This is something we had read about a long time ago, but didn't find how
to reproduce the issue until now.
As mentioned in the Apartment documentation:
> it's important to consider that you may want to maintain the
> "selected" tenant through different parts of the Rack application
> stack. For example, the Devise gem adds the Warden::Manager middleware
> at the end of the stack in the examples above, our
> Apartment::Elevators::Subdomain middleware would come after it.
> Trouble is, Apartment resets the selected tenant after the request is
> finished, so some redirects (e.g. authentication) in Devise will be
> run in the context of the "public" tenant. The same issue would also
> effect a gem such as the better_errors gem which inserts a middleware
> quite early in the Rails middleware stack.
>
> To resolve this issue, consider adding the Apartment middleware at a
> location in the Rack stack that makes sense for your needs, e.g.:
>
> Rails.application.config.middleware.insert_before Warden::Manager,
> Apartment::Elevators::Subdomain
>
> Now work done in the Warden middleware is wrapped in the
> Apartment::Tenant.switch context started in the Generic elevator.
Note this only affects images which can also be customized using the
administration interface; other images like `avatar_admin.png` must be
the same for all tenants. I think this is good enough for now, since the
images that can't be different are the images that aren't customized
that often, and if there's a need to use different images in a certain
CONSUL installation, it can be achieved by changing the code which
renders that image so it uses `image_path_for`.
We were using the same logic in many different places, so we're
simplifying the code. I'm not convinced about the method names, though,
so we might change them in the future.
Note using this method for the default tenant in the `TenantDiskService`
class resulted in a `//` in the path, which is probably harmless but
very ugly and it also generates a different key than the one we got
until now. I've added an extra test to make sure that isn't the case.
We're using the "tenants" subfolder for consistency with the folder
structure we use in ActiveStorage and because some CONSUL installations
might have folders inside the `data` folder which might conflict with
the folders created by tenants.
Note that the Python scripts have a lot of duplication, meaning we need
to change all of them. I'm not refactoring them because I'm not familiar
enough with these scripts (or with Python, for that matter).
Also note that the scripts folder is still shared by all tenants,
meaning it isn't possible to have different scripts for different
tenants. I'm not sure how this situation should be handled; again, I'm
not familiar enough with this feature.
On my machine, seeding a tenant takes about one second, so skipping this
action when it isn't necessary makes tests creating tenants faster
(although creating a tenant still takes about 3-4 seconds on my
machine).
Note we don't need to update the tests; the tests themselves help us
confirm that `Rails.application.secrets` and `Tenant.current_secrets`
return the same object on single-tenant applications.
Right now this is configured using the `secrets.yml` file, which is the
file we've used in the past to configure SMTP settings.
Note that, in the `current_secrets` method, the `if default?` condition
is there so in single-tenant applications it returns the exact same
object as `Rails.application.secrets`, and it makes it immediately clear
for developers reading the code. We're also caching the tenant secrets
(using `||=`) so they behave the same way as Rails secrets; for this to
work properly 100% of the time (for example, in tests) we need to expire
these cached secrets whenever the Rails secrets change.
A similar `unless Tenant.default?` condition is present in the
ApplicationMailer because there's a chance some CONSUL installations
might not be using secrets to define the SMTP settings(they might be
using environment variables, for example) and so in this case we don't
want to force settings based on the secrets.yml file because it would
break the application.
The structure of the SMTP settings in the secrets file should be:
```
production:
tenants:
name_of_the_tenant_subdomain:
smtp_settings:
address:
(...)
```
This way it will be possible to write CSS and JavaScript code that will
only apply to specific tenants.
Note that CSS customization is still limited because it isn't possible
to use different SCSS variables per tenant.
We forgot to do so in commit d827768c0. In order to avoid the same
mistake in the future, we're extracting a method to get these
attributes. We're also adding tests, since we didn't have any tests to
check that the `dir` attribute was properly set.
While this is not strictly necessary, it can help moving the data of one
tenant to a different server or removing it.
Note we're using subfolders inside the `tenants` subfolder. If we simply
used subfolders with the schema names, if the schema names were, for
instance, language codes like `es`, `en`, `it`, ... they would conflict
with the default subfolders used by Active Storage.
Note we aren't allowing to delete a tenant because it would delete all
its data, so this action is a very dangerous one. We might need to add a
warning when creating a tenant, indicating the tenant cannot be
destroyed. We can also add an action to delete a tenant which forces the
admin to write the name of the tenant before deleting it and with a big
warning about the danger of this operation.
