Those customization do not work anymore and it is preventing tests to run,
 so it should be reintroduced later in a way it works.

For upcoming GitLab SR we need Go[1] language support, because one
GitLab service is written in this language:

    https://gitlab.com/gitlab-org/gitlab-workhorse

Here we provide golang component, and helloweb-go service integrated
into helloworld SR.

The patches are based on recent helloworld & helloweb restructuring (see !23).

/reviewed-by @jerome  (on !24)
/cc @kazuhiko, @rafael, @alain.takoudjou, @gabriel, @Camata

[1] http://golang.org

For upcoming GitLab SR we need Ruby support.

We have minimal Ruby support in the form of Ruby component and
`rubygemsrecipe` recipe to install gems. However a lot of top-level
software/services in Ruby world is not released as gem and have to be
installed / worked with via Bundler[1], and for this way we do not have
an example.

Let's add such example in the form of extending helloworld SR to also
say hello to the web in both Python and Ruby via simple Ruby program
which is deployed via Bundler.

For this to happen, we need to make some preparatory changes first:

- move helloweb program(s) to own repository[2].
- modernize instance code and convert it to jinja2 to allow control structures.
- prepare instance infrastructure to support several helloweb program kinds.

and only then show how to do Ruby stuff via Bundler.

To me the effect of `software/helloworld` refactoring is good even without
Ruby part, just to improve reference material about showing how to do.

/generally-reviewed-by @jerome, @vpelletier  (on !23)
/cc @kazuhiko, @rafael, @alain.takoudjou, @cedric.leninivin

[1] http://bundler.io
[2] http://lab.nexedi.com/nexedi/helloweb

 Dump on filesytem the ipv4 used to the node create connections

It is well known that UNIX sockets are faster than TCP over loopback.

E.g. on my machine according to lmbench[1] they have ~ 2 times
lower latency and ~ 2-3 times more throughput compared to TCP over
loopback:

    *Local* Communication latencies in microseconds - smaller is better
    ---------------------------------------------------------------------
    Host                 OS 2p/0K  Pipe AF     UDP  RPC/   TCP  RPC/ TCP
                            ctxsw       UNIX         UDP         TCP conn
    --------- ------------- ----- ----- ---- ----- ----- ----- ----- ----
    teco      Linux 4.2.0-1  13.8  29.2 26.8  45.0  47.9  48.5  55.5  45.

    *Local* Communication bandwidths in MB/s - bigger is better
    -----------------------------------------------------------------------------
    Host                OS  Pipe AF    TCP  File   Mmap  Bcopy  Bcopy  Mem   Mem
                                 UNIX      reread reread (libc) (hand) read write
    --------- ------------- ---- ---- ---- ------ ------ ------ ------ ---- -----
    teco      Linux 4.2.0-1 1084 4353 1493 2329.1 3720.7 1613.8 1109.2 3402 1404.

The same ratio holds for our std shuttle servers.

API to work with unix sockets is essentially the same as with TCP/UDP.
Because of that it is easy to support both TCP and UNIX socket in one
software, and this way a lot of software support unix sockets out of the
box, including Redis.

Because of lower latencies and higher throughput, for performance
reasons, it makes sense to interconnect services on one machine via unix
sockets and talk via TCP only to outside world.

Here we add support for unix sockets to Redis recipe.

[1] http://www.bitmover.com/lmbench/

/reviewed-by @kazuhiko  (on !27)
/cc @alain.takoudjou, @jerome, @vpelletier

Because redis.Redis(...) ctor creates connection pool on initialization
and we can rely on it.

Another reason: Redis ctor (in form of StrictRedis.__init__()) has logic
how to process arguments and does selecting - either it is TCP (`host` and
`port` args), or UNIX socket (`unix_socket_path` arg):

    https://lab.nexedi.com/nexedi/slapos/blob/95dbb5b2/slapos/recipe/redis/MyRedis2410.py#L560

Since we are going to introduce unix socket support to redis recipe in
the next patch, and don't want to duplicate StrictRedis.__init__() logic
in promise code, let's refactor promise to delegate argument processing
logic to Redis.

/reviewed-by @kazuhiko  (on !27)
/cc @alain.takoudjou

- update Redis software to latest upstream in 2.8.* series (which now
  supports IPv6 out of the box);

- update Redis instance template to the one from 2.8.23 and re-merge our
  templating changes to it (file/dir locations, port and binding, master
  password).

