closes https://github.com/TryGhost/Ghost/issues/12791
closes https://github.com/TryGhost/Team/issues/566https://github.com/TryGhost/Ghost/pull/12787 introduced a significant performance regression due to a misunderstanding of when Bookshelf calls `.format()` ([related upstream issue](https://github.com/bookshelf/bookshelf/issues/668)). We expected `.format()` to only be called on save but it's also called when Bookshelf performs fetching and eager loading which happens frequently. `.format()` can be a heavy method as it needs to parse and serialize html and markdown so it should be performed as infrequently as possible.
- override `sync()` in the base model so we can call our own `.formatOnWrite()` method to transform attributes on `update` and `insert` operations
- this was the only feasible location in Bookshelf I could find that is low enough level to not require modifying model instance attributes
- gives models the option to perform heavy transform operations only when writing to the database compared to the usual `.format()` method that is also called on fetch in many situations
refs https://github.com/TryGhost/Ghost/pull/12736
refs https://github.com/TryGhost/Team/issues/467
knex's `parse()` method is only called on data when directly fetched from the db. This was causing problems when model instances are passed around via events for example because `.get('key')` will return data that was directly set on the model without having gone through the `parse()` transformations. The result of this inconsistency was settings appearing correct when Ghost started up but then being broken as soon as a setting was changed.
- moved absolute/relative->transform-ready URL transformations from the API input serializers to the model's `format()` method and replaced with a relative->absolute transform in API input serializers
- results in consistency because `.get()` on a settings model will always return an URL
- removed transform-ready->absolute transforms from the API output serializers as that is now handled at the model-layer
refs https://github.com/TryGhost/Team/issues/467
refs https://github.com/TryGhost/Ghost/pull/12731
- settings are mostly fetched directly from the settings cache rather than via the API so they aren't subject to the API-level output serializers that transform URLs meaning that URLs in the front-end ended up with raw `__GHOST_URL__` replacement strings
- added images to the Settings model's `parse()` method so they are transformed immediately when fetching from the database
no issue
- when saving Stripe keys with `stripeDirect: true` config, if either key didn't match the key format the returned validation error always contained `stripe_secret_key`
- updated to output `stripe_publishable_key` if it was the publishable key that was invalid
no-issue
This removes all references to the members labs setting, any code that was run conditionally behind this flag now runs unconditionally.
* Removed usage of Members labs flag
* Removed tests for Members disabled
* Added dynamic keypair generation for when setting is missing
closes#12181
refs https://github.com/TryGhost/Ghost/pull/12265
- The 'core' value is invalid for settings 'type' column
- The 'core' default value for 'type' column in model is also invalid
- Both need to be removed as they are never used and only introduce confusion into the codebase
closes#12049
Stripe plans used to default to 0, and our new validation of plan
amounts were causing issues when importing from an older version of
Ghost, this updates the validation to be skipped when importing.
- Added regression test for importing plans
closes#12001
* Moved settings validation to the model
This moves the settings validation out of the validation file and into
the model, as it is _only_ used there.
It also sets us up in the future for custom validators on individual
settings.
* Improved validation of stripe_plans setting
- Checks `interval` is a valid string
- Checks `name` & `currency` are strings
* Moved stripe key validation into model
The stripe key settings are all nullable and the regex validation fails
when the input is `null`. Rather than reworking the entirety of how we
validate with default-settings validation objects, this moves the
validation into methods on the Settings model.
* Added tests for new setting validations
Adds tests for both valid and invalid settings, as well as helpers
making future tests easier and less repetitive
refs https://github.com/TryGhost/Ghost/issues/10318
- Updates `boolean` serialization in v2/canary serializers to apply only for `boolean` type settings
- Updates `boolean` transformation in model layer `format`/`parse` to check on `boolean` type setting
- Removes error thrown on Read-only setting for settings edit endpoint
- Updates v2/canary input serializers to remove any Read-only settings (using RO flag) to avoid edits
- Added type/group mappings in the importer when pre-migration settings table import data is present
- Updates tests
refs https://github.com/TryGhost/Ghost/issues/10318
- Updates default settings to contain correct type and validation for each setting
- Updates `populateDefaults` to correctly insert type value for new settings
- Updates settings schema to allow only select types - `array`, `number`, `boolean`, `string`
- `object` is a temporary type allowed till we get rid of all JSON object settings
refs https://github.com/TryGhost/Ghost/issues/10318
`Settings.populateDefaults()` is run before migrations during Ghost's startup. This can cause problems when new settings table columns are added (and populated in `default-settings.json`) because `populateDefaults()` was using the model layer which assumes that those columns are available, resulting in `ER_BAD_FIELD_ERROR: Unknown column` type errors.