For now, we're letting administrators of the "main" (default) tenant to
create other tenants. However, we're only allowing to manage tenants
when the multitenancy configuration option is enabled. This way the
interface won't get in the way on single-tenant applications.
We've thought about creating a new role to manage tenants or a new URL
out of the admin area. We aren't doing so for simplicity purposes and
because we want to keep CONSUL working the same way it has for
single-tenant installations, but we might change it in the future.
There's also the fact that by default we create one user with a known
password, and if by default we create a new role and a new user to
handle tenants, the chances of people forgetting to change the password
of one of these users increases dramatically, particularly if they
aren't using multitenancy.
We had some of the logic in the ApplicationMailer. Since we're going to
use this logic in more places, we're moving it to the Tenant model,
which is the model handling everything related to hosts.
While we ping some search engines (currently, only Google) when
generating the sitemap files, we weren't telling search engines
accessing through the `robots.txt` file where to find the sitemap. Now
we're doing so, using the right sitemap file for the right tenant.
Note that the `sitemap:refresh` task only pings search engines at the
end, so it only does so for the `Sitemap.default_host` defined last. So
we're using the `sitemap:refresh:no_ping` task instead and pinging
search engines after creating each sitemap.
Note we're pinging search engines in staging and preproduction
environments. I'm leaving it that way because that's what we've done
until now, but I wonder whether we should only do so on production.
Since we're creating a new method to get the current url_options, we're
also using it in the dev_seeds.
Testing that the sitemap is valid (which we test in the following test)
also checks that the sitemap has been generated. The test will also fail
with different errors depending on whether no file was generated or the
generated file is invalid.
Some tasks don't have to run on every tenant. The task to calculate the
TSV is only done for records which were present before we added the TSV
column, and that isn't going to happen in any tenants because we added
the TSV column before adding the tenants table. Similarly, the migration
needed for existing polls isn't necessary because there weren't any
tenants before we allowed to set the starting/ending time to polls.
We aren't adding any tests for these tasks because tests for rake tasks
are slow and tests creating tenants are also slow, making the
combination of the two even slower, particularly if we add tests for
every single task we're changing. We're adding tests for the
`.run_on_each` method instead.
The `budgets📧selected` and `budgets📧unselected` tasks are
supposed to be run manually because they only make sense at a specific
point during the life of a budget.
However, they would only run on the default tenant, and it was
impossible to run them on a different tenant.
So we're introducing an argument in the rake task accepting the name of
the tenant whose users we want to send emails to.
We were using `Budget.last`, but the last budget might not be published
yet.
I must admit I don't know whether these tasks are useful, but I'm not
removing them because I'm not sure that won't harm any CONSUL
installations.
Until now, running `db:dev_seed` created development data for the
default tenant but it was impossible to create this data for other
tenants.
Now the tenant can be provided as a parameter.
Note that, in order to be able to execute this task twice while running
the tests, we need to use `load` instead of `require_relative`, since
`require_relative` doesn't load the file again if it's already loaded.
Also note that having two optional parameters in a rake task results in
a cumbersome syntax to execute it. To avoid this, we're simply removing
the `print_log` parameter, which was used mainly for the test
environment. Now we use a different logic to get the same result.
From now on it won't be possible to pass the option to avoid the log in
the development environment. I don't know a developer who's ever used
this option, though, and it can always be done using `> /dev/null`.
When running this task we truncate all tables; however, doing so doesn't
execute the `after_destroy` callback which drops the generated schema.
That meant we could run into a situation where there are schemas in the
database with no associated tenant, leading to data inconsistencies.
So we're now destroying the tenants (alongside their schemas) before
truncating the rest of the database tables.
Since records in different tenants can have the same ID, they can share
the same `cache_key`, and so we need a namespace to differentiate them.
Without them, records from one tenant could expire the cache of a record
from another tenant.
The subdomain elevator we were using, which is included in apartment,
didn't work on hosts already including a subdomain (like
demo.consul.dev, for instance). In those cases, we would manually add
the subdomain to the list of excluded subdomains. Since these subdomains
will be different for different CONSUL installations, it meant each
installation had to customize the code. Furthermore, existing
installations using subdomains would stop working.
So we're using a custom method to find the current tenant, based on the
host defined in `default_url_options`.
