Changes in Varnish 6.0

Usually when we do dot-zero releases in Varnish, it means that users are in for a bit of work to upgrade to the new version, but 6.0 is actually not that scary, because most of the changes are either under the hood or entirely new features.

The biggest user-visible change is probably that we, or to be totally honest here: Geoff Simmons (UPLEX), have added support for Unix Domain Sockets, both for clients and for backend servers.

Because UNIX Domain Sockets have nothing like IP numbers, we were forced to define a new level of the VCL language vcl 4.1 to cope with UDS.

Both vcl 4.0 and vcl 4.1 are supported, and it is the primary source-file which controls which it will be, and you can include lower versions, but not higher versions than that.

Some old variables are not available in 4.1 and some new variables are not available in 4.0. Please see local, server, remote and client for specifics.

There are a few other changes to the vcl 4.0, most notably that we now consider upper- and lower-case the same for symbols.

The HTTP/2 code has received a lot of attention from Dag Haavi Finstad (Varnish Software) and it holds up in production on several large sites now.

There are new and improved VMODs:

The umem stevedore has been brought back on Solaris and it is the default storage method there now.

More error situations now get vcl failure handling, this should make life simpler for everybody we hope.

And it goes without saying that we have fixed a lot of bugs too.

Under the hood (mostly for developers)

The big thing is that the $Abi [vrt|strict] should now have settled. We have removed all the stuff from <cache.h> which is not available under $Abi vrt, and this hopefully means that VMODS will work without recompilation on several subsequent varnish versions. (There are some stuff related to packaging which takes advantage of this, but it didn’t get into this release.)

VMODS can define their own stats counters now, and they work just like builtin counters, because there is no difference.

The counters are described in a .vsc file which is processed with a new python script which does a lot of magic etc. There is a tiny example in vmod_debug in the source tree. If you’re using autotools, a new VARNISH_COUNTERS macro helps you set everything up, and is documented in varnish.m4.

This took a major retooling of the stats counters in general, and the VSM, VSC and VSL apis have all subtly or not so subtly changed as a result.

VMOD functions can take optional arguments, these are different from defaulted arguments in that a separate flag tells if they were specified or not in the call. For reasons of everybody’s sanity, all the arguments gets wrapped in a function-specific structure when this is used.

The vmodtool.py script has learned other new tricks, and as a result also produces nicer .rst output.

VCL types INT and BYTES are now 64bits on all platforms.

VCL ENUM have gotten a new implementation, so the pointers are now constant and can be compared as such, rather than with strcmp(3).

We have a new type of binary VSL records which are hexdumped by default, but on the API side, rather than in varnishd. This saves both VSL bandwidth and processing power, as they are usually only used for deep debugging and mostly turned off.

The VCC compilers has received a lot of work in two areas:

The symbol table has been totally revamped to make it ready for variant symbols, presently symbols which are different in vcl 4.0 and vcl 4.1.

The “prototype” information in the VMOD shared library has been changed to JSON, (look in your vcc_if.c file if you don’t believe me), and this can express more detailed information, presently the optional arguments.

The stuff only we care about

Varnishtest’s process has grown pty(4) support, so that we can test curses-based programs like our own utilities.

This has (finally!) pushed our code coverage, across all the source code in the project up to 90%.

We have also decided to make our python scripts PEP8 compliant, and vmodtool.py is already be there.

The VCL variables are now defined in the .rst file, rather than the other way around, this makes the documentation better at the cost of minor python-script complexity.

We now produce weekly snapshots from -trunk, this makes it easier for people to test all the new stuff.

We have not quite gotten the half-yearly release-procedure under control.

I’m writing this the evening before the release, trying to squeeze out of my brain what I should have written here long time ago, and we have had far more commits this last week than is reasonable.

But we have gotten better at it.

Really!

eof