- query the database for the available `settings` table columns
- switch to using raw knex queries without Bookshelf for insertions so that we're in control of the columns that are added
- use `_.pick` to skip any properties in `default-settings.json` that do not match to an available column - those columns will be added and populated by later migrations
- moving away from using the model to insert settings has the side-effect of not emitting `settings.added/edited` and `settings.x.added/edited` events, this should be fine because `populateDefaults()` is called before anything else is set up and listening
- added a call to `populateDefaults()` in our knex-migrator "before migration" hook so that we have consistent db state across both startup initialised migrations and manually triggered knex migrations
refs https://github.com/TryGhost/Ghost/issues/10318
- `group`
- to replace the `type` column, provides a more descriptive name for the columns use
- for existing sites it will be populated by migrating data from the `type` column in a later migration
- for new sites a minimal update has been added to `parseDefaultSettings()` to populate the `group` field when settings are created during startup - fixes the NOT NULL constraint on `settings.group`
- `flags`
- signifies special handling that is different to other settings in a group
- eg, `PUBLIC,RO` would indicate that the setting is available via unauthenticated endpoints and is read-only
refs #10318
refs https://github.com/TryGhost/Ghost/pull/11942
- Removes force_i18n, permalinks, and members_session_secret usage from the codebase
- We deprecated these flasgs and have not used since Ghost v2. It's good time to remove them before we introduce bigger changes to how `settings` table opeartes.
- Fixed importer test. The test was meant to check if string values were converted properly, the check agains boolean didn't make much sense in this context, so removed it.
- Following this change are going to come ralated migrations to fix existing data (see ref)
* refactored `core/frontend/apps` to destructure common imports
* refactored `core/frontend/services/{apps, redirects, routing}` to destructure common imports
* refactored `core/frontend/services/settings` to destructure common imports
* refactored remaining `core/frontend/services` to destructure common imports
* refactored `core/server/adapters` to destructure common imports
* refactored `core/server/data/{db, exporter, schema, validation}` to destructure common imports
* refactored `core/server/data/importer` to destructure common imports
* refactored `core/server/models/{base, plugins, relations}` to destructure common imports
* refactored remaining `core/server/models` to destructure common imports
* refactored `core/server/api/canary/utils/serializers/output` to destructure common imports
* refactored remaining `core/server/api/canary/utils` to destructure common imports
* refactored remaining `core/server/api/canary` to destructure common imports
* refactored `core/server/api/shared` to destructure common imports
* refactored `core/server/api/v2/utils` to destructure common imports
* refactored remaining `core/server/api/v2` to destructure common imports
* refactored `core/frontend/meta` to destructure common imports
* fixed some tests referencing `common.errors` instead of `@tryghost/errors`
- Not all of them need to be updated; only updating the ones that are
causing failures
* fixed errors import being shadowed by local scope
- All var declarations are now const or let as per ES6
- All comma-separated lists / chained declarations are now one declaration per line
- This is for clarity/readability but also made running the var-to-const/let switch smoother
- ESLint rules updated to match
How this was done:
- npm install -g jscodeshift
- git clone https://github.com/cpojer/js-codemod.git
- git clone git@github.com:TryGhost/Ghost.git shallow-ghost
- cd shallow-ghost
- jscodeshift -t ../js-codemod/transforms/unchain-variables.js . -v=2
- jscodeshift -t ../js-codemod/transforms/no-vars.js . -v=2
- yarn
- yarn test
- yarn lint / fix various lint errors (almost all indent) by opening files and saving in vscode
- grunt test-regression
- sorted!
- Apps are marked as removed in 3.0, never officially launched and have been deprecated for at least 2 years.
- We've slowly removed bits that got in our way or were insecure over time meaning they mostly didn't work
- This cleans up the remainder of the logic
- The tables should be cleaned up in a future major
- Apps are marked as removed in 3.0, never officially launched and have been deprecated for at least 2 years.