  The whole diff to pristine 2.8.23 redis conf is now this:

  diff --git a/.../redis-2.8.23/redis.conf b/slapos/recipe/redis/template/redis.conf.in
  index 870959f..2895539 100644
  --- a/.../redis-2.8.23/redis.conf
  +++ b/slapos/recipe/redis/template/redis.conf.in
  @@ -46 +46 @@ daemonize no
  -pidfile /var/run/redis.pid
  +pidfile %(pid_file)s
  @@ -50 +50 @@ pidfile /var/run/redis.pid
  -port 6379
  +port %(port)s
  @@ -69,0 +70 @@ tcp-backlog 511
  +bind %(ipv6)s
  @@ -108 +109 @@ loglevel notice
  -logfile ""
  +logfile %(log_file)s
  @@ -174 +175 @@ rdbcompression yes
  -# hit to pay (around 10%) when saving and loading RDB files, so you can disable it
  +# hit to pay (around 10%%) when saving and loading RDB files, so you can disable it
  @@ -192 +193 @@ dbfilename dump.rdb
  -dir ./
  +dir %(server_dir)s
  @@ -217 +218 @@ dir ./
  -# masterauth <master-password>
  +%(master_passwd)s

NOTE There are test failures for almost all Redis versions when machine
have not small amount of CPUs:

    https://github.com/antirez/redis/issues/2715#issuecomment-151608948

Because the failure is in replication test, and so far we do not use
replication, and there is no feedback from upstream author to handle
this (for 7 days for my detailed report, and for ~ 3 months for this
issue in general), we can just disable replication test as a temporary
solution.  ( to handle remote patches with md5 hash easily the building
recipe is changed to slapos.recipe.cmmi )

NOTE Redis updated to 2.8 version because GitLab uses this series.
If/when we need more recent one we can add [redis30] in addition to
[redis28].

/reviewed-by @kazuhiko  (on !27 and on !26)
/cc @alain.takoudjou, @jerome

Ruby 2.1.7 contain security and other bugfixes:

    https://www.ruby-lang.org/en/news/2015/08/18/ruby-2-1-7-released/

/reviewed-by @kazuhiko  (on !28)

We added support for Go in the previous patch. Let's now illustrate how
to use the toolchain and build simple service based on it.

/cc @gabriel, @Camata

Add support for Go language in the form of providing component for Go
toolchain and stdlib.

Current Go stable series is 1.5.* but to bootstrap it one need to have
Go 1.4.* available, which itself needs C compiler to bootstrap.

We need Go support for upcoming GitLab SR, where one server is written
in this language:

    https://gitlab.com/gitlab-org/gitlab-workhorse

/cc @gabriel, @Camata

To illustrate how to build Ruby software, let's add helloweb-ruby
component and integrate it into helloworld SR as a service.

helloweb-ruby installation is a bit tricky, because we do not have it
initially as a released gem, and have to build it with Bundler by hand.

Changes for integration into helloworld services infrastructure are
small, thanks for the "kind" rework we did in previous patch.

/cc @kazuhiko

Prepare to have several helloweb programs each written in its own
language, and installed into bin/ as helloworld-<language> and exposed
as separate service to the web on its own port.

For component/helloweb we just rename installed script and section.

For software/helloworld we split helloweb & helloweb-promise section
into base and generate requested helloweb-<kinds> & friends via jinja2
programming; the same for exposing url for each kind.

So far it is only preparatory changes for new kinds - i.e. instance code
now supports it, but the only kind so far is still python.

/cc @jerome

Jinja2 is currently the preferred way to do instance templating, because
it has control structures.

We'll need control structures in the following patches, but even without
it, it is better to show "how to" do instances the preferred way.

/cc @jerome

This

- better illustrates what components do (provide building recipes for
  programs) and what software do (gather components and provide recipes
  for instances).

- illustrates how to build the software properly - usually we build
  eggs or scripts with zc.recipe.egg:* , not generate python programs via
  jinja2 templating. Because when installing via egg, it is possible to
  import other installed eggs relatively straghtforward, and for jinja2
  way it is hard to do.

The helloweb program itself is moved into newly introduced helloweb.git
repository and, as almost all other software, becomes separate from
slapos.git .

/cc @jerome

Compared to our currently-packaged 9.2.8 postgresql 9.2.14 contains a lot of
fixes including security ones. On the path of updating, we cleanup postgresql
component a bit. Details are in commit logs.

