servers to maintain consistency so that we are prepared for consistent fail overs. Robert On Nov 29, 2012, at 8 :02 AM, Baptiste bed...@gmail.com wrote: Hi, This is not HAProxy's role, this is the tool you use to ensure high availability to do that. I could see a way where HAProxy can report one interface failing, maybe this could help you to detect if you're in a split brain situation. cheers On Thu, Nov 29, 2012 at 11:51 AM, Hermes Flying flyingher...@yahoo.com wrote: Hi, I am looking into using HAProxy as our load balancer. I see that you are using a primary/backup approach. I was wondering how does HAProxy (if it does) address split-brain situation? Do you have a mechanism to detect and avoid it? Do you have some standard recommendation to all those using your solution? Thanks Robert Snyder Outreach Technology Services The Pennsylvania State University The 329 Building, Suite 306E University Park PA 16802 Phone: 814-865-0912 E-mail: rsny...@psu.edu
Re: HAproxy and detect split-brain (network failures)
Hi Robert, But with keep alive you can only detect that the 2 nodes can not contact each other (network failure). How do you know if the other node/process actually crashed so that the secondary can become the primary? From: Robert Snyder r...@psu.edu To: Baptiste bed...@gmail.com Cc: Hermes Flying flyingher...@yahoo.com; haproxy@formilux.org haproxy@formilux.org Sent: Thursday, November 29, 2012 3:13 PM Subject: Re: HAproxy and detect split-brain (network failures) Hi, We use Keepalived http://www.keepalived.org/index.html to manage the Virtual IP address management between our two physical HAproxy servers. It maintains heartbeat between the servers, and in the event of failure passes ensures that the VIPs are migrated and the service is brought up. Also handles migration back after a restart of our primary, so that if available, that is the server that owns the IPs. We use Mercurial to manage the configuration files between the two servers to maintain consistency so that we are prepared for consistent fail overs. Robert On Nov 29, 2012, at 8 :02 AM, Baptiste bed...@gmail.com wrote: Hi, This is not HAProxy's role, this is the tool you use to ensure high availability to do that. I could see a way where HAProxy can report one interface failing, maybe this could help you to detect if you're in a split brain situation. cheers On Thu, Nov 29, 2012 at 11:51 AM, Hermes Flying flyingher...@yahoo.com wrote: Hi, I am looking into using HAProxy as our load balancer. I see that you are using a primary/backup approach. I was wondering how does HAProxy (if it does) address split-brain situation? Do you have a mechanism to detect and avoid it? Do you have some standard recommendation to all those using your solution? Thanks Robert Snyder Outreach Technology Services The Pennsylvania State University The 329 Building, Suite 306E University Park PA 16802 Phone: 814-865-0912 E-mail: rsny...@psu.edu
Re: HAproxy and detect split-brain (network failures)
David, Exactly. Robert On Nov 29, 2012, at 1 :57 PM, David Coulson da...@davidcoulson.net wrote: You can do that, but haproxy doesn't have anything to do with the failover process, other than you run an instance of haproxy on one server, and another instance on your backup system. As I said, neither of the haproxy instances communicate anything, so all you need to do is move the IP clients are using from one server to the other in order to handle a failure. Moving the IP around is something keepalived, pacemaker, etc handles - Look at their documentation for specifics and challenges in a two-node config. HAProxy doesn't have a concent of primary and backup in terms of it's own instances. Each of them is stand alone. It's up to you, based on your network/IP config which one has traffic routed to it. David On 11/29/12 1:53 PM, Hermes Flying wrote: But if I install 2 HAProxy as load balancers, doesn't one act as the primary loadbalancer directing the load to the known servers while the secondary takes over load distribution as soon as the heartbeat fails? I remember reading this. Is this wrong? From: David Coulson da...@davidcoulson.net To: Hermes Flying flyingher...@yahoo.com Cc: Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org Sent: Thursday, November 29, 2012 8:39 PM Subject: Re: HAproxy and detect split-brain (network failures) You are mixing two totally different things together. 