- We've slowly removed bits that got in our way or were insecure over time meaning they mostly didn't work
- This cleans up the remainder of the logic
- The tables should be cleaned up in a future major
no-issue
This adds two new endpoints, one at /ghost/.well-known/jwks.json for exposing
a public key, and one on the canary api /identities, which allows the
Owner user to fetch a JWT.
This token can then be used by external services to verify the domain
* Added ghost_{public,private}_key settings
This key can be used for generating tokens for communicating with
external services on behalf of Ghost
* Added .well-known directory to /ghost/.well-known
We add a jwks.json file to the .well-known directory which exposes a
public JWK which can be used to verify the signatures of JWT's created
by Ghost
This is added to the /ghost/ path so that it can live on the admin
domain, rather than the frontend. This is because most of its
uses/functions will be in relation to the admin domain.
* Improved settings model tests
This removes hardcoded positions in favour of testing that a particular
event wasn't emitted which is less brittle and more precise about what's
being tested
* Fixed parent app unit tests for well-known
This updates the parent app unit tests to check that the well-known
route is mounted. We all change proxyquire to use `noCallThru` which
ensures that the ubderlying modules are not required. This stops the
initialisation logic in ./well-known erroring in tests
https://github.com/thlorenz/proxyquire/issues/215
* Moved jwt signature to a separate 'token' propery
This structure corresponds to other resources and allows to exptend with
additional properties in future if needed
no issue
- Added test cases to check edit permission on settings endpoints
- Added test to demonstrate owner-only being able to toggle members flag
- Permission check when editing settings `lab.members`
- Passed additional function to permissions to allow custom selection of unsafe attributes due to settings object structure.
- Fully implementing this check on controller level would be wrong architecturally and not that straight forward because we lack role data in "frame"
- Cleaned up test after moving default_content_visibility to it's own property
no issue
- Updated Test & linting packages
- Updated use of hasOwnProperty
- Using Object.prototype.hasOwnProperty instead (ref. eslint.org/docs/rules/no-prototype-builtins)
- Removed already defined built-in global variable Intl
- Applied `--fix` with lint command on `core/test` folder
- The rules were broken because some of them were made stricter for `eslint: recommended` ruleset (ref. https://eslint.org/docs/user-guide/migrating-to-6.0.0#eslint-recommended-changes)
- Removed redundant global variable declarations to pass linting
closes#10789
* Updated keypair generation to use a memoised fn
This allows us to embed the members dynamic defaults in the object at
definition, and will allow us to only create the keypair when we need
it, in future.
* Added getDefaultValue fn to default setting obj
This will allow us to generate the default values when they're needed
rather than at boot time.
* Ensured dynamic defaults only generated when used
This replaces all the dynamic default values with functions to return
the values, and then calls (if required) that function inside the
getDefaultValue method of the setting object.
This reverts commit 64735693be.
- `rsa-keypair` is a binary dependency that was failing to install for a lot of users, reverting for now so we can look at alternative options for speeding up boot time
* Removed support for cookies in members auth middleware
no-issue
The members middleware will no longer be supporting cookies, the cookie
will be handled by a new middleware specific for serverside rendering,
more informations can be found here:
https://paper.dropbox.com/doc/Members-Auth-II-4WP4vF6coMqDYbSMIajo5
* Removed members auth middleware from site app
no-issue
The site app no longer needs the members auth middleware as it doesn't
support cookies, and will be replaced by ssr specific middleware.
https://paper.dropbox.com/doc/Members-Auth-II-4WP4vF6coMqDYbSMIajo5
* Added comment for session_secret setting
no-issue
We are going to have multiple concepts of sessions, so adding a comment
here to be specific that this is for the Ghost Admin client
* Added theme_session_secret setting dynamic default
no-issue
Sessions for the theme layer will be signed, so we generate a random hex
string to use as a signing key
* Added getPublicConfig method
* Replaced export of httpHandler with POJO apiInstance
no-issue
This is mainly to reduce the public api, so it's easier to document.
* Renamed memberUserObject -> members
no-issue
Simplifies the interface, and is more inline with what we would want to export as an api library.