In order to avoid any side-effects on single-tenant applications, we're
adding a new configuration option to enable multitenancy
We're enabling two ways to handle this configuration option:
a) Change the application_custom.rb file, which is under version control
b) Change the secrets.yml file, which is not under version control
This way people prefering to handle configuration options through
version control can do so, while people who prefer handling
configuration options through te secrets.yml file can do so as well.
We're also disabling the super-annoying warnings mentioning there are no
tenants which we got every time we run migrations on single-tenant
applications. These messages will only be enabled when the multitenancy
feature is enabled too. For this reason, we're also disabling the
multitenancy feature in the development environment by default.
Note we're using the `:HOST` regular expression since subdomains can
contain the same characters as domains do. This isn't 100% precise,
though, since subdomains have a maximum length of 63 characters, but is
good enough for our purposes.
This way all tenants will be able to access them instead of just the
default one.
The apartment gem recommends using a rake task instead of a migration,
but that's a solution which is primarily meant for new installations.
Migrations are easier to execute on existing installations.
However, since this migration doesn't affect the `schema.rb` file, we
still need to make sure the shared schema is created in tasks which do
not execute migrations, like `db:schema:load` or `db:test:prepare`, just
like the apartment gem recommends. That's why we're enhancing these
tasks so they execute this migration.
Note that there might be cases where the database user isn't a superuser
(as it's usually the case on production environments), meaning commands
to create, alter or drop extensions will fail. There's also the case
where users don't have permissions to create schemas, which is needed in
order to create the shared extensions schema and the schemas used by the
tenants. For these reasons, we're minimizing the number of commands, and
so we only alter or create extensions when it is really necessary.
When users don't have permission, we aren't running the commands but
showing a warning with the steps needed to run the migration manually.
This is only necessary on installations which are going to use
multitenancy; single-tenant applications upgrading don't need to run
this migration, and that's why we aren't raising exceptions when we
can't run it.
For new installations, we'll change the CONSUL installer so extensions
are automatically created in the shared schema.
Also note the plpgsql extension is not handled here. This is a special
extension which must be installed on the pg_catalog schema, which is
always in the search path and so is shared by all tenants.
Finally, we also need to change the `database.yml` file in order to
search for shared extensions while running migrations or model tests,
since none of our enabled extensions are executed during migrations;
we're also adding a rake task for existing installations. Quoting the
apartment documentation:
> your database.yml file must mimic what you've set for your default and
> persistent schemas in Apartment. When you run migrations with Rails,
> it won't know about the extensions schema because Apartment isn't
> injected into the default connection, it's done on a per-request
> basis.
Rails 6.0 introduced a `hosts` option which, in the development
environment, defaults to all IP addresses and the `localhost` domain.
However, we can't work with subdomains using `localhost`. For that
purpose, the `lvh.me` domain was created [1].
So we're allowing this domain and its subdomains so we can use them
while working with multitenancy in the development environment.
[1] http://railscasts.com/episodes/123-subdomains-revised
We were using each one half the time, while they both had the same value
by default. It was impossible to know when me meant "use a dark color
here" or "use the secondary brand color" here.
So now we're only using one variable, so it's easy it'll be easy to add
CSS custom properties to overwrite this variable. We're choosing
`brand-secondary` because its name makes it less problematic to use a
light color.
This is consistent with the usage of `$body-background`. This way
Foundation elements using `$body-font-color`, like the `<body>` tag,
will be changed when changing this variable, which wouldn't happen when
using `$text`.
The variables `$anchor-color` and `$anchor-color-hover` are the ones
Foundation uses internally; by using them, we make sure every link will
use the colors we define.
Now we can simplify the default styles for the `<a>` tags, since by
default they already use these variables.
This way we simplify the code a bit.
Note we're only using this function when variables for background colors
are already defined, since that means customizing the variable using the
background color will automatically change the color of the text.
Customization isn't easier when using raw colors.
We were defining (for instance) white text against the `$brand`
background. That meant that, if somebody customized the `$brand` color
so it used a light color, they had to customize the text color as well
in order to guarantee proper contrast between text and background
colors.
So we're using `color-pick-contrast` instead, which means we don't have
to manually calculate whether white or black will be the color which
makes the text more readable.
We were using `#fff` in some places and `$white` in others. We're
choosing `#fff` because it has a better contrast against the lighter SDG
colors and because it's the one officially used by the United Nations.
This way we can simplify the code a bit since Foundation automatically
chooses the text color of the caption so there's enough contrast with
the background color defined in `$orbit-caption-background`.
