Upgrading to Varnish 4

Changes to VCL

The backend fetch parts of VCL have changed in Varnish 4. We've tried to compile a list of changes needed to upgrade here.

Version statement

To make sure that people have upgraded their VCL to the current version, Varnish now requires the first line of VCL to indicate the VCL version number:

vcl 4.0;

req.request is now req.method

To align better with RFC naming, req.request has been renamed to req.method.

vcl_fetch is now vcl_backend_response

Directors have been moved to the vmod_directors

To make directors (backend selection logic) easier to extend, the directors are now defined in loadable VMODs.

Setting a backend for future fetches in vcl_recv is now done as follows:

sub vcl_init {
    new cluster1 = directors.round_robin();
    cluster1.add_backend(b1, 1.0);
    cluster1.add_backend(b2, 1.0);
}

sub vcl_recv {
    set req.backend_hint = cluster1.backend();
}

Note the extra .backend() needed after the director name.

Use the hash director as a client director

Since the client director was already a special case of the hash director, it has been removed, and you should use the hash director directly:

sub vcl_init {
    new h = directors.hash();
    h.add_backend(b1, 1);
    h.add_backend(b2, 1);
}

sub vcl_recv {
    set req.backend_hint = h.backend(client.identity);
}

vcl_error is now vcl_backend_error

To make a distinction between internally generated errors and VCL synthetic responses, vcl_backend_error will be called when varnish encounters an error when trying to fetch an object.

error() is now synth()

And you must explicitly return it:

return (synth(999, "Response"));

Synthetic responses in vcl_synth

Setting headers on synthetic response bodies made in vcl_synth are now done on resp.http instead of obj.http.

The synthetic keyword is now a function:

if (resp.status == 799) {
    set resp.status = 200;
    set resp.http.Content-Type = "text/plain; charset=utf-8";
    synthetic("You are " + client.ip);
    return (deliver);
}

obj in vcl_error replaced by beresp in vcl_backend_error

To better represent a the context in which it is called, you should now use beresp.* vcl_backend_error, where you used to use obj.* in vcl_error.

hit_for_pass objects are created using beresp.uncacheable

Example:

sub vcl_backend_response {
    if (beresp.http.X-No-Cache) {
        set beresp.uncacheable = true;
        set beresp.ttl = 120s;
        return (deliver);
    }
}

req.* not available in vcl_backend_response

req.* used to be available in vcl_fetch, but after the split of functionality, you only have 'bereq.*' in vcl_backend_response.

vcl_* reserved

Any custom-made subs cannot be named 'vcl_*' anymore. This namespace is reserved for builtin subs.

req.backend.healthy replaced by std.healthy(req.backend_hint)

Remember to import the std module if you're not doing so already.

client.port, and server.port replaced by respectively std.port(client.ip) and std.port(server.ip)

client.ip and server.ip are now proper datatypes, which renders as an IP address by default. You need to use the std.port() function to get the port number.

Invalidation with purge

Cache invalidation with purges is now done via return(purge) from vcl_recv. The purge; keyword has been retired.

obj is now read-only

obj is now read-only. obj.last_use has been retired.

Some return values have been replaced

Apart from the new synth return value described above, the following has changed:

  • vcl_recv must now return hash instead of lookup
  • vcl_hash must now return lookup instead of hash
  • vcl_pass must now return fetch instead of pass

Backend restarts are now retry

In 3.0 it was possible to do return(restart) after noticing that the backend response was wrong, to change to a different backend.

This is now called return(retry), and jumps back up to vcl_backend_fetch.

This only influences the backend fetch thread, client-side handling is not affected.

default/builtin VCL changes

The VCL code that is appended to user-configured VCL automatically is now called the builtin VCL. (previously default.vcl)

The builtin VCL now honors Cache-Control: no-cache (and friends) to indicate uncacheable content from the backend.

The remove keyword is gone

Replaced by unset.

X-Forwarded-For is now set before vcl_recv

In many cases, people unintentionally removed X-Forwarded-For when implementing their own vcl_recv. Therefore it has been moved to before vcl_recv, so if you don't want an IP added to it, you should remove it in vcl_recv.

Changes to existing parameters

session_linger

session_linger has been renamed to timeout_linger and it is in seconds now (previously was milliseconds).

sess_timeout

sess_timeout has been renamed to timeout_idle.

sess_workspace

In 3.0 it was often necessary to increase sess_workspace if a lot of VMODs, complex header operations or ESI were in use.

This is no longer necessary, because ESI scratch space happens elsewhere in 4.0.

If you are using a lot of VMODs, you may need to increase either workspace_backend and workspace_client based on where your VMOD is doing its work.

thread_pool_purge_delay

thread_pool_purge_delay has been renamed to thread_pool_destroy_delay and it is in seconds now (previously was milliseconds).

thread_pool_add_delay and thread_pool_fail_delay

They are in seconds now (previously were milliseconds).

New parameters since 3.0

vcc_allow_inline_c

You can now completely disable inline C in your VCL, and it is disabled by default.

Other changes

New log filtering

The logging framework has a new filtering language, which means that the -m switch has been replaced with a new -q switch. See Varnish VSL Query Expressions for more information about the new query language.