1) HAProxy will do periodic health checks of backend systems you are routing to. Depending if you configure something as 'backup' or 'not backup' will determine if/how traffic is routed to it. The backend systems do not 'take over'. Haproxy just routes traffic to systems based on your configuration. The backend systems don't know/care about the other backend nodes, unless your application requires it which is a different story and nothing to do with haproxy. HAproxy only cares about a single instance of itself - If you have more than one haproxy instance, they do NOT communicate anything between each other. 2) In terms of keepalived, pacemaker, etc, it makes no difference which you use with haproxy - all they do is manage the IP address(es) which haproxy is listening on, and perhaps restart haproxy if it dies. Their configuration and how you maintain quorum in a two-node configuration is a question for one of their mailing lists, or just read their documentation. I personally use pacemaker. On 11/29/12 1:35 PM, Hermes Flying wrote: Well I don't follow: You can have a pool of primary that it routes across, then backup systems that are only used when all primary systems are unavailable. When you are saying that the backup systems that are used when primary systems are unavailable, how do they decide to take over? How do they know that the other systems are unavailable? Are you saying that they depend on third party components like the ones you mentioned (Keepalived etc)? In this case, what is the most suitable tool to be used along with HAProxy? Is there a reference manual for this somewhere? From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 8:21 PM Subject: Re: HAproxy and detect split-brain (network failures) HAProxy only does primary and backup in terms of active backend systems - You can have a pool of primary that it routes across, then backup systems that are only used when all primary systems are unavailable. There is no concept of a cluster in terms of haproxy instances, although you can run more than one and manage them via something like pacemaker, keepalived or rgmanager. On 11/29/12 1:19 PM, Hermes Flying wrote: Hi, From a quick look into HAProxy, I see that it is a Primary/backup architecture. So isn't ensuring that both nodes don't become primary part of HAProxy's primary/backup protocol ? From: Baptiste mailto:bed...@gmail.com To: Hermes Flying mailto:flyingher...@yahoo.com Cc: mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 3:02 PM Subject: Re: HAproxy and detect split-brain (network failures) Hi, This is not HAProxy's role, this is the tool you use to ensure high availability to do that. I could see a way where HAProxy can report one interface failing, maybe this could help you to detect if you're in a split brain situation. cheers On Thu, Nov 29, 2012 at 11:51 AM, Hermes Flying flyingher...@yahoo.com wrote: Hi, I am looking into using HAProxy as our load balancer. I see that you are using a primary/backup approach. I was wondering how does HAProxy (if it does) address split-brain situation? Do you have a mechanism to detect and avoid it? Do you have some standard recommendation to all those
Re: HAproxy and detect split-brain (network failures)
Again, you are mixing everything up. HAProxy has it's own configuration - It defines what nodes your port 80 traffic (or whatever) is routed to. Haproxy does periodic health checks of these backend services to make sure they are available for requests. If you have multiple haproxy instances they will all independently do health checks and not share any of that information with each other. HAProxy will route traffic to all systems defined as a backend for a particular service based upon whatever criteria is in the haproxy config. You can run a two-node environment that is active/backup from a VIP perspective, but active/active from a haproxy service perspective - Each node would run Apache (or whatever your service is) and haproxy would distribute requests across both based on your haproxy config. But, at any point in time only one node would actually be routing requests through it's local instance of haproxy. I can't make it any simpler than that. Draw a diagram of what you are trying to do if it doesn't make sense. On 11/29/12 2:06 PM, Hermes Flying wrote: You are saying that one instance of HAProxy runs in each system and one instance is assigned the VIP that clients hit-on (out of scope for HAProxy). But this HAProxy distributes the requests according to the load, either on system-A or system-B for which you seem to refer to as backup system. In what way are you now refering to it as backup system? Because I am interested in distributing the load to all the nodes. *From:* David Coulson da...