ActiveStorage::Record and ApplicationRecord used separate connection
pools (from separate `connects_to` calls). This meant `belongs_to
touch: true` on attachments silently failed — the card couldn't join the
attachment's transaction to receive the deferred touch, so fragment
caches were never invalidated after image removal.
Fix by delegating `connection_pool` from ActiveStorage::Record to
ApplicationRecord so they share a single pool.
These URLs can be problematic or inconsistent, as we might end up with
rich text content with beta URLs for attachments being used in
production, or viceversa. It'll also be problematic when importing or
exporting data between a self-hosted instance and the SaaS version.
An append-only storage ledger replaces deadlock-prone synchronous
counter updates and drift-prone async updates.
All content storage (card images, card/comment/board description embeds)
is tracked. Account exports (which expire) and avatars are not tracked.
Storage is accounted for by Account and Board. Both consume the same
event stream independently: no bubble-up storage bumps triggering
deadlocks due to lock sequencing. Each calculates its own total from
the underlying ledger with independent cursors and materialization.
* Storage::Entry: Append-only ledger recording attach/detach/transfer
events with delta bytes. Single event stream indexed for both
Account and Board cursor queries.
* Storage::Total: Polymorphic snapshot cache with cursor (last_entry_id)
tracking which entries have been materialized.
* Storage::Totaled concern: Provides bytes_used (fast snapshot) and
bytes_used_exact (snapshot + pending) query modes, plus
materialize_storage! to roll up pending entries.
* Storage::Tracked concern: For models owning attachments (Card,
Comment, Board). Provides board_for_storage_tracking for models
where board is determined differently (Board returns self).
Handles board transfers by recording transfer_out/transfer_in entries.
* Storage::AttachmentTracking: Hooks ActiveStorage::Attachment lifecycle
to record attach/detach entries. Handles ActionText::RichText embeds
by traversing to the actual model. Snapshots context in before_destroy
to handle cascading deletes where parent record may be gone by
after_destroy_commit.
* MaterializeJob: Rolls up pending entries into snapshot. Concurrency
limited per owner to prevent duplicate work.
* ReconcileJob: On-demand reconciliation against actual attachment
storage for support/debugging. Compares ledger total to real bytes
from card images, card embeds, comment embeds, and board embeds.
Usage:
* account.bytes_used / board.bytes_used: fast, slightly stale bytesize
* account.bytes_used_exact / board.bytes_used_exact: real-time bytesize
* Storage::Entry: audit trail for debugging and point-in-time queries
When ActiveStorage::Record uses `connects_to` for read replica support,
it creates a separate connection pool from ApplicationRecord. This causes
`after_commit` callbacks to fire in non-deterministic order - the
Attachment's `create_variants` callback can fire before the User model's
upload callback completes, resulting in FileNotFoundError.
The fix removes replica connection configuration from ActiveStorage::Record
so it shares the same connection pool as application models, ensuring
proper callback ordering.
Also reverts test workarounds that were added to work around this issue,
since the root cause is now fixed.
See: https://github.com/rails/rails/issues/53694
This adds an initial configuration for staging. Note that we are doing
this ahead of having the full infrastructure in all 3 DCs. So this
this will result in some cross-DC writes for now (e.g. we have a single cache in
IAD). We'll correct this as the infrastructure becomes available.
We'll also run jobs on every app server, until we split them out to
separate instances.
Mike Dalessio