Database dumps sometimes reference functions in the `discourse_functions` schema. It's possible that some of these functions have been dropped in a newer version of Discourse. In that case, restoring an older backup will fail with a `ERROR: function discourse_functions.something_something() does not exist` error. The restore functionality contains a workaround for that problem, but it didn't work with functions created in plugin migrations.
This commit adds support for temporarily creating missing `discourse_functions` from plugins. And it adds a simple check if the DB migration file even contains the required `DROPPED_TABLES` or `DROPPED_COLUMNS` constant. We don't need to create an instance of the DB migration class unless one of those constants is used. This makes the restore slightly faster and works around a problem with migrations that execute without `up` or `down` methods (e.g. `BackfillChatChannelAndThreadLastMessageIdsPostMigrate`).
Constants should always be only assigned once. The logical OR assignment
of a constant is a relic of the past before we used zeitwerk for
autoloading and had bugs where a file could be loaded twice resulting in
constant redefinition warnings.
I am changing many of these to notes or resolving them as is,
most of these I have not actively worked on in years so someone
else can work on them when we get to these areas again.
* FEATURE: core code, tests for feature to allow backups to removed based on a time window
* FEATURE: getting tests working for time-based backup
* FEATURE: getting tests running
* FEATURE: linting
AWS recommends running buckets without ACLs, and to use resource policies to manage access control instead.
This is not a bad idea, because S3 ACLs are whack, and while resource policies are also whack, they're a more constrained form of whack.
Further, some compliance regimes get antsy if you don't go with the vendor's recommended settings, and arguing that you need to enable ACLs on a bucket just to store images in there is more hassle than it's worth.
The new site setting (s3_use_acls) cannot be disabled when secure
uploads is enabled -- the latter relies on private ACLs for security
at this point in time. We may want to reexamine this in future.
In a multisite Discourse reported that no backup is running after 60 seconds because the Redis key expired. Also, the thread that listens for a shutdown signal stopped running immediately because it didn't detect a running operation.
This also adds an optional `ticket` parameter to `Backuper` which allows identifying the backup in `backup_complete` and `backup_failed` events. Both events contain the logs as payload and moving some methods around ensures that all errors are included in the logs.
This table holds associations between uploads and other models. This can be used to prevent removing uploads that are still in use.
* DEV: Create upload_references
* DEV: Use UploadReference instead of PostUpload
* DEV: Use UploadReference for SiteSetting
* DEV: Use UploadReference for Badge
* DEV: Use UploadReference for Category
* DEV: Use UploadReference for CustomEmoji
* DEV: Use UploadReference for Group
* DEV: Use UploadReference for ThemeField
* DEV: Use UploadReference for ThemeSetting
* DEV: Use UploadReference for User
* DEV: Use UploadReference for UserAvatar
* DEV: Use UploadReference for UserExport
* DEV: Use UploadReference for UserProfile
* DEV: Add method to extract uploads from raw text
* DEV: Use UploadReference for Draft
* DEV: Use UploadReference for ReviewableQueuedPost
* DEV: Use UploadReference for UserProfile's bio_raw
* DEV: Do not copy user uploads to upload references
* DEV: Copy post uploads again after deploy
* DEV: Use created_at and updated_at from uploads table
* FIX: Check if upload site setting is empty
* DEV: Copy user uploads to upload references
* DEV: Make upload extraction less strict
Previously we hardcoded the DOWNLOAD_URL_EXPIRES_AFTER_SECONDS const
inside S3Helper to be 5 minutes (300 seconds). For various reasons,
some hosted sites may need this to be longer for other integrations.
The maximum expiry time for presigned URLs is 1 week (which is
604800 seconds), so that has been added as a validation on the
setting as well. The setting is hidden because 99% of the time
it should not be changed.
* File.exists? is deprecated and removed in Ruby 3.2 in favor of
File.exist?
* Dir.exists? is deprecated and removed in Ruby 3.2 in favor of
Dir.exist?
This PR introduces a new `enable_experimental_backup_uploads` site setting (default false and hidden), which when enabled alongside `enable_direct_s3_uploads` will allow for direct S3 multipart uploads of backup .tar.gz files.
To make multipart external uploads work with both the S3BackupStore and the S3Store, I've had to move several methods out of S3Store and into S3Helper, including:
* presigned_url
* create_multipart
* abort_multipart
* complete_multipart
* presign_multipart_part
* list_multipart_parts
Then, S3Store and S3BackupStore either delegate directly to S3Helper or have their own special methods to call S3Helper for these methods. FileStore.temporary_upload_path has also removed its dependence on upload_path, and can now be used interchangeably between the stores. A similar change was made in the frontend as well, moving the multipart related JS code out of ComposerUppyUpload and into a mixin of its own, so it can also be used by UppyUploadMixin.
