Bug 60948 - Large TCP timeout delays hcheck disabling a node
Summary: Large TCP timeout delays hcheck disabling a node
Status: NEW
Alias: None
Product: Apache httpd-2
Classification: Unclassified
Component: mod_proxy_hcheck (show other bugs)
Version: 2.4.25
Hardware: Sun Solaris
: P2 enhancement (vote)
Target Milestone: ---
Assignee: Apache HTTPD Bugs Mailing List
URL:
Keywords: PatchAvailable
Depends on:
Blocks:
 
Reported: 2017-03-31 14:59 UTC by Michael Renz
Modified: 2022-01-28 15:58 UTC (History)
0 users



Attachments
added new hcconnectiontimeout parameter (1.55 KB, patch)
2017-03-31 14:59 UTC, Michael Renz
Details | Diff
I forgot to allow it in ProxyHCTemplate and the parameter is now optional (3.19 KB, patch)
2017-04-03 13:19 UTC, Michael Renz
Details | Diff

Note You need to log in before you can comment on or make changes to this bug.
Description Michael Renz 2017-03-31 14:59:41 UTC
Created attachment 34892 [details]
added new hcconnectiontimeout parameter

Using latest patched mod_proxy_hcheck (with patch from bug 60071) I encountered a problematic situation. 
If a node goes down due to a complete failure and is not reachable via tcp/ip anymore, the long solaris tcp/ip timeout causes mod_proxy_hcheck to DISABLE the node very late. 
mod_proxy_hcheck does not provide a connection-timeout parameter to shorten this.
On top, the threadpool defined via ProxyHCTPsize quickly fills up and uses all available threads waiting for the timeout. The workaround is to increase ProxyHCTPsize to e.g. 500. But the problem remains, that once the node goes down it is not DISABLED until the first timeout has been reached. Solaris has a timeout of about 120s, therefore the problematic node will still get requests during this time. These requests will run into the "connectiontimeout", but this is still not a good situation as it slows down many requests.

I have patched (well, more copy/paste) the mod_proxy_hcheck.c and added a new parameter called "hcconnectiontimeout". With this new parameter my tests look good now.
Example configuration would look like this:

   SSLProxyEngine On
   SSLProxyVerify none
   SSLProxyCheckPeerCN off
   SSLProxyCheckPeerName off
   SSLProxyCheckPeerExpire off

   ProxyHCTPsize 400
   ProxyHCExpr get {hc('body') =~ /OK/}
   ProxyHCTemplate server hcmethod=GET hcexpr=get hcfails=1 hcinterval=2 hcpasses=1 hcuri=/tester
   <Proxy balancer://group>
      BalancerMember https://192.168.0.2:8080 connectiontimeout=1 hcconnectiontimeout=1 hctemplate=server
      BalancerMember https://192.168.0.3:8080 connectiontimeout=1 hcconnectiontimeout=1 hctemplate=server
   </Proxy>
<VirtualHost *:80>

   ProxyPass "/" "balancer://group/" failontimeout=On timeout=2
   ProxyPassReverse "/" "balancer://group/"

</VirtualHost>

I hope this helps anyone.
Comment 1 Michael Renz 2017-04-03 13:19:15 UTC
Created attachment 34893 [details]
I forgot to allow it in ProxyHCTemplate and the parameter is now optional

I forgot to allow it in ProxyHCTemplate and the parameter is now optional
Comment 2 Thomas Meyer 2019-06-03 07:42:03 UTC
Hi, any updates on this?

an independent timeout for the health check http request would be really helpful!

the patch looks okay, any thing that I can do to get this merged?
Comment 3 jfclere 2019-06-27 05:47:32 UTC
it is confusing to have connectiontimeout and hcconnectiontimeout. 
I have committed http://svn.apache.org/viewvc?view=revision&revision=1862014
if that helps please close the BZ.
Comment 4 goldyliang 2021-12-24 04:40:49 UTC
This fix seems to only fixes the case when there is a high latency in the TCP connection.  

However, there is a case we encountered now, that the back-end server does not respond to the health check http request due to an internal issue or resource exhaustion after the connection has been established and the request has been sent. In such a case, the health check from httpd seems to hang there without any read timeout. 

The consequence is that the load balancer is not able to mark a backend as HCFL, and it seems to be blocking the health check towards other balancer members as well.  In this case, the health check is not able to function as it should be.

Please advice if my understanding is right. If yes, how can we add the read timeout for health checks.
Comment 5 Ruediger Pluem 2022-01-07 10:46:15 UTC
Looks like the backport of r1889936 is missing. Can you check if http://svn.apache.org/viewvc/httpd/httpd/trunk/modules/proxy/mod_proxy_hcheck.c?r1=1889936&r2=1889935&pathrev=1889936&view=patch fixes your issue?
Comment 6 goldyliang 2022-01-28 15:58:11 UTC
Thank you for pointing out this, it is helpful. I will test it and hopefully if it works fine, it can be in the coming release.

I quickly looked in the patch code, looks like it tries to use the same timeout setting from the balancer config. The concern of that is, in some load balancer config we have to set a large timeout (like 5 minutes) to fit some heavy requests which take minutes to complete. But as per health check, we really expect it to timeout in seconds.  It would be better if the health check timeout can be set differently.