This patch series teaches ERP5 software release to automatically instantiate
Jupyter notebook web UI and tune it to connect to ERP5 by default. When Jupyter
is enabled, it also installs on-server erp5_data_notebook bt5 (see
nexedi/erp5!29 and nexedi/erp5@f662b5a2) which handles code execution requested
for Jupyter.

For ERP5 - for security and backward compatibility reasons - Jupyter
instantiation and erp5_data_notebook bt5 install happen only if jupyter is
explicitly enabled in instance parameters. The default is not to have Jupyter
out of the box.

On the other hand for Wendelin SR, which inherits from ERP5 SR, the
default is to have Jupyter out of the box, because Wendelin SR is fresh
enough without lots of backward compatibility needs, and Jupyter is
usually very handy for people who use Wendelin.

--------

NOTE Currently erp5-data-notebook bt5 has the following limitations (see
details on nexedi/slapos!43 and nexedi/erp5!29):

- errors are not reported properly to users;
- state is not fully saved to ZODB.

the latter point means notebook works only if it is connected to Zope family
with only 1 zope process. Hopefully this will be resolved some day.

Technical overview about how the integration is done itself on slapos part and
other notes are in 0a446263.

/proposed-for-review-on nexedi/slapos!43

Query strings used on GET requests have size limitations
on servers and this causes big code cells to not be executed
at all, returning only an Internal Server Error with no
further explanation.

/reviewed-by @kirr, @Tyagov (on nexedi/slapos!43)

Pandas and scikit-learn are handy to have for data processing which we
are going to do more and more in ERP5 context. Matplotlib is very handy
to have when one works with Jupyter, but also can be used by just
backend code to generate graphs.

From this point of view it makes sense to have this eggs always
available out of the box. To do so move them from Wendelin to ERP5.

/suggested-by @tiwariayush
/reviewed-by @Tyagov (on nexedi/slapos!43)

This patch teaches ERP5 software release to automatically instantiate Jupyter
notebook web UI and tune it to connect to ERP5 by default. When Jupyter is
enabled, it also installs on-server erp5_data_notebook bt5 (erp5!29)
which handles code execution requested for Jupyter.

For ERP5 - for security and backward compatibility reasons - Jupyter
instantiation and erp5_data_notebook bt5 install happen only if jupyter is
explicitly enabled in instance parameters. The default is not to have Jupyter
out of the box.

On the other hand for Wendelin SR, which inherits from ERP5 SR, the
default is to have Jupyter out of the box, because Wendelin SR is fresh
enough without lots of backward compatibility needs, and Jupyter is
usually very handy for people who use Wendelin.

~~~~

For integration, we reuse already established in ERP5 infrastructure, to
request various slave instances, and request Jupyter in a way so it
automatically tunes and connects to balancer of one of Zope family.

Jupyter code itself is compiled by reusing
software/ipython_notebook/software.cfg, and Jupyter instance code is
reused by hooking software/ipython_notebook/instance.cfg.in into ERP5 SR
properly (the idea to override instance-jupyter not to render into
default template.cfg is taken from previous work by @tiwariayush).

~~~~

I tested this patch inside webrunner with create-erp5-site software type and
various configurations (whether to have or not have jupyter, to which zope
family to connect it, etc).

I have not tested frontend instantiation fully - because tests were done only
in webrunner, but I've tried to make sure generated buildout code is valid for
cases with frontend.

NOTE the code in this patch depends erp5_data_notebook bt5 (erp5!29) which just got merged to erp5.git recently (see erp5@f662b5a2)

NOTE even when erp5_data_notebook bt5 is installed, on a freshly installed ERP5, it
is required to "check site consistency" first, so that initial bt5(s) are
actually installed and erp5 is ready to function.

/cc @vpelletier, @Tyagov, @klaus, @Camata, @tiwariayush, @Kreisel, @jerome, @nexedi
/proposed-for-review-on !43

Gitlab uses github-markup to render various text-based markups
(markdown, rst, ...) into html. For rst github-markup wants to run
python and have docutils egg available:

    https://github.com/github/markup/blob/5393ae93/lib/github/markups.rb#L36

as we were not having docutils installed and path to proper python
interpreter setup, rst documents were not automatically rendered and
were show just as plain text.

We do a lot of documents in rst - that case is important for us.

