Repository navigation
v1.9.0-rc.1
Pre-release
Pre-release
v1.9.0-rc.1
[1.9.0-rc.1] - 2026-11-20
Warning
Downgrading from 1.9 to 1.8 needs the pre-upgrade backup. 1.9 applies store schema migrations 15 and 16 (the audit actor column and an update-operation index) on first start, and 1.8 supports migrations up to 14. 1.8 refuses a database with a newer schema and stops at startup with a store-schema-too-new error instead of opening it. Stop Drydock and back up the complete store directory before upgrading, and restore that backup to go back to 1.8. Changes made after the upgrade do not carry back.
Added
- GitHub Dispatch and GitLab Pipeline triggers ask your CI to rebuild a locally built image when its upstream base image has a newer tag.
githubdispatchsends arepository_dispatchevent (GitHub Enterprise Server throughURL) andgitlabpipelinecalls the pipeline trigger API (self-hosted GitLab throughURL). Both are notification triggers with theDD_NOTIFICATION_*prefix, fire only for base image updates of locally built images unlessBASEIMAGEONLY=false, send the container, image, local tag, base image and current and new base tag and digest for a single update or a batch, support a proxy and an insecure TLS option, and keep the token (and GitLab variable values) out of logs and the configuration API. - A watcher can now read a static list of images from a YAML file instead of watching a Docker host. Set
DD_WATCHER_{name}_PROVIDER=fileandDD_WATCHER_{name}_PATHto the file. Each entry is an image with an optional name,dd.*labels andtaginclude, exclude, transform and family fields, and becomes a synthetic container, so tag filters, registry authentication and notification routing work as they do for a Docker container. It covers pre-pull staging images, CI base images and DockerfileFROMlines. The file is re-read on the schedule and shortly after it changes. A bad entry, a duplicate name or a deleted file keeps the last good list and logs what is wrong. These containers are notification-only: action triggers skip them, the update API refuses them, and the container API reportssource: file. - Container state alerts now cover containers on Portwing agents, over both the Standard and Edge connections. The controller reads the die, oom and stop events a Portwing agent forwards (agents that advertise the
dd:container-eventcapability), and raises the samecontainer-stopped,container-exited,container-oom-killedandcontainer-restart-loopalerts, with the restart loop defaults of 3 crashes in 300 seconds. Only containers Drydock watches alert, and updates and rollbacks Drydock runs stay quiet. Edge can drop events under load and nothing is replayed after a Docker reconnect, so an alert can be missed. - Container state alerts now cover containers on remote agents. The agent runs the same classifier next to its Docker events and sends the finished
container-stopped,container-exited,container-oom-killedorcontainer-restart-loopalert to the controller, so one crash is still one alert, restart loops use the agent'sRESTARTLOOP_THRESHOLDandRESTARTLOOP_WINDOW, and updates and rollbacks on that agent stay quiet. Agents from before this release send nothing and raise no alerts, with no errors on either side. - A locally built image can now be watched against its upstream base. Set
dd.registry.lookup.imageon a container whose image was built on the host and never pushed, and Drydock looks that base up, treats your local tag as the base version and reports a newer tag as a base image update (updateKind.baseImagein the API, with the default notification worded to match). It is notification-only: the Docker and Docker Compose actions refuse to run for a locally built image, by hand with a409or automatically, so a build is never replaced by the stock upstream image. A command action can be wired to rebuild. Locally built images without the label behave as before (#897). - Containers now list the best update candidate for each kind of change. Next to the highest allowed tag,
result.updateCandidatesholds the newestpatch,minorandmajortag measured against the running version, each one filtered by the include and exclude patterns, tag family,dd.tag.ceiling, transforms and rejected release lines, and the container detail view shows them. The main update is unchanged. A new trigger option,SEPARATEUPDATETYPES=trueon an action or notification trigger, makes a trigger whoseTHRESHOLDrules out the highest update act on the best candidate inside the threshold instead of doing nothing; the notification, approval item and update then name that candidate. It defaults to off. Digest-only and non-semver updates only have the main update, and the list is only recorded when at least two kinds are on offer. - Updates can now be capped at the running version. Set the
dd.tag.ceilinglabel (orTAG_CEILINGon a watcher or imgset) tomajorto stay inside the running major, orminorto stay inside the running major.minor. The ceiling is read from the transformed tag, composes withdd.tag.family, include and exclude, and is ignored with a debug log when the running tag isn't semver.dd.tag.includeanddd.tag.excludecan also use${major},${minor},${patch}and${flavor}from the running tag. Values are escaped before the RE2 pattern is compiled, and a filter that uses a variable the running tag can't supply offers no candidates and logs a warning instead of matching everything. - Edge agent keys can now be managed from the Agents view. A new Edge agent keys panel registers a Portwing agent's public key (paste the bare key or the line
