Backend servers

Varnish has a concept of “backend” or “origin” servers. A backend server is the server providing the content Varnish will accelerate.

Our first task is to tell Varnish where it can find its backends. Start your favorite text editor and open the relevant VCL file.

Somewhere in the top there will be a section that looks a bit like this.:

# backend default {
#     .host = "127.0.0.1";
#     .port = "8080";
# }

We remove the comment markings in this text stanza making the it look like.:

backend default {
    .host = "127.0.0.1";
    .port = "8080";
}

Now, this piece of configuration defines a backend in Varnish called default. When Varnish needs to get content from this backend it will connect to port 8080 on localhost (127.0.0.1).

Varnish can have several backends defined you can even join several backends together into clusters of backends for load balancing purposes.

The “none” backend

Backends can also be declared as none with the following syntax::

backend default none;

none backends are special:

• All backends declared none compare equal:

  backend a none;
  backend b none;
  
  sub vcl_recv {
      set req.backend_hint = a;
      if (req.backend_hint == b) {
          return (synth(200, "this is true"));
      }
  }
  

• The none backend evaluates to false when used in a boolean context:

  backend nil none;
  
  sub vcl_recv {
      set req.backend_hint = nil;
      if (! req.backend_hint) {
          return (synth(200, "We get here"));
      }
  }
  

• When directors find no healthy backend, they typically return the none backend

Multiple backends

At some point you might need Varnish to cache content from several servers. You might want Varnish to map all the URL into one single host or not. There are lot of options.

Lets say we need to introduce a Java application into out PHP web site. Lets say our Java application should handle URL beginning with /java/.

We manage to get the thing up and running on port 8000. Now, lets have a look at the default.vcl.:

backend default {
    .host = "127.0.0.1";
    .port = "8080";
}

We add a new backend.:

backend java {
    .host = "127.0.0.1";
    .port = "8000";
}

Now we need tell Varnish where to send the difference URL. Lets look at vcl_recv.:

sub vcl_recv {
    if (req.url ~ "^/java/") {
        set req.backend_hint = java;
    } else {
        set req.backend_hint = default;
    }
}

It’s quite simple, really. Lets stop and think about this for a moment. As you can see you can define how you choose backends based on really arbitrary data. You want to send mobile devices to a different backend? No problem. if (req.http.User-agent ~ /mobile/) .. should do the trick.

Without an explicit backend selection, Varnish will continue using the default backend. If there is no backend named default, the first backend found in the vcl will be used as the default backend.

Backends and virtual hosts in Varnish

Varnish fully supports virtual hosts. They might however work in a somewhat counter-intuitive fashion since they are never declared explicitly. You set up the routing of incoming HTTP requests in vcl_recv. If you want this routing to be done on the basis of virtual hosts you just need to inspect req.http.host.

You can have something like this:

sub vcl_recv {
    if (req.http.host ~ "foo.com") {
        set req.backend_hint = foo;
    } elsif (req.http.host ~ "bar.com") {
        set req.backend_hint = bar;
    }
}

Note that the first regular expressions will match “foo.com”, “www.foo.com”, “zoop.foo.com” and any other host ending in “foo.com”. In this example this is intentional but you might want it to be a bit more tight, maybe relying on the == operator in stead, like this:

sub vcl_recv {
    if (req.http.host == "foo.com" || req.http.host == "www.foo.com") {
        set req.backend_hint = foo;
    }
}

Connecting Through a Proxy

As of this release, Varnish can connect to an actual destination through a proxy using the PROXY2 protocol. Other protocols may be added.

For now, a typical use case of this feature is to make TLS-encrypted connections through a TLS onloader. The onloader needs to support dynamic connections with the destination address information taken from a PROXY2 preamble. For example with haproxy Version 2.2 or higher, this snippet can be used as a basis for configuring an onloader:

# to review and adjust:
# - maxconn
# - bind ... mode ...
# - ca-file ...
#
listen sslon
       mode    tcp
       maxconn 1000
       bind    /path/to/sslon accept-proxy mode 777
       stick-table type ip size 100
       stick   on dst
       server  s00 0.0.0.0:0 ssl ca-file /etc/ssl/certs/ca-bundle.crt alpn http/1.1 sni fc_pp_authority
       server  s01 0.0.0.0:0 ssl ca-file /etc/ssl/certs/ca-bundle.crt alpn http/1.1 sni fc_pp_authority
       server  s02 0.0.0.0:0 ssl ca-file /etc/ssl/certs/ca-bundle.crt alpn http/1.1 sni fc_pp_authority
       # ...
       # A higher number of servers improves TLS session caching

Varnish running on the same server/namespace can then use the onloader with the .via feature (see Attribute .via):

backend sslon {
  .path = "/path/to/sslon";
}

backend destination {
  .host = "my.https.service";
  .port = "443";
  .via = sslon;
}

The .authority attribute can be used to specify the SNI for the connection if it differs from .host.

