Take the 2-minute tour ×
Server Fault is a question and answer site for system and network administrators. It's 100% free, no registration required.

The Background

We're running a Varnish cache server in front of a CF/Apache2 backend server. The varnish box runs a healthcheck probe every two seconds as follows:

probe healthcheck {
        .url = "/probe.cfm";
        .timeout = 5s;
        .interval = 2s;
        .window = 10;
        .threshold = 5;
        .initial = 5;
        .expected_response = 200;
}

backend web1 {
        .host = "<backend ip>";
        .port = "80";
        .probe = healthcheck;
}

The probe.cfm does this:

<cfoutput>
<!doctype html>
<!--[if lt IE 7]> <html class="no-js ie6 oldie" lang="en"> <![endif]-->
<!--[if IE 7]>    <html class="no-js ie7 oldie" lang="en"> <![endif]-->
<!--[if IE 8]>    <html class="no-js ie8 oldie" lang="en"> <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en"> <!--<![endif]-->
<head>
        <meta charset="utf-8">
        <title>CF Probe</title>
</head>

<body>
</cfoutput>

<cfquery name="qryProbe" datasource="#Request.DSN#">
        SELECT TOP 1 [PageID] FROM [Page] WHERE [PageID] > 6
</cfquery>

<cfoutput>
        #Variables.qryProbe.RecordCount#
        </body>
</html>
</cfoutput>

This snippet selects a single record from the underlying DB (mapped in data sources) and returns a 200 if it's successful.

Later in the Varnish config there's a section that tests if the backend is up and if not, then it sets the grace period on the cache to 24 hours and for any pages not in cache, it should generate a synthetic maintenance page.

sub vcl_recv

if (req.backend.healthy) { set req.grace = 30s; } else { set req.grace = 24h; }

sub vcl_error

if (!req.backend.healthy && obj.status != 200 && obj.status != 403 && obj.status != 404 && obj.status != 301 && obj.status != 302) {
    synthetic{"<some HTML here>"}
}

The Problem

We recently had a condition wherein the CF instance wasn't strictly-speaking unresponsive but it wasn't serving pages. According to the varnish logs, however, the backend was still healthy, so Varnish quickly stopped serving content, too.

Additionally, I saw at least one instance where the backend was generating error 500s while we restarted the CF instance and Varnish was still reporting it as healthy, despite the health test line in the log saying it received a 500 from the backend.

The Question

How can I more accurately test the health status of the CF backend so Varnish responds correctly to outages/reboots/etc.

I suppose, additionally, can anyone see any glaring flaws in the Varnish tests for backend health that I've set up and the tests that determine whether the synthetic HTML gets rendered or not?

share|improve this question

1 Answer 1

Well, it's configured such that it takes 12 seconds of consecutive erroring (6 checks, 2 seconds apart) for the backend to be marked as unhealthy, so there's certainly the potential for user requests to come in and be served bad content while the probe collects enough bad responses to mark the backend down.

You can adjust your probe's window/threshold settings as desired to cut that timeframe down, and consider setting up saint mode as a way to handle immediately treating bad responses as such without needing to collect failures on a probe or send those bad responses to users.

share|improve this answer

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.