/bin/sh and /usr/bin/env are the only two binaries that NixOS provides
at a fixed path (outside a buildFHSUserEnv sandbox).
This discussion was split from #11004.
Signed-off-by: Anders Kaseorg <andersk@mit.edu>
In tools/test-install/destroy-all line 31:
| while read c
^-- SC2162: read without -r will mangle backslashes.
In tools/test-install/install line 57:
installer_dir="$(readlink -f $INSTALLER)"
^-- SC2086: Double quote to prevent globbing and word splitting.
In tools/test-install/lxc-wait line 30:
for i in {1..60}; do
^-- SC2034: i appears unused. Verify use (or export if used externally).
Signed-off-by: Anders Kaseorg <andersk@mit.edu>
Otherwise prepare-base is likely to fail when first run (but then
succeed when rerun, because the container is left running), because
the container isn't up yet when we try to operate in it.
Also clean up the placement of `set -e` vs `set -x`.
(This is a small fixup to the main change, which was accidentally
included in a previous commit:
08bbd7e61 "settings: Slightly simplify EMAIL_BACKEND logic."
Oops. See there for most of the changes described here.)
The installer works out of a release-tarball tree. We typically want
to share this tree between successive test-install runs (with an rsync
or similar command to update source files of interest) because
rebuilding a release tree from scratch is slow. But the installer
will munge the tree; so instead of directly bind-mounting the tree
into the container, we need to give it an overlay over the tree, as a
sandbox to play in.
Previously we used lxc-copy's `-m overlay=...` feature to do this,
mounting an overlay in the container. But then sometimes in
development we want to reach in and edit some code in the tree,
e.g. before rerunning the installer after something failed. Reaching
inside the container for this is a pain (`ssh` would add latency, and
I haven't installed sshd in the containers; and getting rsync to work
with `lxc-attach` was beyond what I could figure out in a few minutes
of fiddling); and editing the base tree often doesn't work.
So, create the overlay with our own `mount -t overlay`, and have
`lxc-copy` just bind-mount that in. Now the host has direct access to
the same overlay which the guest is working from.
Also this makes it past time to help the user out in finding the fresh
names we've created: first the container, now this shared tree. Print
those at the end, rather than make the user scroll to the top and find
the right `set -x` line to copy-paste from.
Before this fix, the installer has an extremely annoying bug where
when run inside a container with `lxc-attach`, when the installer
finishes, the `lxc-attach` just hangs and doesn't respond even to
C-c or C-z. The only way to get the terminal back is to root around
from some other terminal to find the PID and kill it; then run
something like `stty sane` to fix the messed-up terminal settings
left behind.
After bisecting pieces of the install script to locate which step
was causing the issue, it comes down to the `service camo restart`.
The comment here indicates that we knew about an annoying bug here
years ago, and just swept it under the rug by skipping this step
when in Travis. >_<
The issue can be reproduced by running simply `service camo restart`
under `lxc-attach` instead of the installer; or `service camo start`,
following a `service camo stop`. If `lxc-attach` is used to get an
interactive shell, these commands appear to work fine; but then when
that shell exits, the same hang appears. So, when we start camo
we're evidently leaving some kind of mess that entangles the daemon
with our shell.
Looking at the camo initscript where it starts the daemon, there's
not much code, and one flag jumps out as suspicious:
start-stop-daemon --start --quiet --pidfile $PIDFILE -bm \
--exec $DAEMON --no-close -c nobody --test > /dev/null 2>&1 \
|| return 1
start-stop-daemon --start --quiet --pidfile $PIDFILE -bm \
--no-close -c nobody --exec $DAEMON -- \
$DAEMON_ARGS >> /var/log/camo/camo.log 2>&1 \
|| return 2
What does `--no-close` do?
-C, --no-close
Do not close any file descriptor when forcing the daemon
into the background (since version 1.16.5). Used for
debugging purposes to see the process output, or to
redirect file descriptors to log the process output.
And in fact, looking in /proc/PID/fd while a hang is happening finds
that fd 0 on the camo daemon process, aka stdin, is connected to our
terminal.
