v1 queried on an already-committed session; v2 opened storage while
still holding the transaction. Both now load meta under the tx, commit,
then open the file. Addresses pr-swarm findings.
Preview generation reads the whole file (a network round trip on s3),
decodes and resizes it — seconds per image — while the request
transaction pinned a pool connection. A board full of image previews
could drain the pool. Commit first; the storage read needs no
transaction.
Querying the global engine while the request already holds an open
transaction acquires a second pool connection. Under concurrent
attachment downloads (e.g. a kanban board full of image previews) all
pool connections end up held by transactions that each wait for an
extra connection that can never be freed, hanging every request until
restart.
The redundant meta load in the upload avatar provider is removed
entirely; the decoded image is all that path uses.
SQLite supports ALTER TABLE DROP COLUMN since 3.35 and the bundled
driver ships 3.50.x, so dropTableColum no longer needs to skip SQLite
with a warning that leaves dead columns behind.
The unconditional type=raw,value=latest applied the latest tag to every
pushed git tag, including prereleases like v1.0.0-rc1. Removing it lets
metadata-action's default latest=auto flavor handle it, which only tags
latest for non-prerelease semver versions.
Fixes#2161
Todoist returns opaque identifiers instead of urls in file_url for
attachments it does not host itself (mail attachments for example).
Passing those to the http client failed with "unsupported protocol
scheme" and aborted the entire migration.
Skip attachments without an http(s) url and log-and-continue when a
single download fails instead of failing the whole migration.
The auto-label workflow's LLM call succeeds, but the docker action ran
as a non-root user (`USER appuser` in the Dockerfile) and could not
write the runner's `GITHUB_OUTPUT` file:
```
Error: failed to set output: ... open /github/file_commands/set_output_...: permission denied
```
GitHub docker actions must run as root — the runner's file-command files
are owned by the runner user. Fixed in the fork
(tink-bot/LLM-action@8526bab removes the non-root user) and bumped the
pin here.
Failing run:
https://github.com/go-vikunja/vikunja/actions/runs/30904804523/job/91977288580
Co-authored-by: kolaente <k@knt.li>
The config was shipped as plain `%config`, so rpm installs the packaged
file on every upgrade and moves the user's version aside as `.rpmsave`.
The postinstall script then re-ran its seds against the fresh sample and
generated a new JWT secret, logging everyone out — with the real config
no longer in place.
Mark it `config|noreplace` so rpm keeps the existing file and writes
`.rpmnew` instead. deb and apk treat the type exactly like `config`, so
nothing changes there.
nfpm only maps `scripts.postinstall` to apk's `.post-install` and
archlinux's `post_install`, both of which run on a fresh install. Their
upgrade hooks come from the packager-specific `apk.scripts.postupgrade`
and `archlinux.scripts.postupgrade` keys, which we never set — so
upgrading on Alpine or Arch executed no maintainer script whatsoever, and
the restart added in the previous commit would never have fired there.
Point both at the matching install script.
The postinstall script only ran `systemctl enable`, which creates the
wants-symlink and nothing else. On an upgrade the package manager
replaced the binary while the old process kept running the old inode, so
users had to restart vikunja by hand to get the new version. Changes to
vikunja.service were never picked up either, for lack of a
`daemon-reload`.
Add a daemon-reload plus `try-restart` (`rc-service restart` on OpenRC).
try-restart is a no-op while the unit is stopped, so fresh installs keep
their current behaviour of being enabled but not started.
The config fixups move above the service handling so a restart never
races a half-patched config.
renovate[bot]andkolaente
dependabot[bot]andkolaente
Frederick [Bot]
TinkandGitHub