On Wed, Nov 12, 2014 at 05:11:27PM +0900, Simon Horman wrote: > On Wed, Nov 12, 2014 at 08:22:05AM +0100, Willy Tarreau wrote: > > Hi Simon, > > > > On Wed, Nov 12, 2014 at 03:55:53PM +0900, Simon Horman wrote: > > > disable starts a server in the disabled state, however setting the health > > > of an agent implies that the agent is disabled as well as the server. > > > > > > This is a problem because the state of the agent is not restored if > > > the state of the server is subsequently updated leading to an > > > unexpected state. > > > > > > For example, if a server is started disabled and then the server > > > state is set to ready then without this change show stat indicates > > > that the server is "DOWN (agent)" when it is expected that the server > > > would be UP if its (non-agent) health check passes. > > > > Interesting case. I believe I caused it myself while trying to address > > a different case : health checks are disabled, only agent checks are > > enabled, and the server is disabled in the configuration. Could you > > please check that this use case still works properly with your patch ? > > I'd rather avoid to see the server continue to show up! > > Thanks, will do. > > I was aware you had done some work in this area but I wasn't entirely > sure what case you were trying to fix. Thanks for filling in that gap > in my knowledge.
Hi Willy, I have tested the following scenario which I hope matches the one that you describe: 1. Start haproxy with server disabled in config 2. Disable health checks using: echo "disable health VIP_Name/RIP_Name" | socat .. 2. Enable server using: echo "set server VIP_Name/RIP_Name state ready" | socat ... The results are as follows: 1. With this patch applied and agent-check enabled Server status is reported as "" 2. With this patch applied and agent-check not configured in config Server status is reported as "" 3. Without this patch applied and agent-check enabled Server status is reported as "DOWN (agent)" 4. With this patch applied and agent-check not configured in config Server status is reported as "" My working assumption is that unless the agent-check explicitly marks a backend as down then the backend should not be considered down due to the agent-check. This includes this scenario when the agent-check is not responding. This seems to be reflected in the implementation other than the area of HTTP statistics. As both 1) and 2) are consistent with 4) it seems to me that this patch is correct in the context of the scenario you describe (assuming my test is correct).