@davidcoulson.net *To:* Hermes Flying flyingher...@yahoo.com *Cc:* Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org *Sent:* Thursday, November 29, 2012 8:57 PM *Subject:* Re: HAproxy and detect split-brain (network failures) You can do that, but haproxy doesn't have anything to do with the failover process, other than you run an instance of haproxy on one server, and another instance on your backup system. As I said, neither of the haproxy instances communicate anything, so all you need to do is move the IP clients are using from one server to the other in order to handle a failure. Moving the IP around is something keepalived, pacemaker, etc handles - Look at their documentation for specifics and challenges in a two-node config. HAProxy doesn't have a concent of primary and backup in terms of it's own instances. Each of them is stand alone. It's up to you, based on your network/IP config which one has traffic routed to it. David On 11/29/12 1:53 PM, Hermes Flying wrote: But if I install 2 HAProxy as load balancers, doesn't one act as the primary loadbalancer directing the load to the known servers while the secondary takes over load distribution as soon as the heartbeat fails? I remember reading this. Is this wrong? *From:* David Coulson mailto:da...@davidcoulson.net *To:* Hermes Flying mailto:flyingher...@yahoo.com *Cc:* Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org *Sent:* Thursday, November 29, 2012 8:39 PM *Subject:* Re: HAproxy and detect split-brain (network failures) You are mixing two totally different things together. 1) HAProxy will do periodic health checks of backend systems you are routing to. Depending if you configure something as 'backup' or 'not backup' will determine if/how traffic is routed to it. The backend systems do not 'take over'. Haproxy just routes traffic to systems based on your configuration. The backend systems don't know/care about the other backend nodes, unless your application requires it which is a different story and nothing to do with haproxy. HAproxy only cares about a single instance of itself - If you have more than one haproxy instance, they do NOT communicate anything between each other. 2) In terms of keepalived, pacemaker, etc, it makes no difference which you use with haproxy - all they do is manage the IP address(es) which haproxy is listening on, and perhaps restart haproxy if it dies. Their configuration and how you maintain quorum in a two-node configuration is a question for one of their mailing lists, or just read their documentation. I personally use pacemaker. On 11/29/12 1:35 PM, Hermes Flying wrote: Well I don't follow: You can have a pool of primary that it routes across, then backup systems that are only used when all primary systems are unavailable. When you are saying that the backup systems that are used when primary systems are unavailable, how do they decide to take over? How do they know that the other systems are unavailable? Are you saying that they depend on third party components like the ones you mentioned (Keepalived etc)? In this case, what is the most suitable tool to be used along with HAProxy? Is there a reference manual for this somewhere? *From:* David Coulson mailto:da...@davidcoulson.net *To:* Hermes Flying mailto:flyingher...@yahoo.com *Cc:* Baptiste mailto:bed...@gmail.com; mailto:haproxy
Re: HAproxy and detect split-brain (network failures)
Something like the following: HAProxy1 Tomcat1 | +/\ | + |+---Tomcat2 + /+\ + + HAProxy2+++ HAProxy1 is in the same machine as Tomcat1 HAproxy2 is in the same machine as Tomcat2 HAProxy1 distributes the load among Tomcat1 and Tomcat2. I erroneously thought that HAProxy2 would take over when HAProxy1 crashed to distribute the load among Tomcat1/Tomcat2. So if both are independent what can I do? From: David Coulson da...@davidcoulson.net To: Hermes Flying flyingher...