portwing keygen -pub-fromprints), lists keys as active or revoked, and revokes them after a confirmation, which disconnects the agent. It also shows a Compose file and adocker runcommand for the agent with the controller URL filled in from your browser address, editable before you copy. Only public keys are registered, so nothing secret is displayed. It uses the existing key registry API, so it needs a signed-in session, and API keys and recovery logins see a notice instead of the controls. - Security fast lane, an opt-in policy for updates that fix vulnerabilities in the image you're running. Set the
dd.security.fastlane=truelabel on a container, orDD_SECURITY_FASTLANE=truefor all of them, and a pending update that fixes at least one critical or high CVE or GHSA from the running image's latest scan, without adding a new critical, skips the maturity min age, snooze and maintenance window holds. It never liftsdd.action.exclude, update thresholds, approval, skip and reject entries or digest pinning, and an update that needs approval still waits in the queue marked as a security fix. The candidate has to be scanned first, by digest, and both scans must be complete, taken by real sha256 digests (the running image is now scanned by its digest too), bound to the running and pending digests, from the same scanner and in order, or the update isn't fast-laned. The scheduled scan now scans each pending candidate digest of an opted-in container once per running scan, shared by every container on that digest, and the on-demand scan does it for held updates too. Scans now also record their full critical and high ID lists before the stored list is cut off, and the reference they scanned. After the pull, a fast-laned install is aborted and audited if the pulled digest isn't the one that qualified; for multi-arch images the candidate also records its manifest list digest, which a pull by tag stores locally, and the install accepts either. Each fast-laned auto update writes asecurity-fast-laneaudit entry, and the container (so notifications and events) carries asecurityFastLaneobject with the fixed IDs, capped at 10. See Security fast lane. - The MQTT trigger's Home Assistant integration now sends release notes and availability, and can add buttons. The update entity gets
release_summary(cut to Home Assistant's 255 characters) andrelease_urlwhen drydock knows them. The client registers a Last Will and publishes a retained online message on{topic}/status, and every discovery config names that topic as itsavailability_topic, so entities go unavailable when drydock disconnects. WithHASS_BUTTONS=true(off by default) each container also gets a Check for updates button and an Update now button; Update now goes through the same admission rules as a manual update in the UI, and a refused press publishes the reason on a result topic. - Docker and Docker Compose actions now pace image pulls per registry. A new
PULLCONCURRENCYoption (default2) bounds simultaneous daemon pulls per registry host across all actions, with Docker Hub aliases counted as one registry. When a pull is throttled with HTTP 429 ortoomanyrequests, that registry's waiting pulls are deferred, honoringRetry-Afterwhen present and otherwise backing off exponentially up to 5 minutes, while other registries keep pulling. Each backoff is logged once. - Scheduled and disk-threshold cleanup policies. Set
DD_CLEANUP_<POLICY>_CRONand/or_THRESHOLDBYTESand opt in to each prune type separately: dangling images, unused images (with an older-than filter), stopped containers, unused networks and build cache. Anonymous volumes are supported only behind a second opt-in. Policies cover local and agent hosts, honor add.cleanup.exclude=truelabel, never touch the rollback backup of a recent update or a-old-<timestamp>container still inside the rollback window, and skip a host while an update is running on it. Each run writes an audit entry with the bytes reclaimed per type,DRYRUN=truereports without removing anything, andGET /api/v1/cleanupandPOST /api/v1/cleanup/runshow the last results and run a policy on demand. The threshold is a reclaimable-bytes figure because Docker does not report filesystem usage. - The two-factor status read at
GET /api/v1/auth/totp-factornow reportsmanagementAllowed, plusreason: "https-required"when it is false, using the same HTTPS-or-local check the enrollment routes apply. The Profile tab uses it to show the HTTPS hint and hide the Turn on button as soon as it loads, instead of after you've typed a password and been refused. drydock scanruns one fresh, observation-only scan and exits, for CI jobs and scripts. It prints a single JSON document or, with--format ndjson, one line per container, with stable field names and ordering, and--updates-onlylists only containers with an update. Exit codes are0for no updates,3for updates found,1for a technical failure such as a registry error,2for a usage error and4when--timeoutruns out. It starts no HTTP listener, scheduler or events stream, keeps its database in memory, loads no triggers so nothing can notify or update, and sends logs to stderr so stdout holds only the output.- Opt-in liveness watch for containers without a