/reviewed-by @kazuhiko  (on !25)

Compared to 9.2.8 postgresql 9.2.14 contains a lot of fixes including
security ones:

    http://www.postgresql.org/docs/current/static/release-9-2-9.html
    http://www.postgresql.org/docs/current/static/release-9-2-10.html
    http://www.postgresql.org/docs/current/static/release-9-2-11.html
    http://www.postgresql.org/docs/current/static/release-9-2-12.html
    http://www.postgresql.org/docs/current/static/release-9-2-13.html
    http://www.postgresql.org/docs/current/static/release-9-2-14.html

Currently we have recipes to build both postgresql 9.1 and 9.2 (see
9f1f0759 "provide both postgres 9.1 and 9.2") which mostly duplicate
each other with minor difference that 9.1 is built with perl and 9.2
without (perl added to 9.1 in dbbd9a96 "Include Perl in Postgres").

To reduce the duplication let's move common compilation bits to common
section.

NOTE I've tried to add perl to postgresql 9.2 as well and got the
following compilation error:

    ld: .../perl/libs-c/libperl.a(op.o): relocation R_X86_64_32 against `.rodata.str1.8' can not be used when making a shared object; recompile with -fPIC
    .../parts/perl/libs-c/libperl.a: could not read symbols: Bad value

which happens because libperl.a is not compiled with fPIC.

For now we don't need perl in postgresql92 and this is not handled, and
we just deduplicate building recipes without any actual change for
resulting commands how the software is built.

component/postgresql/buildout.cfg defines both postgresql92 and
postgresql sections, with postgresql being just

    [postgresql]
    <= postgresql92

On the other hand, software/postgresql uses postgresql92 in parts, but
bin = ${postgresql:location}/bin in instance, which leads to _both_
postgresql and postgresql92 being built, and postgresql92 not used at all.

Let's fix it and use postgresql everywhere.

/cc @kazuhiko

This is an important update for us because we use http-send-name-header.
A bug caused HAProxy to segfault.

Since 7cd2f953 (helloworld: Turn it into real web-service) helloworld SR
implements real web-service which talks to outside world with the help
of shipped-with-it "hello web" program.

Buildout sections which create helloweb service and promise were named
based on helloworld prefix though, which is not very consistent. Thus the
change for such sections to the same name as the "hello web" program uses.

Also for style consistency the "hello web" program is renamed from
hello-web to helloweb, as this naming style is usually used for
helloworld.{c,py.rb,go,etc}

So the final naming is:

    helloweb

for both program and related sections.

Slaprunner tries to leverage multicore, and spawns multiple jobs when compiling software, based on `cpu-usage-ratio` parameter. But this currently have effect only on Make-based projects (via setting `MAKEFLAGS=-j<n>`) and does not affect software with different build systems.

Let's also provide support for parallel building for NumPy-based software and Ruby gems out of the box.

/cc @cedric.leninivin, @kazuhiko, @alain.takoudjou
/reviewed-by @jerome, @rafael  (on !22)

Most Ruby projects are built via Bundler[1], e.g. the upcoming GitLab,
and Bundler supports building/fetching dependency gems in parallel:

    http://bundler.io/v1.10/bundle_install.html#jobs

via using `--jobs <n>` command line option.

All bundler options can be also passed in via environment variable, and
this way we can use BUNDLE_JOBS=<n> for default `--jobs <n>`.

Let's use it, and this way speedup Ruby-related builds (like we already
do for Make- and NumPy- based software).

[1] http://bundler.io/

/cc @jerome, @cedric.leninivin, @kazuhiko

Starting from NumPy 1.10 numpy's distutils support parallel building

    http://docs.scipy.org/doc/numpy/user/install.html#basic-installation
    https://github.com/numpy/numpy/commit/23d54617

and this way software which uses numpy's distutils (scipy,
scikit-learn, etc) should support it too.

Let's use it, like we currently already use MAKEFLAGS for speeding up
make-based projects.

/cc @jerome, @cedric.leninivin, @kazuhiko

Currently we use '%d' and string formatting on max(1, ncpu / cpu-usage-ratio),
because `ncpu / cpu-usage-ratio` is float:

    In [1]: from __future__ import division

    In [2]: 8 / 4
    Out[2]: 2.0

and jinja2 uses future division by default:

    {{ 8 / 4 }}   ->   2.0

We can however make things more explicit, by explicitly using integer
division (// operator) and this way avoid the need for '%d' and string
formatting.

/cc @jerome, @cedric.leninivin