@yahoo.com Cc: Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org Sent: Thursday, November 29, 2012 9:12 PM Subject: Re: HAproxy and detect split-brain (network failures) Again, you are mixing everything up. HAProxy has it's own configuration - It defines what nodes your port 80 traffic (or whatever) is routed to. Haproxy does periodic health checks of these backend services to make sure they are available for requests. If you have multiple haproxy instances they will all independently do health checks and not share any of that information with each other. HAProxy will route traffic to all systems defined as a backend for a particular service based upon whatever criteria is in the haproxy config. You can run a two-node environment that is active/backup from a VIP perspective, but active/active from a haproxy service perspective - Each node would run Apache (or whatever your service is) and haproxy would distribute requests across both based on your haproxy config. But, at any point in time only one node would actually be routing requests through it's local instance of haproxy. I can't make it any simpler than that. Draw a diagram of what you are trying to do if it doesn't make sense. On 11/29/12 2:06 PM, Hermes Flying wrote: You are saying that one instance of HAProxy runs in each system and one instance is assigned the VIP that clients hit-on (out of scope for HAProxy). But this HAProxy distributes the requests according to the load, either on system-A or system-B for which you seem to refer to as backup system. In what way are you now refering to it as backup system? Because I am interested in distributing the load to all the nodes. From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 8:57 PM Subject: Re: HAproxy and detect split-brain (network failures) You can do that, but haproxy doesn't have anything to do with the failover process, other than you run an instance of haproxy on one server, and another instance on your backup system. As I said, neither of the haproxy instances communicate anything, so all you need to do is move the IP clients are using from one server to the other in order to handle a failure. Moving the IP around is something keepalived, pacemaker, etc handles - Look at their documentation for specifics and challenges in a two-node config. HAProxy doesn't have a concent of primary and backup in terms of it's own instances. Each of them is stand alone. It's up to you, based on your network/IP config which one has traffic routed to it. David On 11/29/12 1:53 PM, Hermes Flying wrote: But if I install 2 HAProxy as load balancers, doesn't one act as the primary loadbalancer directing the load to the known servers while the secondary takes over load distribution as soon as the heartbeat fails? I remember reading this. Is this wrong? From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 8:39 PM Subject: Re: HAproxy and detect split-brain (network failures) You are mixing two totally different things together. 1) HAProxy will do periodic health checks of backend systems you are routing to. Depending if you configure something as 'backup' or 'not backup' will determine if/how traffic is routed to it. The backend systems do not 'take over'. Haproxy just routes traffic to systems based on your configuration. The backend systems don't know/care about the other backend nodes, unless your application requires it which is a different story and nothing to do with haproxy. HAproxy only cares about
Re: HAproxy and detect split-brain (network failures)
Both haproxy instances have the same config, with the tomcat instances with the same weight, etc. Run something like keepalived or pacemaker to manage a VIP between the two boxes. That's it. Not sure about keepalived, but pacemaker can make sure haproxy is running, then either restart it or move the VIP if it is not running. David On 11/29/12 2:27 PM, Hermes Flying wrote: Something like the following: HAProxy1 Tomcat1 |+/\ | + |+---Tomcat2 +/+\ + + HAProxy2+++ HAProxy1 is in the same machine as Tomcat1 HAproxy2 is in the same machine as Tomcat2 HAProxy1 distributes the load among Tomcat1 and Tomcat2. I erroneously thought that HAProxy2 would take over when HAProxy1 crashed to distribute the load among Tomcat1/Tomcat2. So if both are independent what can I do? *From:* David Coulson da...@davidcoulson.net *To:* Hermes Flying flyingher...@yahoo.com *Cc:* Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org *Sent:* Thursday, November 29, 2012 9:12 PM *Subject:* Re: HAproxy and detect split-brain (network failures) Again, you are mixing everything up. HAProxy has it's own configuration - It defines what nodes your port 80 traffic (or whatever) is routed to. Haproxy does periodic health checks of these backend services to make sure they are available for requests. If you have multiple haproxy instances they will all independently do health checks and not share any of that information with each other. HAProxy will route traffic to all systems defined as a backend for a particular service based upon whatever criteria is in the haproxy