So fix it by providing gitlab with properly setup python interpreter
with all needed eggs installed.

/cc @kazuhiko, @jerome
/proposed-for-review-on nexedi/slapos!39

The reason is: we are going to integrate Jupyter into ERP5 SR, and in
ERP5 SR code, there is already established infrastructure, which
requests slave instances. That infrastructure passes parameters in
serialized (= json) form, that's why we need to switch.

On the other hand, slapos.cookbook:slapconfiguration.serialised handles
non-json instance parameters as well: when there parameters format is
not

    {'_' -> <json>}

it decodes them as slapos.cookbook:slapconfiguration would do:

    https://lab.nexedi.com/nexedi/slapos/blob/69229988/slapos/recipe/slapconfiguration.py#L232
    https://lab.nexedi.com/nexedi/slapos/blob/69229988/slapos/recipe/librecipe/__init__.py#L51

So from this point of view, though we don't really need it here (Jupyter
SR is relatively new), we are staying backward compatible.

/cc @tiwariayush

Both numpy and ipython are included in ERP5 SR which wendelin inherits
from, and are pinned there. Here are e.g. latest pin-ups for numpy and
ipython in erp5:

    e3144a8a    (version up eggs.)
    135570c9    (version up eggs.)

Furthermore: this is not only a cleanup. As e3144a8a shows current
version of numpy in erp5 is 1.10.4 and in wendelin we still have 1.9.2
which is unintentional downgrade compared to erp5.

Don't do that.

/cc @kazuhiko
/reviewed-by @Tyagov  (on !41)

@rafael added this in 971d0bb7 (erp5: Make possible extent the list of
initial business templates to install), but we dropped that change while
merging erp5-cluster to master - see:

    6bbb61a8 "Merge branch 'master' into erp5-cluster", and
    e84d5e83 "Merge branch 'erp5-cluster'"

6bbb61a8 claimed that it

    Dropped commit 971d0bb7
    ("erp5: Make possible extent the list of initial business templates to install").

but it actually dropped changes only under stack/erp5/ , not software/wendelin/

Fix it.

/cc @rafael, @jm, @Tyagov
/reviewed-by TrustMe

In slapos.package.git/obs, we need to build binaries for specific paths,
without changing where buildout actually install them.

Hello up there,

Here comes SlapOS port of GitLab. We start from GitLab 8.2.X as that is what we
currently run on KVM on lab.nexedi.com, so that our data can be
straightforwardly migrated.

The SR compiles all needed software and organized all (sub-)services in one
partition and interconnects them with unix sockets for security and speed
reasons (see patch "gitlab: Make a plan to base instance layout on
gitlab-omnibus and to interconnect all internal services").

Services configuration files are originally taken from omnibus-gitlab
"distribution" and incrementally ported to slapos variant. This way we
establish a (imho) good path on how to track upstream changes and minimize our
delta & effort supporting it.

GitLab itself is patched (above patches that were already applied by upstream):

  - to support HTTP(S) only
  - to show site's ICP number
  - to speedup raw blob downloading ~ 17x times
    ( see patch "gitlab: Optimize raw blob downloading" for details and
      https://gitlab.com/gitlab-org/gitlab-workhorse/merge_requests/17 )

Overall it should work and we should finally be able to migrate slapos.git
(because of raw blob downloading is not slow now) to GitLab and all other
Nexedi git repositories.

Thanks,
Kirill

P.S.

Somewhat outdated, but this picture on GitLab architecture might help to
understand how parts are glued together:

https://gitlab.com/gitlab-org/gitlab-ce/blob/master/doc/development/architecture.md

P.P.S.

Native resiliency is not implemented yet, but we should be able to use gitlab
inside resilient webrunner already.

/proposed-for-review-on !39
/partly-reviewed-by @kazuhiko, @jerome, @Yanni, @jp
/cc @rafael, @jm

We've reached a state where first gitlab SR version should work. So as
promised let's freeze the md5 checksums.

All later patches should update corresponding md5 info when they change
a file.

/cc @kazuhiko, @jerome

In slapos we do a lot of automated software rebuild constantly, and thus
there is constant flow of requests to get raw blobs from git service,
e.g. like this

    https://lab.nexedi.com/nexedi/slapos/raw/master/software/wendelin/software.cfg

A lot of requests comes to slapos.git repository and currently gitlab,
out of the box, cannot keep up with that load.