* Removed use of require options inside members
no-issue
This was too tight of a coupling between Ghost and Members
* Simplified apiInstance definition
no-issue
* Added getMember method to members api
* Added MembersSSR instance to members service
* Wired up routes for members ssr
* Updated members auth middleware to use getPublicConfig
* Removed publicKey static export from members service
* Used real session secret
no-issue
* Added DELETE /members/ssr handler
no-issue
This allows users to log out of the theme layer
* Fixed missing code property
no-issue
Ignition uses the statusCode property to forward status codes to call sites
* Removed superfluous error middleware
no-issue
Before we used generic JWT middleware which would reject, now the
middleware catches it's own error and doesn't error, thus this
middleware is unecessary.
* Removed console.logs
no-issue
* Updated token expirty to hardcoded 20 minutes
no-issue
This returns to our previous state of using short lived tokens, both for
security and simplicity.
* Removed hardcoded default member settings
no-issue
This is no longer needed, as defaults are in default-settings.json
* Removed stripe from default payment processor
no-issue
* Exported `getSiteUrl` method from url utils
no-issue
This keeps inline with newer naming conventions
* Updated how audience access control works
no-issue
Rather than being passed a function, members api now receives an object
which describes which origins have access to which audiences, and how
long those tokens should be allowed to work for. It also allows syntax
for default tokens where audience === origin requesting it. This can be
set to undefined or null to disable this functionality.
{
"http://site.com": {
"http://site.com": {
tokenLength: '5m'
},
"http://othersite.com": {
tokenLength: '1h'
}
},
"*": {
tokenLength: '30m'
}
}
* Updated members service to use access control feature
no-issue
This also cleans up a lot of unecessary variable definitions, and some
other minor cleanups.
* Added status code to auth pages html response
no-issue
This was missing, probably default but better to be explicit
* Updated gateway to have membersApiUrl from config
no-issue
Previously we were parsing the url, this was not very safe as we can
have Ghost hosted on a subdomain, and this would have failed.
* Added issuer to public config for members
no-issue
This can be used to request SSR tokens in the client
* Fixed path for gateway bundle
no-issue
* Updated settings model tests
no-issue
* Revert "Removed stripe from default payment processor"
This reverts commit 1d88d9b6d73a10091070bcc1b7f5779d071c7845.
* Revert "Removed hardcoded default member settings"
This reverts commit 9d899048ba7d4b272b9ac65a95a52af66b30914a.
* Installed @tryghost/members-ssr
* Fixed tests for settings model
refs #10582
- I don't think this is a good idea
- If a user passses "null", we should treat it as a string
- I am not aware of a use case why people have "null" in their database
- If people send "null" via the API, we should respect this and accept a string
no issue
- Reported here: https://forum.ghost.org/t/in-version-2-16-3-found-bug/6065/3
---
Admin Client sends false or true booleans for `is_private` key.
The settings table has two columns "key" and "value". And "value" is always type TEXT.
If you pass value=false, the db will transform this value into "0".
`settingsCache.get('is_private')` is then always true, even though the value is meant to be false.
We should add a migration in v3 and normalize all setting values to ensure consistent database values. Furthermore, we should improve the handling around settings values in general.
For now, we protect parsing values from DB, which we anyway need to transform the values into the correct data type, because we always save strings. This will protect values being stored as "false" or "1" or whatever.
no issue
- the event chain works like this:
- if a model registers an event, it get's triggered, because it's stronger than the base model
- but you have to call the base model to agree on a contract, because base model implements generic logic in event handlers
- this was inconsistently used
no issue
- if multiple queries run in a transaction, the model events are triggered before the txn finished
- if the txn rolls back, the events are anyway emitted
- the events are triggered too early
- solution:
- `emitChange` needs to detect that a transaction is happening
- it listens on a txn event to determine if events should be triggered
refs https://github.com/TryGhost/Ghost/issues/3658
- the `validateSchema` helper was a bit broken
- if you add a user without email, you will receive a database error
- but the validation error should catch that email is passed with null
- it was broken, because:
- A: it called `toJSON` -> this can remove properties from the output (e.g. password)
- B: we only validated fields, which were part of the JSON data (model.hasOwnProperty)
- we now differentiate between schema validation for update and insert
- fixed one broken import test
- if you import a post without a status, it should not error
- it falls back to the default value
- removed user model `onValidate`
- the user model added a custom implementation of `onValidate`, because of a bug which we experienced (see https://github.com/TryGhost/Ghost/issues/3638)
- with the refactoring this is no longer required - we only validate fields which have changed when updating resources
- also, removed extra safe catch when logging in (no longer needed - unit tested)
- add lot's of unit tests to proof the code change
- always call the base class, except you have a good reason
no issue
- this commit cleans up the usages of `include` and `withRelated`.