config. You can run a two-node environment that is active/backup from a VIP perspective, but active/active from a haproxy service perspective - Each node would run Apache (or whatever your service is) and haproxy would distribute requests across both based on your haproxy config. But, at any point in time only one node would actually be routing requests through it's local instance of haproxy. I can't make it any simpler than that. Draw a diagram of what you are trying to do if it doesn't make sense. On 11/29/12 2:06 PM, Hermes Flying wrote: You are saying that one instance of HAProxy runs in each system and one instance is assigned the VIP that clients hit-on (out of scope for HAProxy). But this HAProxy distributes the requests according to the load, either on system-A or system-B for which you seem to refer to as backup system. In what way are you now refering to it as backup system? Because I am interested in distributing the load to all the nodes. *From:* David Coulson mailto:da...@davidcoulson.net *To:* Hermes Flying mailto:flyingher...@yahoo.com *Cc:* Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org *Sent:* Thursday, November 29, 2012 8:57 PM *Subject:* Re: HAproxy and detect split-brain (network failures) You can do that, but haproxy doesn't have anything to do with the failover process, other than you run an instance of haproxy on one server, and another instance on your backup system. As I said, neither of the haproxy instances communicate anything, so all you need to do is move the IP clients are using from one server to the other in order to handle a failure. Moving the IP around is something keepalived, pacemaker, etc handles - Look at their documentation for specifics and challenges in a two-node config. HAProxy doesn't have a concent of primary and backup in terms of it's own instances. Each of them is stand alone. It's up to you, based on your network/IP config which one has traffic routed to it. David On 11/29/12 1:53 PM, Hermes Flying wrote: But if I install 2 HAProxy as load balancers, doesn't one act as the primary loadbalancer directing the load to the known servers while the secondary takes over load distribution as soon as the heartbeat fails? I remember reading this. Is this wrong? *From:* David Coulson mailto:da...@davidcoulson.net *To:* Hermes Flying mailto:flyingher...@yahoo.com *Cc:* Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org *Sent:* Thursday, November 29, 2012 8:39 PM *Subject:* Re: HAproxy and detect split-brain (network failures) You are mixing two totally different things together. 1) HAProxy will do periodic health checks of backend systems you are routing to. Depending if you configure something as 'backup' or 'not backup' will determine if/how traffic is routed to it. The backend systems do not 'take over'. Haproxy just routes traffic to systems based on your configuration. The backend systems don't know/care about the other backend nodes, unless your application requires it which is a different story and nothing to do with haproxy. HAproxy only cares about a single instance of itself - If you have more than one haproxy instance, they do
Re: HAproxy and detect split-brain (network failures)
I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment From: David Coulson da...@davidcoulson.net To: Hermes Flying flyingher...@yahoo.com Cc: Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org Sent: Thursday, November 29, 2012 9:29 PM Subject: Re: HAproxy and detect split-brain (network failures) Both haproxy instances have the same config, with the tomcat instances with the same weight, etc. Run something like keepalived or pacemaker to manage a VIP between the two boxes. That's it. Not sure about keepalived, but pacemaker can make sure haproxy is running, then either restart it or move the VIP if it is not running. David On 11/29/12 2:27 PM, Hermes Flying wrote: Something like the following: HAProxy1 Tomcat1 | +/\ | + |+---Tomcat2 + /+\ + + HAProxy2+++ HAProxy1 is in the same machine as Tomcat1 HAproxy2 is in the same machine as Tomcat2 HAProxy1 distributes the load among Tomcat1 and Tomcat2. I erroneously thought that HAProxy2 would take over when HAProxy1 crashed to distribute the load among Tomcat1/Tomcat2. So if both are independent what can I do? From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 9:12 PM Subject: Re: HAproxy and detect split-brain (network failures) Again, you are mixing everything up. HAProxy has it's own configuration - It defines what nodes your port 80 traffic (or whatever) is routed to. Haproxy does periodic health checks of these backend services to make sure they are available for requests. If you have multiple haproxy instances they will all independently do health checks and not share any of that information with each other. HAProxy will route traffic to all systems defined as a backend for a particular service based upon whatever criteria is in the haproxy config. You can run a two-node environment that is active/backup from a VIP perspective, but active/active from a haproxy service perspective - Each node would run Apache (or whatever your service is) and haproxy would distribute requests across both based on your haproxy config. But, at any point in time only one node would actually be routing requests through it's local instance of haproxy. I can't make it any simpler than that. Draw a diagram of what you are trying to do if it doesn't make sense. On 11/29/12 2:06 PM, Hermes Flying wrote: You are saying that one instance of HAProxy runs in each system and one instance is assigned the VIP that clients hit-on (out of scope for HAProxy). But this HAProxy distributes the requests according to the load, either on system-A or system-B for which you seem to refer to as backup system. In what way are you now refering to it as backup system? Because I am interested in distributing the load to all the nodes. From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 8:57 PM Subject: Re: HAproxy and detect split-brain (network failures) You can do that, but haproxy doesn't have anything to do with the failover process, other than you run an instance of haproxy on one server, and another instance on your backup system. As I said, neither of the haproxy instances communicate anything, so all you need to do is move the IP clients are using from one server to the other in order to handle a failure. Moving the IP around is something keepalived, pacemaker, etc handles - Look at their documentation for specifics and challenges in a two-node config. HAProxy doesn't have a concent of primary and backup in terms of it's own instances. Each of them is stand alone. It's up to you, based on your network/IP config which one has traffic routed to it. David On 11/29/12 1:53 PM, Hermes Flying wrote: But if I install 2 HAProxy as load balancers, doesn't one act as the primary loadbalancer directing the load to the known servers while the secondary
Re: HAproxy and detect split-brain (network failures)
we have exactly that setup with heartbeat, and 2 floating IPs. Working in production for 3 years now Owen On 29/11/2012 3:26 PM, David Coulson wrote: On 11/29/12 3:11 PM, Hermes Flying wrote: I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? Any two node cluster has risk of split brain. if you implement fencing/STONITH, you are in a better place. If you have a third node, that's even better, even if it does not actually run any services beyond the cluster software I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that. In my config I have haproxy configured as a cloned resource in pacemaker, so all nodes have the same pacemaker config for haproxy and it keeps haproxy running on all nodes all of the time.
Re: HAproxy and detect split-brain (network failures)
Hi Owen, How does the heartbeat this help for splitbrain? With heartbeat the nodes know that it can't talk to each other. They don't know if the other is down. If there is a different communication path between the nodes and the incoming requests, both can become primary assuming the other is down due to network failure of the communcation link So how does this work for your system? From: Owen MArinas omari...@woozworld.com To: haproxy@formilux.org Sent: Thursday, November 29, 2012 10:40 PM Subject: Re: HAproxy and detect split-brain (network failures) we have exactly that setup with heartbeat, and 2 floating IPs. Working in production for 3 years now Owen On 29/11/2012 3:26 PM, David Coulson wrote: On 11/29/12 3:11 PM, Hermes Flying wrote: I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? Any two node cluster has risk of split brain. if you implement fencing/STONITH, you are in a better place. If you have a third node, that's even better, even if it does not actually run any services beyond the cluster software I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that. In my config I have haproxy configured as a cloned resource in pacemaker, so all nodes have the same pacemaker config for haproxy and it keeps haproxy running on all nodes all of the time.