HEALTHCHECK. Set thedd.rollback.livenesslabel, or theLIVENESSWINDOWoption on the Docker or Docker Compose action, to a window in milliseconds (unlikeRESTARTLOOP_WINDOW, which is in seconds). After an update, Drydock watches the recreated container for that long and rolls back to the previous image if it exits, is OOM-killed or restarts, using the same rollback path and audit entry as health-based auto-rollback. Containers with a healthcheck and installs that leave it unset behave as before. - Rejecting an update from the approvals queue can now skip a whole minor or major line instead of only the exact tag. A scope picker beside Reject offers this version only, the whole minor line or the whole major line, and the API takes
scopeon the reject routes and on the containerskip-currentaction. Candidate selection leaves the rejected line out, so older patches and the next newer line are still offered. Line scopes need a semver tag update; digest updates and non-semver tags keep the exact skip. - OIDC sign-in can be restricted with
ALLOWEDEMAILS,ALLOWEDEMAILDOMAINSandREQUIREDGROUP(claim name set byGROUPCLAIM, defaultgroups, array or string values). Email matching is case-insensitive and rejects an address when the IdP reportsemail_verifiedas false. A denied user gets a403and no session. Unset, behavior is unchanged. - Docker watchers can now include or exclude containers by name or label with
DD_WATCHER_{name}_CONTAINERS_INCLUDE,_CONTAINERS_EXCLUDE,_LABELS_INCLUDEand_LABELS_EXCLUDE. Patterns use RE2, exclude wins over include, a container'sdd.watchlabel still overrides the filters, and an invalid pattern stops startup with a clear message. Filtered containers are skipped before any registry lookup. - Container state alerts. Four new notification rules,
container-stopped,container-exited,container-oom-killedandcontainer-restart-loop, report when a watched container stops, exits non-zero, is killed for running out of memory, or crashes repeatedly (3 times in 300 seconds by default, set per watcher withRESTARTLOOP_THRESHOLDandRESTARTLOOP_WINDOW, where the window is in seconds). They use the same routing, trigger lists, template overrides anddd.notification.excludeopt-out as the other rules, and templates getevent.exitCode,event.reason,event.restartCountandevent.windowSeconds. All four are off by default. One crash sends one alert, and stops or recreates Drydock does during an update or rollback never raise one. They go to notification triggers only, so an empty trigger list means all notification triggers and does not run Command actions. - Alert breadth. Two new notification rules,
auth-abuseandagent-version-mismatch, both off by default.auth-abusefires once when failed logins (passwords or second-factor codes) lock out an account or a source address, using the existing lockout thresholds, and names the same username the audit log already records, never a password.agent-version-mismatchfires once per agent when it connects on a different major or minor version than the controller (setDD_ALERTS_AGENTVERSIONPATCH=trueto include patch differences) and clears when it reconnects matching. - Quiet hours and silences for notifications.
DD_ALERTS_QUIETHOURS_SCHEDULE(HH:MM-HH:MMranges),_TZ,_RULESand_EXEMPThold notifications back during a window, andPOST /api/v1/notifications/silencesmutes a rule, or every rule, until an expiry you choose, withGETandDELETEto list and remove them. Held notifications are dropped and counted per rule, not queued, andsecurity-alertandupdate-failedare exempt by default. Creating and removing a silence is recorded in the audit log. Action triggers are never muted. - The audit log now records who did what. Each entry has an
actor, the username of the signed-in user orapi-key:<key name>for an API key, taken from the authenticated request on the server. Automated events such as auto-updates and schedulers have no actor and show as System, as do entries from before this release. The Audit view has an Actor column, a user filter and CSV and JSON export buttons. The API takesactoronGET /api/v1/auditand addsGET /api/v1/audit/exportwith the samereadscope. CSV cells that start with=,+,-or@are prefixed with an apostrophe so spreadsheets don't run them. Retention is now configurable withDD_AUDIT_RETENTION_DAYS(1 to 3650 days, default 30 as before). - Long security scans now show progress. Scanner and cosign commands log "still scanning {image} ({elapsed}s)" every 30 seconds until they finish, fail or time out, and a container shows a Scanning chip in the list while a scan of it runs, whether it was started by the schedule, the Scan button or Scan All. The state is held in memory only, clears on success, failure and timeout, and rides the existing container-updated event, so there is no new stream.