I've prepared patches to offload raw blobs download requests handling
from unicorn (ruby) to gitlab-workhorse (go), and that resulted in ~ 17x
speedup - e.g. previously our std shuttle can handle ~ 70 raw-blob
requests/s and with my changes it is now ~ 1200 requests/s.

The patches were sent upstream

    https://gitlab.com/gitlab-org/gitlab-workhorse/merge_requests/17

and we discussed with GitLab people and made a plan how to proceed
incrementally. It will probably take some time for gitlab team to fully
accept the approach though.

For now we can use our gitlab-workhorse fork. The patches itself are:

    kirr/gitlab-workhorse@1b274d0d
    kirr/gitlab-workhorse@2beb8c95

/cc @kazuhiko, @jerome, @jm

GitLab Nexedi Edition is currently upstream 8.2.X + the following
patches:

    - HTTP(S) is made to be default clone protocol

        kirr/gitlab-ce@5c1f2fb3

      and SSH info is completely removed from UI

        kirr/gitlab-ce@dfe9fb16
        kirr/gitlab-ce@f3f84743

      so essentially the only way to access a repository is via HTTP(S).

    - Rake check tasks are adjusted to exit with non-zero code if there
      is a failure

        kirr/gitlab-ce@a93ae418

      We need this for promises to work correctly with failures being
      detected, not silently skipped. The patch was sent upstream:

        https://gitlab.com/gitlab-org/gitlab-ce/merge_requests/1885

    - GitLab supports setting up site's ICP License in gitlab.yml and
      shows it in appropriate places together with info about GitLab
      itself:

        kirr/gitlab-ce@e7e0fd88
        kirr/gitlab-ce@79c127e6

    + other cosmetic/minor changes.

More patches will probably come (e.g. apply a single patch from a
merge-request with `git am` without creating merge commit for just 1
patch, etc) but for now that's all.

NOTE ICP is non-ascii text with hieroglyphs. slapos.core was taught to
    be able to pass parameters with non-ascii values to instance:

        slapos.core@347d33d6

    That patch is included in slapos.core 1.3.15, but as we currently
    have a lot of older slapos.core deployed (e.g. 1.3.5 on my
    development webrunner) a workaround is (hopefully temporarily) used
    to pass non-ascii values as URL-encoded strings.

/cc @kazuhiko, @jerome, @rafael

In the previous patch we setup nginx service which listens to the world
and as such gitlab service becomes to be ready to used - so publish
backend URL.

NOTE we'll need to optimise and tweak gitlab a bit further in upcoming
    patches, so it can be really used under load and with our use-cases,
    but even now it listens to http ok and generally works.

/cc @kazuhiko, @jerome

Go through nginx configuration templates and convert them to jinja2 with
slapos parameters (reminder: names and default values are imported from
omnibus-gitlab 8.2.3+ce.0-0-g8eda093), except commenting out features we
do not want to support (yet ?).

As nginx is a reverse-proxy, i.e. it integrates all internal services
and works as frontend to them, our gitlab service is now ready to listen
and talk to the world over (standard to slapos services backend) IPv6.

Nginx also acts as SSL termination point - for it to work by default we
setup self-signed certificate for the backend, which can be manually
changed to proper certificate if needed. Backend certificate is used
if gitlab is configured to work in HTTPS mode (and frontend certificate
is another story).

NOTE ssl certificate is generated with just `openssl req ...` - yes, there
    is slapos.cookbook:certificate_authority.request but it requires
    to start whole service and has up to 60 seconds latency to generate
    certificate. And we only need to run 1 command to do that...

The features disabled are:

    - http -> https redirection

      not needed for us at nginx level - the frontend can do the
      redirection and also gitlab speaks HSTS on https port so when we access
      https port via http protocol, it gets redirected to https.

    - kerberos
    - ssl_dhparam
    - providing custom nginx configuration via instance parameter

/cc @kazuhiko, @jerome

Like with Rails configuration files, hook nginx configuration files into
SR / instance build process; rename *.erb -> *.in and add our header.

The templates are still not valid - a lot of erb code is left there -
we'll slapos'ify it incrementally in the following patches.