Re: HAproxy and detect split-brain (network failures)
Again, you need to talk to the pacemaker people for actual clustering information. The ping was so a node could detect it lost upstream connectivity, and move the VIP, otherwise the VIP may continue to run on a system which does not have access to your network. This has nothing at all to do with split brain. If you want to deal with split brain, add a third node. Period. You also want to have redundant heartbeat communication paths. You also want STONITH/fencing so if one node detects the other is down it'll power it off or crash it. I've not had issues with a two-node cluster with two diverse backend communication links and fencing enabled. David On 11/29/12 3:58 PM, Hermes Flying wrote: You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that How does this help for splitbrain? If I understand what you say, pacemaker will ping an IP and if successfull will assume that the other node has crashed. But what if the other node hasn't and it is just their communication link that failed? Won't both become primary? How does the ping help? *From:* David Coulson da...@davidcoulson.net *To:* Hermes Flying flyingher...@yahoo.com *Cc:* Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org *Sent:* Thursday, November 29, 2012 10:26 PM *Subject:* Re: HAproxy and detect split-brain (network failures) On 11/29/12 3:11 PM, Hermes Flying wrote: I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? Any two node cluster has risk of split brain. if you implement fencing/STONITH, you are in a better place. If you have a third node, that's even better, even if it does not actually run any services beyond the cluster software I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that. In my config I have haproxy configured as a cloned resource in pacemaker, so all nodes have the same pacemaker config for haproxy and it keeps haproxy running on all nodes all of the time.
Re: HAproxy and detect split-brain (network failures)
Thank you for your help. I take it that you are find Pacemaker reliable in your experience? Should I look into it? From: David Coulson da...@davidcoulson.net To: Hermes Flying flyingher...@yahoo.com Cc: Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org Sent: Thursday, November 29, 2012 11:04 PM Subject: Re: HAproxy and detect split-brain (network failures) Again, you need to talk to the pacemaker people for actual clustering information. The ping was so a node could detect it lost upstream connectivity, and move the VIP, otherwise the VIP may continue to run on a system which does not have access to your network. This has nothing at all to do with split brain. If you want to deal with split brain, add a third node. Period. You also want to have redundant heartbeat communication paths. You also want STONITH/fencing so if one node detects the other is down it'll power it off or crash it. I've not had issues with a two-node cluster with two diverse backend communication links and fencing enabled. David On 11/29/12 3:58 PM, Hermes Flying wrote: You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that How does this help for splitbrain? If I understand what you say, pacemaker will ping an IP and if successfull will assume that the other node has crashed. But what if the other node hasn't and it is just their communication link that failed? Won't both become primary? How does the ping help? From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 10:26 PM Subject: Re: HAproxy and detect split-brain (network failures) On 11/29/12 3:11 PM, Hermes Flying wrote: I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? Any two node cluster has risk of split brain. if you implement fencing/STONITH, you are in a better place. If you have a third node, that's even better, even if it does not actually run any services beyond the cluster software I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that. In my config I have haproxy configured as a cloned resource in pacemaker, so all nodes have the same pacemaker config for haproxy and it keeps haproxy running on all nodes all of the time.
Re: HAproxy and detect split-brain (network failures)
In general, yes, Pacemaker is reliable. If your config is wrong, you may still have an outage in the event of a failure. That said, if you are a business and need support, you probably want to use whatever clustering software ships with the distribution you use. I belive SuSE uses pacemaker, but RedHat still uses rgmanager. Pacemaker is tech preview in RHEL6 but will be mainline in 7. I believe RedHat employ some core developers of pacemaker. David On 11/29/12 4:10 PM, Hermes Flying wrote: Thank you for your help. I take it that you are find Pacemaker reliable in your experience? Should I look into it? *From:* David Coulson da...@davidcoulson.net *To:* Hermes Flying flyingher...@yahoo.com *Cc:* Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org *Sent:* Thursday, November 29, 2012 11:04 PM *Subject:* Re: HAproxy and detect split-brain (network failures) Again, you need to talk to the pacemaker people for actual clustering information. The ping was so a node could detect it lost upstream connectivity, and move the VIP, otherwise the VIP may continue to run on a system which does not have access to your network. This has nothing at all to do with split brain. If you want to deal with split brain, add a third node. Period. You also want to have redundant heartbeat communication paths. You also want STONITH/fencing so if one node detects the other is down it'll power it off or crash it. I've not had issues with a two-node cluster with two diverse backend communication links and fencing enabled. David On 11/29/12 3:58 PM, Hermes Flying wrote: You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that How does this help for splitbrain? If I understand what you say, pacemaker will ping an IP and if successfull will assume that the other node has crashed. But what if the other node hasn't and it is just their communication link that failed? Won't both become primary? How does the ping help? *From:* David Coulson mailto:da...@davidcoulson.net *To:* Hermes Flying mailto:flyingher...@yahoo.com *Cc:* Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org *Sent:* Thursday, November 29, 2012 10:26 PM *Subject:* Re: HAproxy and detect split-brain (network failures) On 11/29/12 3:11 PM, Hermes Flying wrote: I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? Any two node cluster has risk of split brain. if you implement fencing/STONITH, you are in a better place. If you have a third node, that's even better, even if it does not actually run any services beyond the cluster software I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that. In my config I have haproxy configured as a cloned resource in pacemaker, so all nodes have the same pacemaker config for haproxy and it keeps haproxy running on all nodes all of the time.
