* Fix Delete and Create-related locks expiring too fast
Fixes#16238
By default, RedisLock expires after 10 seconds, which may not be enough to
process statuses, especially when those have attached media files.
This commit extends those 10 seconds to 15 minutes, which should be plenty
enough to handle any status, while being short enough to not waste many
sidekiq job retries in the exceedingly rare case in which a sidekiq process
would crash when processing a `Create` or `Delete`.
* Fix other RedisLock autorelease durations
Fixes#15645
- things that only perform a few simple database queries (e.g. finding and
saving a record) have been left unchanged, so they'll still use the default
10s duration
- things that perform significantly more complex database queries have been
changed to a 5 minutes timeout
- things that perform multiple HTTP queries have been changed to a 15 minutes
timeout
* Fix media processing getting stuck on too much stdin/stderr
See thoughtbot/terrapin#5
* Remove dependency on paperclip-av-transcoder gem
* Remove dependency on streamio-ffmpeg gem
* Disable stdin on ffmpeg process
* Fix URI of repeat follow requests not being recorded
In case we receive a “repeat” or “duplicate” follow request, we automatically
fast-forward the accept with the latest received Activity `id`, but we don't
record it.
In general, a “repeat” or “duplicate” follow request may happen if for some
reason (e.g. inconsistent handling of Block or Undo Accept activities, an
instance being brought back up from the dead, etc.) the local instance thought
the remote actor were following them while the remote actor thought otherwise.
In those cases, the remote instance does not know about the older Follow
activity `id`, so keeping that record serves no purpose, but knowing the most
recent one is useful if the remote implementation at some point refers to it
by `id` without inlining it.
* Add tests
Unlike locally-issued blocks, they weren't clearing follow
relationships in both directions, follow requests or notifications.
Co-authored-by: Claire <claire.github-309c@sitedethib.com>
* Add tests
* Ensure deleted statuses are marked as such
* Save some redis memory by not storing URIs in delete_upon_arrival values
* Avoid possible race condition when processing incoming Deletes
* Avoid potential duplicate Delete forwards
* Lower lock durations to reduce issues in case of hard crash of the Rails process
* Check for `lock.aquired?` and improve comment
* Refactor RedisLock usage in app/lib/activitypub
* Fix using incorrect or non-existent sender for relaying Deletes
* Delete status records by batches of 50
* Do not precompute values that are only used once
* Do not generate redis events for removal of public toots older than two weeks
* Filter reported toots a priori for polls and status deletion
* Do not process reblogs when cleaning up public timelines
As in Mastodon proper, reblogs don't appear in public TLs
* Clean the deleted account's own feed in one go
* Refactor Account#clean_feed_manager and List#clean_feed_manager
* Delete instead of destroy a few more associations
* Fix preloading
Co-authored-by: Claire <claire.github-309c@sitedethib.com>
Extract logic for determining ActivityPub inboxes to send deletes
to to its own class and explicitly include the person the status
replied to (even if not mentioned), people who favourited it, and
people who replied to it (though that one is still not recursive)
Nginx can be configured to bypass proxy cache when a special header
is in the request. If the response is cacheable, it will replace
the cache for that request. Proxy caching of media files is
desirable when using object storage as a way of minimizing bandwidth
costs, but has the drawback of leaving deleted media files for
a configured amount of cache time. A cache buster can make those
media files immediately unavailable. This especially makes sense
when suspending and unsuspending an account.
When failing to fetch the target account, the ProcessingWorker fails
as expected, but since it hasn't cleared the `move_in_progress` flag,
the next attempt at processing skips the `Move` activity altogether.
This commit changes it to clear the flag when encountering any
unexpected error on fetching the target account. This is likely to
occur because, of, e.g., a timeout, when many instances query the
same actor at the same time.
* Add support for followers synchronization on the receiving end
Check the `collectionSynchronization` attribute on `Create` and `Announce`
activities and synchronize followers from provided collection if possible.
* Add tests for followers synchronization on the receiving end
* Add support for follower synchronization on the sender's end
* Add tests for the sending end
* Switch from AS attributes to HTTP header
Replace the custom `collectionSynchronization` ActivityStreams attribute by
an HTTP header (`X-AS-Collection-Synchronization`) with the same syntax as
the `Signature` header and the following fields:
- `collectionId` to specify which collection to synchronize
- `digest` for the SHA256 hex-digest of the list of followers known on the
receiving instance (where “receiving instance” is determined by accounts
sharing the same host name for their ActivityPub actor `id`)
- `url` of a collection that should be fetched by the instance actor
Internally, move away from the webfinger-based `domain` attribute and use
account `uri` prefix to group accounts.