Directors

You can also group several backend into a group of backends. These groups are called directors. This will give you increased performance and resilience.

You can define several backends and group them together in a director. This requires you to load a VMOD, a Varnish module, and then to call certain actions in vcl_init.:

import directors;    # load the directors

backend server1 {
    .host = "192.168.0.10";
}
backend server2 {
    .host = "192.168.0.11";
}

sub vcl_init {
    new bar = directors.round_robin();
    bar.add_backend(server1);
    bar.add_backend(server2);
}

sub vcl_recv {
    # send all traffic to the bar director:
    set req.backend_hint = bar.backend();
}

This director is a round-robin director. This means the director will distribute the incoming requests on a round-robin basis. There is also a random director which distributes requests in a, you guessed it, random fashion. If that is not enough, you can also write your own director (see Writing a Director).

But what if one of your servers goes down? Can Varnish direct all the requests to the healthy server? Sure it can. This is where the Health Checks come into play.

Health checks

Lets set up a director with two backends and health checks. First let us define the backends:

backend server1 {
    .host = "server1.example.com";
    .probe = {
        .url = "/";
        .timeout = 1s;
        .interval = 5s;
        .window = 5;
        .threshold = 3;
    }
}

backend server2 {
    .host = "server2.example.com";
    .probe = {
        .url = "/";
        .timeout = 1s;
        .interval = 5s;
        .window = 5;
        .threshold = 3;
    }
}

What is new here is the probe. In this example Varnish will check the health of each backend every 5 seconds, timing out after 1 second. Each poll will send a GET request to /. If 3 out of the last 5 polls succeeded the backend is considered healthy, otherwise it will be marked as sick.

Refer to the Backend health probes section in the VCL documentation for more information.

Now we define the ‘director’:

import directors;

sub vcl_init {
    new vdir = directors.round_robin();
    vdir.add_backend(server1);
    vdir.add_backend(server2);
}

You use this vdir director as a backend_hint for requests, just like you would with a simple backend. Varnish will not send traffic to hosts that are marked as unhealthy.

Varnish can also serve stale content if all the backends are down. See Grace mode and keep for more information on how to enable this.

Please note that Varnish will keep health probes running for all loaded VCLs. Varnish will coalesce probes that seem identical - so be careful not to change the probe config if you do a lot of VCL loading. Unloading the VCL will discard the probes. For more information on how to do this please see ref:reference-vcl-director.

Layering

By default, most directors’ .backend() methods return a reference to the director itself. This allows for layering, like in this example:

import directors;

sub vcl_init {
    new dc1 = directors.round_robin();
    dc1.add_backend(server1A);
    dc1.add_backend(server1B);

    new dc2 = directors.round_robin();
    dc2.add_backend(server2A);
    dc2.add_backend(server2B);

    new dcprio = directors.fallback();
    dcprio.add_backend(dc1);
    dcprio.add_backend(dc2);
}

With this initialization, dcprio.backend() will resolve to either server1A or server1B if both are healthy or one of them if only one is healthy. Only if both are sick will a healthy server from dc2 be returned, if any.

Director Resolution

The actual resolution happens when the backend connection is prepared after a return from vcl_backend_fetch {} or vcl_pipe {}.

In some cases like server sharding the resolution outcome is required already in VCL. For such cases, the .resolve() method can be used, like in this example:

set req.backend_hint = dcprio.backend().resolve();

When using this statement with the previous example code, req.backend_hint will be set to one of the server* backends or the none backend if they were all sick.

.resolve() works on any object of the BACKEND type.

Connection Pooling

Opening connections to backends always comes at a cost: Depending on the type of connection and backend infrastructure, the overhead for opening a new connection ranges from pretty low for a local Unix domain socket (see Backend definition .path attribute) to substantial for establishing possibly multiple TCP and/or TLS connections over possibly multiple hops and long network paths. However relevant the overhead, it certainly always exists.

So because re-using existing connections can generally be considered to reduce overhead and latencies, Varnish pools backend connections by default: Whenever a backend task is finished, the used connection is not closed but rather added to a pool for later reuse. To avoid a connection from being reused, the Connection: close http header can be added in vcl_backend_fetch.

While backends are defined per VCL, connection pooling works across VCLs and even across backends: By default, the identifier for pooled connections is constructed from the .host/.port or .path attributes of the Backend definition (VMODs can make use of custom identifiers). So whenever two backends share the same address information, irrespective of which VCLs they are defined in, their connections are taken from a common pool.

If not actively closed by the backend, pooled connections are kept open by Varnish until the backend_idle_timeout expires.