/cc @kazuhiko, @jerome

Like with Rails configuration this first step is pristine import of
nginx configuration files from omnibus-gitlab. All files were imported
as-is in their ERB form and filenames from omnibus-gitlab
8.2.3+ce.0-0-g8eda093 from here:

    https://gitlab.com/gitlab-org/omnibus-gitlab/tree/8eda093/files/gitlab-cookbooks/gitlab/templates/default

We import only nginx main http configuration - nginx's CI and Mattermost
configurations are not imported, as we do not support CI and Mattermost (yet ?).

As with Rails configuration files, we will convert the templates to
jinja2 and adjust them to slapos version in the following patches.  We
will also use the same (commit from last-erb-mod commit + merge)
approach to track upstream changes.

/cc @kazuhiko, @jerome

Sidekiq[1] is used in GitLab as background jobs manager - i.e. if a
request handler needs to spawn some non-light job - it adds it to
sidekiq queue (in Redis) and relies on sidekiq service to later pick
this job up and execute it.

The service is setup with just to run bin/gitlab-sidekiq with
appropriate queues (extracted from omnibus-gitlab) and appropriate
settings to controlling GitLab's sidekiq Out-Of-Memory killer[2].

NOTE Unlike unicorn OOM killer, Sidekiq memory killer just makes sidekiq
    processes to be SIGKILL terminated and relies on managing service to
    restart it. In slapos we don't have mechanism to set autorestart=true,
    nor bang/watchdog currently work with slapproxy, so we setup to do
    such monitoring ourselves manually with here-introduced
    watcher-sigkill program.

NOTE2 sidekiq promise, because it is rake/gitlab based, is slow to
    load/run and thus is put into etc/promise.slow/

[1] http://sidekiq.org/
[2] https://gitlab.com/gitlab-org/gitlab-ce/blob/1322bd78/doc/operations/sidekiq_memory_killer.md

/cc @kazuhiko, @jerome

As was described in the previous patch, we need $HOME to be propagated
by this programs so that git can find partition's .gitconfig.

Specifically we need the following patches to be present in our build:

    https://gitlab.com/gitlab-org/gitlab-shell/commit/9e087f64
    https://gitlab.com/gitlab-org/gitlab-workhorse/commit/b5f1b803

They both have been applied upstream very close to revisions we
previously had in software.cfg, so we only need to update the revisions
to get them.

/cc @kazuhiko, @jerome

Convert gitconfig template to jinja2 (reusing already-there
`email_display_name` and `email_from` parameters for commits generated by
gitlab).

System-level git config from gitlab-omnibus is also imported to this
file (on slapos we cannot tweak system-level git config -
software/.../parts/git/... is read-only for programs in instance
partitions - so we move all gitlab's system-wide git settings to this
"user-level" gitconfig.

System gitconfig in omnibus is defined here:

    https://gitlab.com/gitlab-org/omnibus-gitlab/blob/8eda093/files/gitlab-cookbooks/gitlab/attributes/default.rb#L23

so it is

    pack.threads = 1    and
    receive.fsckObjects = true

which makes sense to not waste a lot of memory when packing and not to
allow corrupt objects to enter to system by evil users intentionally.

To make the file foundable by git - we put it into partition root
directory and set $HOME to point to partition root when running
appropriate programs / services.

NOTE we'll need to upgrade gitlab-shell and gitlab-workhorse to
    propagate $HOME for this setting to actually have effect.
    See the next patch.

/cc @kazuhiko, @jerome

Like with Rails configuration files, this is pristine import of template
gitconfig from omnibus GitLab from

    https://gitlab.com/gitlab-org/omnibus-gitlab/tree/8eda093/files/gitlab-cookbooks/gitlab/templates/default

This is only a "user" part of git configuration. System-wide
configuration is generated dynamically:

    https://gitlab.com/gitlab-org/omnibus-gitlab/blob/8eda093/files/gitlab-cookbooks/gitlab/attributes/default.rb#L23

and we'll import it by hand in the follow-up patches.

/cc @kazuhiko, @jerome

Gitlab-workhorse[1] is a service which offloads Ruby-on-Rails based
GitLab from long-running and slow requests. It is written in Go.

Now as we have unicorn service set up, we can setup gitlab-workhorse
service (which uses unicorn as authentication backend).

Gitlab-workhorse setup is easy - it is just one program and several
command line options to point to unicorn socket and to configure on
which unix socket gitlab-workhorse will listen itself.