* Add environment variable to disable followers synchronization
Since the whole mechanism relies on some new preconditions that, in some
extremely rare cases, might not be met, add an environment variable
(DISABLE_FOLLOWERS_SYNCHRONIZATION) to disable the mechanism altogether and
avoid followers being incorrectly removed.
The current conditions are:
1. all managed accounts' actor `id` and inbox URL have the same URI scheme and
netloc.
2. all accounts whose actor `id` or inbox URL share the same URI scheme and
netloc as a managed account must be managed by the same Mastodon instance
as well.
As far as Mastodon is concerned, breaking those preconditions require extensive
configuration changes in the reverse proxy and might also cause other issues.
Therefore, this environment variable provides a way out for people with highly
unusual configurations, and can be safely ignored for the overwhelming majority
of Mastodon administrators.
* Only set follower synchronization header on non-public statuses
This is to avoid unnecessary computations and allow Follow-related
activities to be handled by the usual codepath instead of going through
the synchronization mechanism (otherwise, any Follow/Undo/Accept activity
would trigger the synchronization mechanism even if processing the activity
itself would be enough to re-introduce synchronization)
* Change how ActivityPub::SynchronizeFollowersService handles follow requests
If the remote lists a local follower which we only know has sent a follow
request, consider the follow request as accepted instead of sending an Undo.
* Integrate review feeback
- rename X-AS-Collection-Synchronization to Collection-Synchronization
- various minor refactoring and code style changes
* Only select required fields when computing followers_hash
* Use actor URI rather than webfinger domain in synchronization endpoint
* Change hash computation to be a XOR of individual hashes
Makes it much easier to be memory-efficient, and avoid sorting discrepancy issues.
* Marginally improve followers_hash computation speed
* Further improve hash computation performances by using pluck_each
There are edge cases where requests to certain hosts timeout when
using the vanilla HTTP.rb gem, which the goldfinger gem uses. Now
that we no longer need to support OStatus servers, webfinger logic
is so simple that there is no point encapsulating it in a gem, so
we can just use our own Request class. With that, we benefit from
more robust timeout code and IPv4/IPv6 resolution.
Fix#14091
* Add bell button
Fix#4890
* Remove duplicate type from post-deployment migration
* Fix legacy class type mappings
* Improve query performance with better index
* Fix validation
* Remove redundant index from notifications
* Check for and record reblog info atomically
Instead of using ZREVRANK to determine whether a reblog is a new reblog or not,
use ZADD's NX option to perform the check/addition option atomically.
* Replace ZREVRANK call with ZSCORE key which is more efficient
* Make tests a bit stricter
* Fix off-by-one
* Add database support for list show-reply preferences
* Add backend support to read and update list-specific show_replies settings
* Add basic UI to set list replies setting
* Add specs for list replies policy
* Switch "cycling" reply policy link to a set of radio inputs
* Capitalize replies_policy strings
* Change radio button design to be consistent with that of the directory explorer
Follow-up to #14359
In the case of limited toots, the receiver may not be explicitly part of the
audience. If a specific user's inbox URI was specified, it makes sense to
dereference the toot from the corresponding user, instead of trying to find
someone in the explicit audience.
* Fix not handling Undo on some activity types when they aren't inlined
When receiving an Undo for a non-inlined activity, try looking it up in
database using the URI. The queries are ad-hoc because we don't have a global
index of object URIs, and not all activity types are stored in database with
an index on their URI.
Announces are just statuses, and have an index on URIs, so this check can
be done efficiently.
Accepts cannot be handled at all because we don't record their URI at any
point.
Follows don't have an index on URI, but they have an index on the issuing
account, which should make such queries largely manageable.
Likes don't have an index on URI, they have an index on the issuing account,
but the number of favs per account may be very high, so I decided not to
handle that.
Blocks don't have an index on URI, but they have an index on the issuing
account, which should make such queries largely manageable.
In all cases, if an Undo could not be handled properly, we call `delete_later!`
because that does not require us to know more than the URI of the undone
property.
* Add tests
* Make newer blocks overwrite older ones
Allows re-synchronizing block info by re-blocking and un-blocking again
when the original Undo Block has been lost.
* Fix the local group's followers collection
* Fix to accept followed relayed_through_account
* Add local delivery to the group's followers
* Fix code style
* Revert "Add local delivery to the group's followers"
This reverts commit 3237effc19.