So, stop that by denying the initscript our stdin in the first place.
This fixes the problem.
The Debian maintainer turns out to be "Zulip Debian Packaging Team",
at debian@zulip.com; so this package and its bugs are basically ours.
This is a tool that throws away `fsync` calls and other requests for
the system to sync files to disk. It may make the install faster; for
example, if it has to install a number of system packages, `dpkg` is
known to make a lot of `fsync` calls which slow things down
significantly. Conversely, if there's a power failure in the middle
of running a test install, we really don't mind if the test install's
data becomes corrupt.
When the install script is successful, one of the final things it
wants to do is to move the tree that Zulip was installed from into the
deployments directory. It can't do that, at least not in a naive way
with `mv`, if the tree is actually a mount point. So, stick the tree
inside some other directory that we create just for the purpose of
being the mount point and containing the install tree.
This saves several minutes off the install time. Sadly pip still
clones Git repos for dependencies that point to them, but for many
others (not all? not sure) it just gets a wheel from the cache.
The third-party `install-yarn.sh` script uses `curl`, and we invoke it
in `install-node`. So we need to install it as a dependency.
We've mostly gotten away with this because it's common for `curl` to
already be installed; but it isn't always.
This greatly simplifies iterating on changes to the installer and
associated code: just edit in the shared directory (or edit in your
worktree and rsync to the directory), and rerun.
With this change, the form with a directory is now really the main
way to run the script; the form accepting a tarball is really just
a convenience feature, unpacking the tarball and then proceeding with
that directory.
This will facilitate testing interesting installer features
using its own CLI.
On my laptop, with a recent base image (updated a few days ago with
`prepare-base`), it takes just 7 or 8 seconds to get to the installer
running, as timed by passing `--help` so that the installer promptly
exits.
In order to do development on the installer itself in a sane way,
we need a reasonably fast and automatic way to get a fresh environment
to try to run it in.
This calls for some form of virtualization. Choices include
* A public cloud, like EC2 or Digital Ocean. These could work, if we
wrote some suitable scripts against their APIs, to manage
appropriate base images (as AMIs or snapshots respectively) and to
start fresh instances/droplets from a base image. There'd be some
latency on starting a new VM, and this would also require the user
to have an account on the relevant cloud with API access to create
images and VMs.
* A local whole-machine VM system (hypervisor) like VirtualBox or
VMware, perhaps managing the configuration through Vagrant. These
hypervisors can be unstable and painfully slow. They're often the
only way to get development work done on a Mac or Windows machine,
which is why we use them there for the normal Zulip development
environment; but I don't really want to find out how their
instability scales when constantly spawning fresh VMs from an image.
* Containers. The new hotness, the name on everyone's lips, is Docker.
But Docker is not designed for virtualizing a traditional Unix server,
complete with its own init system and a fleet of processes with a
shared filesystem -- in other words, the platform Zulip's installer
and deployment system are for. Docker brings its own quite
different model of deployment, and someday we may port Zulip from
the traditional Unix server to the Docker-style deployment model,
but for testing our traditional-Unix-server deployment we need a
(virtualized) traditional Unix server.
* Containers, with LXC. LXC provides containers that function as
traditional Unix servers; because of the magic of containers, the
overhead is quite low, and LXC offers handy snapshotting features
so that we can quickly start up a fresh environment from a base
image. Running LXC does require a Linux base system. For
contributors whose local development machine isn't already Linux,
the same solutions are available as for our normal development
environment: the base system for running LXC could be e.g. a
Vagrant-managed VirtualBox VM, or a machine in a public cloud.
This commit adds a first version of such a thing, using LXC to manage
a base image plus a fresh container for each test run. The test
containers function as VMs: once installed, all the Zulip services run
normally in them and can be managed in the normal production ways.
This initial version has a shortage of usage messages or docs, and
likely has some sharp edges. It also requires familiarity with the
basics of LXC commands in order to make good use of the resulting
containers: `lxc-ls -f`, `lxc-attach`, `lxc-stop`, and `lxc-start`,
in particular.
Anders Kaseorg
Greg Price
Anders Kaseorg
Aditya Bansal