The migration failed on MySQL with a 1062 duplicate-entry error: recollating
name/slug to utf8mb4_unicode_520_ci makes previously-distinct values collide
on the unique indexes, so the ALTER TABLE was rejected.
Merge colliding rows first, comparing values under the target collation. For
each collision group the lowest id is kept, references in status_hashtags,
hashtag_follows, hashtag_related and discover_category_hashtags are repointed
to it (UPDATE IGNORE + cleanup), and the losing rows are deleted before the
collation is applied. Adds a regression test for the merge path.
MySQL/MariaDB's utf8mb4_unicode_ci collation treats all characters outside
the Basic Multilingual Plane as equal, conflating distinct same-length
hashtags (e.g. Shavian vs cuneiform) on the unique name/slug indexes.
Migrate the hashtags name/slug columns to utf8mb4_unicode_520_ci, which
differentiates supplementary-plane characters. Improvements over the
original PR:
- Use an explicit ALTER ... MODIFY, since Laravel's fluent ->change() emits
no collation change on MySQL/MariaDB and silently no-ops.
- Match both 'mysql' and 'mariadb' drivers (Laravel 11+ reports MariaDB as a
distinct driver, so a mysql-only check would skip the fix on MariaDB).
- Provide an accurate, reversible down() and preserve NOT NULL + unique keys.
Add feature tests covering the no-op path on non-MySQL drivers and, on
MySQL/MariaDB, that distinct BMP-outside hashtags coexist while same-slug
and case-insensitive dedup still work.
Cover source-of-truth resync of likes/boosts/comments, dry-run, no-op on
correct data, --type restriction, argument validation, and bulk --all
mode. Includes regression tests for the two reporting bugs: the summary
now lists only drifted metrics, and a null reply_count renders as 0.
The resynced summary printed all three counts unconditionally, which
made an untouched metric (e.g. an already-correct comments count) look
like it had been resynced. Drive the summary from the drifted set and
show before->after values, so it matches the drift detection exactly.
reply_count is a nullable column, so NULL rendered as an empty string in
the resync summary. Cast the summary output to int so a null/absent
comment count prints as 0. No behavior change to the reconcile logic.
FixProfileCounts and FixPostCounts use the admin: signature prefix and
are operator-run maintenance tools, so move them from FixBugs/ to Admin/
(namespace updated) and refresh the README tables to match.
Add a FixPostCounts command mirroring admin:fixProfileCounts (single-id,
--all --scope, --active, --type, --dry-run, --force). It reconciles the
statuses likes_count, reblogs_count, and reply_count columns against
source-of-truth tables.
Add canonical recompute helpers and reconcileStatusCounts() to
StatusService (mirroring AccountStatService), busting the status cache
only when a column actually drifted.
status:dedup and fix:avatars address historical data states that can no
longer occur (unique statuses.uri index since 2019; SVG identicon avatars
no longer generated). Move both to a Deprecated/ folder and update the
README audit accordingly. media:fix stays in FixBugs/ since image filters
are still an active feature.
Group console commands into Admin, Dev, FixBugs, Install, Internal, and
User subfolders (matching the earlier reorganization), and add a new
Status subfolder for the status:user, status:profile, and status:post
debug commands. Namespaces updated to match; command signatures and the
total command count are unchanged.
Rename user:status, profile:status, and post:status console commands
to status:user, status:profile, and status:post. Rename the command
files and classes to match (StatusUser, StatusProfile, StatusPost) and
update the cross-reference tip in StatusProfile.
Helpers::importNoteAttachment unconditionally inserted a new Media row per
attachment, so re-importing a remote status (an Announce racing another
inbox job, a re-fetch, or a duplicate url within one activity) hit the
media_status_id_media_path_unique constraint and crashed the queue job with
a 1062 UniqueConstraintViolationException, dropping the boost/import.
Make createMediaAttachment idempotent on (status_id, media_path): skip when
a row already exists, and catch the unique-constraint violation as a
lost-race no-op, returning null so the caller skips re-dispatching storage.
Adds regression tests (re-import no-op, distinct urls still stored,
concurrent-insert returns null).
Cold-migrate existing media from an old S3 bucket to the current cloud
bucket, one media row at a time (like MigrateLocalS3MediaURL):
- Source = --sourceDisk (default s3-old, reads AWS_OLD_*); destination = the
current cloud disk (config filesystems.cloud). No .env editing: operators
point AWS_* at the new bucket first (restarting workers as usual) so new
uploads/downloads land on the new bucket, then run this to backfill old data.
- Copies media (+thumbnail) source->destination, verifies by size and by
sha256 of the freshly-written destination object (against original_sha256),
rewrites cdn_url/optimized_url/thumbnail_url to the destination host, and
GCs the source objects (unless --keep-source). Busts caches.
- Only touches rows whose cdn_url still points at the source host; idempotent.
- --sourceDisk / --limit / --dry-run / --force.
- Adds the s3-old disk (AWS_OLD_*) to config/filesystems.php and feature tests.
Add admin:MediaMoveStorageLocalToCloud and admin:MediaMoveStorageCloudToLocal:
- Copy media (+thumbnail) between local and cloud disks, verify by size (and
sha256 against original_sha256 when present) before deleting the source.