### API layer (`include`)
- as request parameter e.g. `?include=roles,tags`
- as theme API parameter e.g. `{{get .... include="author"}}`
- as internal API access e.g. `api.posts.browse({include: 'author,tags'})`
- the `include` notation is more readable than `withRelated`
- and it allows us to use a different easier format (comma separated list)
- the API utility transforms these more readable properties into model style (or into Ghost style)
### Model access (`withRelated`)
- e.g. `models.Post.findPage({withRelated: ['tags']})`
- driven by bookshelf
---
Commits explained.
* Reorder the usage of `convertOptions`
- 1. validation
- 2. options convertion
- 3. permissions
- the reason is simple, the permission layer access the model layer
- we have to prepare the options before talking to the model layer
- added `convertOptions` where it was missed (not required, but for consistency reasons)
* Use `withRelated` when accessing the model layer and use `include` when accessing the API layer
* Change `convertOptions` API utiliy
- API Usage
- ghost.api(..., {include: 'tags,authors'})
- `include` should only be used when calling the API (either via request or via manual usage)
- `include` is only for readability and easier format
- Ghost (Model Layer Usage)
- models.Post.findOne(..., {withRelated: ['tags', 'authors']})
- should only use `withRelated`
- model layer cannot read 'tags,authors`
- model layer has no idea what `include` means, speaks a different language
- `withRelated` is bookshelf
- internal usage
* include-count plugin: use `withRelated` instead of `include`
- imagine you outsource this plugin to git and publish it to npm
- `include` is an unknown option in bookshelf
* Updated `permittedOptions` in base model
- `include` is no longer a known option
* Remove all occurances of `include` in the model layer
* Extend `filterOptions` base function
- this function should be called as first action
- we clone the unfiltered options
- check if you are using `include` (this is a protection which could help us in the beginning)
- check for permitted and (later on default `withRelated`) options
- the usage is coming in next commit
* Ensure we call `filterOptions` as first action
- use `ghostBookshelf.Model.filterOptions` as first action
- consistent naming pattern for incoming options: `unfilteredOptions`
- re-added allowed options for `toJSON`
- one unsolved architecture problem:
- if you override a function e.g. `edit`
- then you should call `filterOptions` as first action
- the base implementation of e.g. `edit` will call it again
- future improvement
* Removed `findOne` from Invite model
- no longer needed, the base implementation is the same
refs #9192
- Each setting is saved individually
- Update this to only happen on import, or when a value changes
- Reduces the amount of work Ghost does on every setting change
refs #9001
When a blog is in private mode there is now an unguessable URL that allows access to the RSS feed for internal use, commenting systems, etc.
- add public hash for private blogging
- auto generate on bootstrap if missing
- global hash, we can re-use in the future
- update private blogging middleware to detect the private RSS URL and rewrite it so that the normal rss route/code is used for display
- if a normal `/rss/` route is accessed with a private session return a 404
no issue
- if you upload a huge import file, parallel operations can throw errors e.g. lock wait exceeds
- this can happen if multiple transactions run in parallel
- there is no need to run:
1. the removal of active tokens on import, because imported users have no active session
2. rescheduling logic on timezone, because importing scheduled posts works out of the box via the model layer (if a published date is detected and it's in the future, the post get's scheduled)
refs #8093✨ Add activate theme permission
- add permission to activate themes
- update tests
- also: update tests for invites
TODO: change how the active theme setting is updated to reduce extra permissions
✨ Move theme validation to gscan
- add a new gscan validation method and use it for upload
- update activate endpoint to do validation also using gscan
- change to using SettingsModel instead of API so that we don't call validation or permissions on the settings API
- remove validation from the settings model
- remove the old validation function
- add new invalid theme message to translations & remove a bunch of theme validation related unused keys
📖 Planned changes
🚨 Tests for theme activation API endpoint
🐛 Don't allow deleting the active theme
🚫 Prevent activeTheme being set via settings API
- We want to control how this happens in future.