Fixed
- Release builds no longer run the UI build under emulation on arm64. The Dockerfile's
ui-buildstage now runs on the build host's own platform, since the static assets it produces are the same for every architecture, which fixes the multi-arch release build timing out while vite hung under QEMU. - The image entrypoint's notices (the
group_addGID line and the insecure-root warning) now go to stderr, soscanoutput on stdout stays valid JSON under those settings. - The OpenAPI schema for
PATCH /api/v1/containers/{id}/update-policynow lists thescopefield (tag,minorormajor) thatskip-currentalready accepted. - Portwing agents no longer get the "upgrade to 1.7.0" warning and Update agent badge, or an agent-version-mismatch alert. The minimum-version check now applies only to Drydock agents, since Portwing versions on its own line.
- Env reveal and the agent log view now answer Portwing agents with a clear 501 saying they are not supported, instead of an opaque failure, and the UI hides the env reveal button for containers on Portwing agents and the Logs tab on the agent detail panel. Container log viewing and download already work over the edge connection.
- Container list performance - The container store no longer re-parses and re-validates every stored container (including its vulnerability list) each time its query cache is rebuilt after a write; only rows that changed are re-read. The container list endpoint also stops running three update-operation queries per container when no update is active, the by-identity lookup only reads active operations, and a new migration indexes
update_operations.new_container_idso the remaining lookup no longer scans the table. - GHCR, LSCR and Docker Hub publish-date lookups are now shared per image and tag for each scan, so containers on the same image no longer repeat the GitHub package versions paging or the Docker Hub tag request, and each GHCR versions page now goes through the api.github.com rate limiter. This stops a large fleet from burning the anonymous GitHub API quota in one scan.
- The Home Assistant MQTT trigger no longer clones and redacts the whole container list for every sensor refresh when agent topic segments are on, which made startup discovery quadratic in the container count.
- The trigger docs no longer say notification rules only reach notification triggers. An empty trigger list on
container-unhealthyalso runs Command triggers, and the docs now say so. The container state alerts are the exception and go to notification triggers only. - A Portwing edge agent's over-cap response now fails with a clear error instead of a generic agent failure. Buffered requests and legacy container log replies that exceed Portwing's 16 MiB cap come back as
response_too_large; the log download endpoint answers 413 for it, and other buffered requests report that the response was too large and to use a streaming request. Older Portwing versions that don't send the code behave as before. - A
dd.tag.includeordd.tag.excluderegex that fails to compile no longer widens the candidate tags. Before, the bad filter was skipped with a log warning and the container could be offered tags the filter was meant to block. Now it gets no update candidates and shows an error naming the pattern until you fix it. - Container state alerts (
container-stopped,container-exited,container-oom-killedandcontainer-restart-loop) now go to notification triggers only. With an empty trigger list the rule means all notification triggers, but a Command action trigger was also running its command for these alerts. Command and the update actions are skipped for them now, as the docs already said. - Reloading the configuration from the API or the config editor no longer attributes later scheduled scans and auto-updates to the admin who triggered the reload. Components re-registered by a reload now run outside the request's audit actor, so their cron jobs and timers record as System, not as that user until the next restart.
- An invalid watcher setting, such as a bad
CONTAINERS_INCLUDEorLABELS_EXCLUDEpattern or aRESTARTLOOP_THRESHOLDbelow 2, now stops startup with a clear message on installs that use environment variables only, the same as it already did with adrydock.yml. Before, those installs logged "Some watchers failed to register" and kept running without the watcher. - Digest-only tag comparison now ignores include and exclude tag patterns longer than 1024 characters, the same as normal tag filtering. Before, one overlong label could give two different tag sets. The three filter compilers now share one implementation.
