Package Management For Lusers
If you run Linux on your own machines, you're used to having root
and doing what you jolly well like. But, if you've ever spent more
than about 10 minutes in a large coporate IT environment, you learn
pretty quickly that root
is hard to get, it takes a ton of
paperwork to get anything done, and you usually have to wait forever.
I've actually had the experience of waiting for 6 weeks to get
permission to install as symlink ... and I had root
!
There is a good reason for this, of course. Security threats are very real, lawsuits are ominipresent, and the Geniuses In Charge (tm) are writing regulation and audit compliance rules that make root canals seem like fun. Information Security people may feel like they are the IRS of the business, but they perform an important and necessary task: Saying "No".
So ... is there a better way? Is there a way to eliminate the
requirement for root
in most day-to-day things we need to do as
users and developers. Is there a way we can comply with the required
corporte security constraints, but still run our own happy show? The
answer is a qualified "Yes".
Some things do- and always will need root
: Managing devices,
storage, ulimits, and security configuration leap to mind. But, say,
all you want is a newer version of java
on your servers. Or
suppose you want a package that isn't part of your standard OS load.
vi
is everywhere, but suppose you want to use emacs
instead
(as you should).
You could, of course, download the source for the programs you want,
configure and compile them, and run them, say, out of your home
directory. Oops ... standard IT corporate security practice is to
never allow a compiler to exist on a production host. There are ways
around this, but it's fairly painful to have to do that for every
single package you may want. (If you don't think so, I encourage you
to try and bootstrap the gcc
compiler chain from scratch. It's a
ton of fun. No, really, it is ...)
Wouldn't it be nice if we could implement package management in
userland in a way that is repeatable, can be automated, and gives us
control of our own universe without having to beg for root
changes
or have to wait for the vendor to release a new package. Well,
Sparky, we have the techology to do just that.
It's worth mentioning that the approach outlined below is especially
handy with cloud and on-demand computing. It makes automating your
deploys pretty simple. It's also actually pretty handy on your own
machines when you do have root
. The less you use superuser, the
less chance you'll screw something up.
Warning
What follows has been implemented on an experimental basis. It's been tested in only a very limited number of systems but seems to work well. However, you should do your own detailed testing before deploying this into a production environment. Failure to do so may result in broken systems, hallway snikering, hives, and being transferred to your new development shop in Adak, AK.
The approach we're going to describe got started in the Mac OSX world.
Back when Apple finally came to their senses, and switched their OS to
a Unix-base (FreeBSD 4.4), they only partly implemented the shell
tools everyone had come to know and love. The brew
project got
spun up to allow any OSX user to install the command line applications
they knew and loved from Unix. brew
is essentially a userland
package management system which can be run and modified without superuser
power. Many of the brew
packages (these days, perhaps all, I haven't
checked) actually download a pre-compiled version under /usr/local
.
This ended up being pretty popular with advanced Mac users. So much
so, that a derivative project, linuxbrew
, got spun up to take the
Mac stuff, but apply it to Linux. That is, give the Linux user
userland package management system. It too, has found success among
the Linux literati.
But ... there is a fly in the ointment. When I first undertook this
project, I thought I could just pick a directory on Linux machine and
use linuxbrew
to install what I wanted. No habla Senor Frog.
Many Linux binaries are sensitive to where they are installed, where
they can find their supporting libraries and a host of other things.
So, if I install a binary with linuxbrew
somewhere other than the
default /home/linuxbrew
, it's likely not going to work. And I
wanted that to work. I wanted to have a way of creating a tools
tree wherever I jolly well felt like putting it.
"So", sez me, "I'll just use linuxbrew
to automate the download,
configuration, compliation, and installation of all the packages."
i.e. "I'll automate the build from source." (That roaring laughing
you hear is coming from every Linux engineer who ever tried something
like this.)
I will spare you sensitive readers the subsequent cursing, whining,
begging, crying and caterwauling that ensured. Let's just say that
making a position-dependent package management system work in a
position-independent way is ... er, non-trivial. In fairness, it's
not the fault of the linuxbrew
people. They were super supportive
and helpful with all this. A lot of the issues have to do with the
packages themselves having embedded assumptions about where they can
find tools during the compilation phase. That's right, the source
code and configurations have hardwired assumptions about where they
would find things like perl
and make
.
At this point, the whole process had taken me a few dozen hours and I was sufficiently enraged that I just had to figure out. As we'll see shortly, I think I finally go there. But, in the mean time ...
Note
If you write software, config files, makefiles, test cases,
or any part of the software delivery ecosystem with hardwired
paths to things emebedded in them, you are officially a
big bozo. Not the fun kind with a red nose and big shoes either.
The only hardwired path that's
OK is /bin/sh
on a shebang line. But if you do things
like this:
#!/usr/bin/python
You should be sent to work 1st level phone support on the midnight shift in Somalia until you learn better. Grrrrrr.
This is the right way to do this is:
#!/usr/bin/env python
env
can reliably be found there and it will "discover"
where python
happens to actually be installed on that
machine, so long as it is in $PATH somewhere. Similarly,
learn to use constructs like:
SED=`which sed`
$SED something
DATE=`which date`
DATE=${DATE:-"/bin/date"
In short, NEVER make assumptions where things are. Always discover it at configuration time.
What I eventually (after many hours of whining, etc.) disovered, was that getting this to work required a number of key things:
Everything has to be built from source in the directory location being targeted. The only exception is the
brew
program itself, which is position agnostic. So, if I want to build a tools tree under/my/fine/tools
, then I have to clonelinuxbrew
into that directory and do the build from there.The initial build requires the OS compiler chain and related development tools to bootstrap up a minimal
linuxbrew
environment capable of compiling everything else. You can do this on your own machine (not recommended because you shouldn't be fidding around as root there), but a better way is to do it in a VM. In my case, I made it even simpler by doing everything indocker
containers.Once you have a bootstrapped
linuxbrew
environment running - i.e., One that has a functioninggcc
and supporting tool chain - you make atar
backup of it. You then untar that onto a machine that has (almost) no native OS development tools on it and do the remainder of the installations from there.It's "almost" because - due to the aforementioned dain bramaged open source packages, You have to have the OS copies of
autoconfig
,automake
,perl
, andmake
installed on your build machine. These open source packages just insist thatperl
is always to be found under/usr/bin
, for example.
When you're all done installing and configuring your
linuxbrew
environment, you just `tar`` it off somewhere safe. You can then untar it onto any other Linux machine (with a reasonably current kernel) so long as you do so at the same directory location under which it was built.This lends itself nicely to automated deploys via tools like
tsshbatch
or ``ansible`. You build a master tarball of your "standard" tools tree and then use automated deployment to put it everywhere.
docker
WayLike I said, you can do this in a VM, but the step-by-step
approach below uses docker
containers which are easy to
setup and tear down for testing.