NOTE we have to care that git and ruby to be on PATH when running
    gitlab-workhorse - because on e.g. git push'ing workhorse will run `git
    receive-pack` and a hook will be called which calls gitlab-shell,
    which is written in ruby.

NOTE2 promise to check whether gitlab-workhorse is alive is to ping it
    via URL to non-existent endpoint and check for proper 403 HTTP code
    returned.

[1] https://gitlab.com/gitlab-org/gitlab-workhorse

/cc @kazuhiko, @jerome

gitlab: Compile assets on instantiation and make sure DB is properly setup/migrated before unicorn runs

There are several actions that needs to be done on gitlab instance
upgrade:

    - we have to (re-)compile assets
    - we have to migrate DB

and also before the first run

    - we have to initialize DB

We can compile assets as part of instantiation process, but regarding
DB migration / setup - it is not currently possible to do that as part
of instantiation - for that operations we need PG & Redis to be already
running, but the first time slapos instantiates an SR it first prepares
all services, and only after instantiation is done, starts them all. There is
currently no way to hook into starting process, and run some scripts
after one service is started but before another service startup...

So the solution is: to perform such actions in delayed mode as part of
application - unicorn service - startup: it makes sure PG is running and
initializes it and does other actions which needs to be done to migrate
the DB. Only if/after they succeed the main application is started.

NOTE the comment about unicorn/gitlab startup slowness from the previous
    patch still holds true - so in order to get "all ok" after
    instantiation, it is required to perform the instantiation several
    times, because unicorn promise initially fails.

/cc @kazuhiko, @jerome

Now that all gitlab Rails application configuration files are ready, we
can setup unicorn service to start it.

NOTE there is a promise to check unicorn by url which works, but there
    are also rake tasks to check gitlab itself, e.g. like

        gitlab:app:check
        gitlab:gitlab_shell:check
        gitlab:repo:check

    Unfortunately this tasks are slow to run (and gitlab:repo:check is
    very slow to run). That's why we do not put them into etc/promise/ -
    if we do - slapos reports promises time outs.

    What we do is we put them into etc/promise.slow/ so we have those
    scripts ready, but currently no one automatically checks them.

    Again, the promise to check unicorn just by accessing it by URL is
    there and is checked automatically out of the box.

NOTE2 GitLab is very slow to load. That's why it can take some time
    after unicorn starts that it's promise start to report ok. This can
    show itself as temporary instantiation errors which say promise such and
    such failed.

NOTE3 Unicorn start, but so far we did not cared to setup GitLab DB
    schema on instantiation. That's why unicorn remains not very usable
    and a lot of requests fail. We'll teach instance to setup DB and
    perform all other needed settings in the next patch.

/cc @kazuhiko, @jerome

Like with [promise-wrapper] a recipe could do

    [promise-<service>]
    <= promise-byurl
    url     = ...

and a script to check such ur will be generated and automatically put
into etc/promise/<service>.

/cc @kazuhiko, @jerome

Automatically configure unicorn to spawn as much worker processes as
there are CPUs on the system by default.

GitLab omnibus pre-hardcodes this value default to 2 (which we copied)
and then also tweaks it this way in active code

    https://gitlab.com/gitlab-org/omnibus-gitlab/blob/8eda0933/files/gitlab-cookbooks/gitlab/attributes/default.rb#L230

which we also do here.

/cc @kazuhiko, @jerome

Explicitly point gitlab-shell to location where we keep secrets.

We already pointeg gitlab to that place and now we do that for
gitlab-shell so those 2 peieces can connect to each other ok.

Regarding the setting itself - there is no such block in omnibus-gitlab,
but it is present in gitlab-shell configuration example:

    https://gitlab.com/gitlab-org/gitlab-shell/blob/82b3a4e8/config.yml.example#L35

/cc @kazuhiko, @jerome

Convert gitlab-shell configuration file to slapos:

    - convert to jinja2,
    - connect gitlab-shell to unicorn & redis unix sockets

NOTE

    - http_settings are left to be default (empty) ones - as that works ok.
    - `auth_file` is still configured to point to wont-be-used sshkeys
      file, as without it gitlab-shell check will fail.
    - support for audit_usernames and git_annex is disabled and
      remains not configurable.

/cc @kazuhiko, @jerome