- We still want to store the information in settings, via the model.
- We just don't want to be able to change this info via the settings edit endpoint
🐛✨ Fix warnings for uploads & add for activations
- warnings for uploads were broken in f8b498d
- fix the response + adds tests to cover that warnings are correctly returned
- add the same response to activations + more tests
- activations now return a single theme object - the theme that was activated + any warnings
🎨 Improve how we generate theme API responses
- remove the requirement to pass in the active theme!
- move this to a specialist function, away from the list
🎨 Do not load gscan on boot
no issue
🔥 Remove unnecessary cache update
🎨 simplify updateSettingsCache()
🎨 Simplify readSettingsResult
- although this is more code, it's now much clearer what happens in the two cases
🎨 Don't use readSettingResult for edit
🎨 Simplify updateSettingsCache further
🔥 Remove now unused readSettingsResult
🎨 Change populateDefault to return all
🎨 Move the findAll call out of updateSettingsCache
🔥 Remove updateSettingsCache!!
🎨 Restructure init & finish up settingsCache
- move initialisation into settingsCache.init AT LAST
- change settingCache to use cloneDeep, so that the object can't be modified outside of the functions
- add lots of docs to settings cache
🎨 Cleanup db api endpoints
🔥 Don't populate settings in migrations
no issue
🔥 remove unused loadThemes API method
🚨 Add tests for themes.readOne
🔥 Don't update settings cache for imports
- this isn't needed as of #8057
- settings.edit fires an event, that will result in the update happening automatically
🎨 Move validation to themes
- slowly collecting all theme-related code together
🔥 Reduce DEBUG output
- all this info is a bit tooooo much!
closes#8037🔥 Remove API-level default settings population
- This is a relic!
- We ALWAYS populate defaults on server start therefore this code could never run.
- This was a lot of complicated code that wasn't even needed!!
🎨 Move settings cache
- Move settings cache to be its own thing
- Update all references
- Adds TODOs for further cleanup
🎨 Create settings initialisation step
- Create new settings library, which will eventually house more code
- Unify the interface for initialising settings (will be more useful later)
- Reduce number of calls to updateSettingsCache
refs #2182
* 🔥 Remove unused options from server init
- this is left over from old code and is now unused
* 🎨 Move knex-migrator check to db health
- Move complex check function into own module
- Call module from server/index.js
- This just improves the readability of server/index.js
* 🔥 Remove old comments
- These comments all make no sense now!
* 🎨 ⏱ Move model init out of promise chain
- Model.init() does not return a promise
- Therefore, we can move it to the top of the init function, outside of the promise change
- This should be a minor optimisation, and again improves readability /clarity of what's happening
* ✨⁉️ Move DBHash init / first run to Settings model
- this structure is left over from when we had code we executed on the first run of Ghost
- the implementation used the API to initialise one setting before populateDefaults is called
- this had lots of dependencies - the whole model, API, and permissions structure had to be initialised for it to work
- the new implementation is simpler, it captures the dbHash getting initialised during populateDefaults()
- it also adds an event, so we can do first-run code later if we really want to (or maybe apps can?!)
- perhaps this is hiding behaviour, and there's a nicer way to do it, but populateDefaults seems like a sane place to populate a default setting 😁
* ⏱ Optimise require order so config is first
- the first require to config will cause the files to be read etc
- this ensures that it happens early, and isn't confusingly timed as part of loading a different module
* 🎨 Simplify settings model changes
refs #7489
- as we are now using a different migration approach (knex-migrator), we don't need to remember the database version anymore
- it was once used to check the state of a database and based on it we decided to migrate or not
- with knex-migrator everything depends on the migration table entries and the current ghost version you are on
- on current master the leftover usage is to add the db version when exporting the database, which can be replaced by reading the ghost version
- removing this solves also an interesting migration case with knex-migrator:
- you are on 1.0
- you update to 1.1, but 1.1 has no migrations
- the db version would remain in 1.0
- because the db version was only updated when knex migrator executed a migration