- Integrated GC: delete the verified source copy (local on upload, cloud on
download), set version=4 / reset to 3, and bust MediaService/StatusService
caches. --keep-local / --keep-cloud opt out.
- Manage PF_ENABLE_CLOUD in .env AND the live runtime + config cache so new
uploads route to the correct backend mid-migration on a hot server. Uses the
installer's atomic .env writer (shared ManagesMediaStorageEnv trait).
- --limit / --dry-run / --force.
Replaces media:migrate2cloud (CloudMediaMigrate) and media:s3gc
(MediaS3GarbageCollector); scheduler now runs MediaMoveStorageLocalToCloud
hourly for straggler upload + GC. Keeps media:fix-nonlocal-driver.
Adds feature tests (download+GC, --keep-cloud, dry-run, env-flag flip both
directions, unknown-disk guard).
config_cache() falls through to config() when instance.enable_cc is off
(ENABLE_CONFIG_CACHE=false, as in CI/.env.testing), so ConfigCacheService::put()
alone did not toggle pixelfed.cloud_storage and the command's cloud-enabled
guard aborted with exit 1. Set the underlying config value too (both in
beforeEach and the local-storage refusal test).
Rename the command (and test) to admin:MigrateLocalS3MediaURL to reflect its
scope: rewriting stale S3/cloud media URLs only. Remove avatar handling and
the --avatars option; the command now focuses solely on status media
(cdn_url, thumbnail_url, optimized_url).
Rebuilds stale local media URLs (cdn_url, thumbnail_url, optimized_url) and
avatar cdn_urls from their storage paths via the configured cloud disk.
- Default target host comes from the configured cloud disk (AWS_URL);
requires confirmation (or --force) and can be overridden with --newDomain.
- Optional --oldDomain filters to a single old backend host; by default all
stale hosts are rewritten.
- Refuses to run when PF_ENABLE_CLOUD is false (local storage) and, when
auto-detecting, refuses a target equal to the app domain — so local-storage
instances are never rewritten.
- Single status id / post URL, --all, --avatars; --dry-run; busts
MediaService/StatusService caches for affected statuses.
- Removes the superseded media:cloud-url-rewrite command.
- Adds feature tests covering rewrite/skip/dry-run/oldDomain/newDomain/
remote-skip/local-storage-refusal.
Dumps a Status and its media for debugging. Accepts a post id or URL
(/p/username/ID). Shows status columns, author, every media row's storage
fields (media_path, thumbnail_path, cdn_url, thumbnail_url, optimized_url,
remote_url, etc.), computed url()/thumbnailUrl()/expected-from-path, a URL
health check comparing stored URL hosts against the configured cloud disk
host (flags stale hosts), and the cached MediaService media_attachments
actually served to clients.
Bulk --all reconciliation previously scanned both local and remote profiles
implicitly. Now --all requires an explicit --scope of local, remote, or
both. --active stays local-only and rejects a non-local --scope. Adds the
BelongsTo return type to Profile::user() so the whereHas('user') scope
filter passes Larastan, and adds tests for scope requirement/validation and
local/remote filtering.
The <intersect> in sections/Notifications.vue wrapped four <placeholder>
elements directly. vue-intersect requires exactly one child (it checks
$slots.default.length and observes $slots.default[0]), so it logged
'[VueIntersect] You may only wrap one element in a <intersect> component.'
and only observed the first placeholder. Wrap the placeholders in a single
<div> so the slot has one root element.
- Rename command signature fix:profilecounts -> admin:fixProfileCounts.
- --active is now its own bulk mode (recently-active local accounts),
mutually exclusive with --all and a single id.
- Add --type=followers|following|statuses to restrict reconciliation to a
single metric (validated).
- Update/extend tests for the new name, --type restriction and invalid-type
rejection.
- Add --force flag to fix:profilecounts for unattended runs and schedule
'fix:profilecounts --all --force' weekly (Sun 03:37) as a safety-net
reconcile. Kept as a low-frequency full scan rather than a new event-driven
dirty-set; it only writes profiles that actually drifted.
- Add Feature tests for AccountStatService recompute helpers and
reconcileProfileCounts (media-type status_count semantics, follower/
following counts, drift/no-drift/no-write, metric restriction, missing
profile) plus fix:profilecounts command behavior (silent-when-synced,
dry-run makes no changes).
Extract canonical source-of-truth count logic into AccountStatService:
recalculateStatusCount/FollowerCount/FollowingCount and a
reconcileProfileCounts() that fixes only drifted columns and busts caches.
Both the scheduled app:account-post-count-stat-update (status-only, its
correct scope) and fix:profilecounts now use these instead of duplicating
the SQL. Also corrects the status_count definition to match the actual
increment logic in StatusEntityLexer/StatusDelete (media post types only:
photo/video albums), rather than the previous inconsistent all-statuses /
scoped counts that could themselves cause drift.
The scheduled updater keeps its incremental, dirty-set design and remains
status-only; follower/following stay owned by FollowServiceWarmCache.
Your Name
Shlee
Daniel Supernault