Re: HAproxy and detect split-brain (network failures)
Great help! Thank you for your time! Much appreciated! From: David Coulson da...@davidcoulson.net To: Hermes Flying flyingher...@yahoo.com Cc: Baptiste bed...@gmail.com; haproxy@formilux.org haproxy@formilux.org Sent: Thursday, November 29, 2012 11:13 PM Subject: Re: HAproxy and detect split-brain (network failures) In general, yes, Pacemaker is reliable. If your config is wrong, you may still have an outage in the event of a failure. That said, if you are a business and need support, you probably want to use whatever clustering software ships with the distribution you use. I belive SuSE uses pacemaker, but RedHat still uses rgmanager. Pacemaker is tech preview in RHEL6 but will be mainline in 7. I believe RedHat employ some core developers of pacemaker. David On 11/29/12 4:10 PM, Hermes Flying wrote: Thank you for your help. I take it that you are find Pacemaker reliable in your experience? Should I look into it? From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 11:04 PM Subject: Re: HAproxy and detect split-brain (network failures) Again, you need to talk to the pacemaker people for actual clustering information. The ping was so a node could detect it lost upstream connectivity, and move the VIP, otherwise the VIP may continue to run on a system which does not have access to your network. This has nothing at all to do with split brain. If you want to deal with split brain, add a third node. Period. You also want to have redundant heartbeat communication paths. You also want STONITH/fencing so if one node detects the other is down it'll power it off or crash it. I've not had issues with a two-node cluster with two diverse backend communication links and fencing enabled. David On 11/29/12 3:58 PM, Hermes Flying wrote: You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that How does this help for splitbrain? If I understand what you say, pacemaker will ping an IP and if successfull will assume that the other node has crashed. But what if the other node hasn't and it is just their communication link that failed? Won't both become primary? How does the ping help? From: David Coulson mailto:da...@davidcoulson.net To: Hermes Flying mailto:flyingher...@yahoo.com Cc: Baptiste mailto:bed...@gmail.com; mailto:haproxy@formilux.org mailto:haproxy@formilux.org Sent: Thursday, November 29, 2012 10:26 PM Subject: Re: HAproxy and detect split-brain (network failures) On 11/29/12 3:11 PM, Hermes Flying wrote: I see now! One last question since you are using Pacemaker. Do you recommend it for splitbrain so that I look into that direction? Any two node cluster has risk of split brain. if you implement fencing/STONITH, you are in a better place. If you have a third node, that's even better, even if it does not actually run any services beyond the cluster software I mean when you say that pacemaker restart HAProxy, does it detect network failures as well? Or only SW crashes? I assume pacemaker will be aware of both HAProxy1 and HAProxy2 in my described deployment You can have pacemaker ping an IP (gateway for example) and migrate the VIP based on that. In my config I have haproxy configured as a cloned resource in pacemaker, so all nodes have the same pacemaker config for haproxy and it keeps haproxy running on all nodes all of the time.