Some changes to ExternalUploadManager had to be made here as well. The backup direct uploads do not need an Upload record made for them in the database, so they can be moved to their final S3 resting place when completing the multipart upload.
This changeset is not perfect; it introduces some special cases in UploadController to handle backups that was previously in BackupController, because UploadController is where the multipart routes are located. A subsequent pull request will pull these routes into a module or some other sharing pattern, along with hooks, so the backup controller and the upload controller (and any future controllers that may need them) can include these routes in a nicer way.
This commit introduces a new s3:ensure_cors_rules rake task
that is run as a prerequisite to s3:upload_assets. This rake
task calls out to the S3CorsRulesets class to ensure that
the 3 relevant sets of CORS rules are applied, depending on
site settings:
* assets
* direct S3 backups
* direct S3 uploads
This works for both Global S3 settings and Database S3 settings
(the latter set directly via SiteSetting).
As it is, only one rule can be applied, which is generally
the assets rule as it is called first. This commit changes
the ensure_cors! method to be able to apply new rules as
well as the existing ones.
This commit also slightly changes the existing rules to cover
direct S3 uploads via uppy, especially multipart, which requires
some more headers.
In preparation for adding automatic CORS rules creation
for direct S3 uploads, I am adding tests here and moving the
CORS rule definitions into a dedicated class so they are all
in the one place.
There is a problem with ensure_cors! as well -- if there is
already a CORS rule defined (presumably the asset one) then
we do nothing and do not apply the new rule. This means that
the S3BackupStore.ensure_cors method does nothing right now
if the assets rule is already defined, and it will mean the
same for any direct S3 upload rules I add for uppy. We need
to be able to add more rules, not just one.
This is not a problem on our hosting because we define the
rules at an infra level.
Discourse automatically sends a private message after backup or
restore finished. The private message used to contain the log inline
even when it was very long. A very long log can create issues because
the length of the post will be over the maximum allowed length of a
post. When that happens, Discourse will try to create an upload with
the logs. If that fails, it will trim the log and inline it.
You can use `discourse restore --location=local FILENAME` if you want to restore a backup that is stored locally even though the `backup_location` has the value `s3`.
It pauses Sidekiq, clears Redis (namespaced to the current site), clears Sidekiq jobs for the current site, restores the database and unpauses Sidekiq. Previously it stayed paused until the end of the restore.
Redis is cleared because we don't want any old data lying around (e.g. old Sidekiq jobs). Most data in Redis is prefixed with the name of the multisite, but Sidekiq jobs in a multisite are all stored in the same keys. So, deleting those jobs requires a little bit more logic.
After restoring a backup it takes up to 48 hours for uploads stored on S3 to appear in the S3 inventory. This change prevents alerts about missing uploads by preventing the EnsureS3UploadsExistence job from running in the first 48 hours after a restore. During the restore it deletes the count of missing uploads from the PluginStore, so that an alert isn't triggered by an old number.
This is a follow-up for f51ccea028 because specs were failing on macOS. macOS uses bsdtar by default and the --transform option isn't supported by bsdtar. But the -s option is and it works the same.
This also ensures that restoring a backup works when it was created with the wrong upload paths in the time between ab4c0a4970 (shortly after v2.6.0.beta1) and this fix.
The `EXECUTE FUNCTION` syntax for `CREATE TRIGGER` statements was introduced in PostgreSQL 11. We need to replace `EXECUTE FUNCTION` with `EXECUTE PROCEDURE` in order to be able to restore backups created with PG12 on PG10.
* DEV: new S3 backup layout
Currently, with $S3_BACKUP_BUCKET of "bucket/backups", multisite backups
end up in "bucket/backups/backups/dbname/" and single-site will be in
"bucket/backups/".
Both _should_ be in "bucket/backups/dbname/"
- remove MULTISITE_PREFIX,
- always include dbname,
- method to move to the new prefix
- job to call the method
* SPEC: add tests for `VacateLegacyPrefixBackups` onceoff job.
Co-authored-by: Vinoth Kannan <vinothkannan@vinkas.com>
This reverts commit 20780a1eee.
* SECURITY: re-adds accidentally reverted commit:
03d26cd6: ensure embed_url contains valid http(s) uri
* when the merge commit e62a85cf was reverted, git chose the 2660c2e2 parent to land on
instead of the 03d26cd6 parent (which contains security fixes)