- Portainer triggers are now treated as update actions in the UI. The Notifications page no longer offers them as notification triggers (the server rejected them), and a Portainer trigger in dry-run mode now shows the preview label and badge, with its Run button disabled in notify mode like the Docker triggers.
- The full-page container view now shows the reason when revealing a masked environment variable fails, the same as the side panel.
- A failed Command action now reports the failure instead of resolving. Before, a command that exited non-zero or timed out was recorded as delivered, so under the default
oncesetting it was never retried, and a manual run returned success. The error names the exit code or timeout and never includes the command string. - Registry bearer authentication requests now use the configured outbound HTTP timeout (
DD_OUTBOUND_HTTP_TIMEOUT_MS, 30 seconds by default). Before, a registry or token realm that accepted the connection and never answered left the lookup hanging. - A same-tag digest rebuild no longer shows a skip-tag hold or confirmation when the tag was on the skip list. The skip-tag blocker now applies to tag updates only, matching what actually suppresses an update.
- An approved update that failed now goes back to the approvals queue on the next watch cycle while its candidate is still on offer, instead of staying in Decided until retention pruned it. Rolled-back and applied approvals stay decided.
- Approvals bind a decision to the candidate, not to its ref spelling - a row minted while the digest lookup failed can be approved, rejected or deferred once the digest returns, and a rejected candidate is no longer requeued and re-notified after one failed digest lookup. Two different known digests still never match.
- Cancelling a queued update now sticks. The cancelled operation is no longer reopened and run when its turn comes up, and a cancelled entry is skipped at dispatch.
- A container with
dd.depends_on.action=restartand no update, requested on its own or without its upstream in the batch, is now rejected with "No update available" instead of being pulled and recreated. - In a dependency group update, a member that simply has no update no longer rejects every container that depends on it. A restart-only dependent still needs an upstream that will actually run.
- Signing in from the UI with a non-ASCII password (for example
pässwordor秘密) works. The Basic credentials are now sent as UTF-8 bytes, matching what the server decodes, instead of failing in the browser or arriving as Latin-1. - The login page no longer shows Connection Lost when the server answers 429 (rate limited). It shows how long to wait, from the
Retry-Afterheader, and retries when that time is up. Real network failures still show the connection-lost overlay. (#1592) - The Docker Compose action keeps the
mac_addressyou configure per network (and at service level) when it recreates a container. Drydock now reads the explicit address from the Compose service definition, including override files, and pins it on the matching network. Generated MAC addresses are still never copied, and a secondary network that pins a MAC is attached at create time because network connect ignores it on older Docker APIs. (#1581) - A failed digest lookup now stores its error on the container and keeps the last good verdict, including a digest update found earlier, instead of showing the container as up to date. The error was being set on a copy the watcher threw away. (#1590)
- Docker Hub images keep their
hub.publicprovider across repeated image normalization, because the Hub provider now also matches its own canonicalhttps://registry-1.docker.io/v2URL. An image whose registry matches no configured provider now records a "no configured registry matches" error instead of looking up to date. (#1591) - After an update from the UI or API, the watcher's fast resync now matches the bare container name, looks up the recreated container by name and refreshes just that container, so the row no longer keeps the removed container's id and old tag when
WATCHEVENTSis false. (#1595) - A liveness rollback now starts the container it restores. It used to recreate the container from the dead replacement's state, so the service stayed down while the audit log recorded a successful rollback. Rollbacks restore the running state the container had before the update and only record success once the restored container is running, otherwise the audit entry records a failure with the reason. A container that was stopped before the update is no longer watched, so updating a stopped container with
LIVENESSWINDOWset no longer gets "rolled back" on the first poll. - Automatic rollbacks after a health or liveness failure and manual rollbacks now remember the candidate they undid, so the "last update rolled back" check blocks the same bad update instead of re-applying it once the recently-applied window passes.
- The Portainer action no longer runs the liveness watch. It rolled back through Docker directly and bypassed the stack.
LIVENESSWINDOWis rejected on a Portainer action anddd.rollback.livenessis ignored with a warning for containers it updates. - Rollbacks now take the same per-container lock updates do, and starting an update or a manual rollback cancels the health or liveness watch still running for that container, so an old watch can't undo a newer update.
- Rollbacks are recorded as operations, so container state alerts no longer report a stop for the container Drydock itself is stopping to roll back. The manual rollback route also honors
DD_SERVER_FEATURE_CONTAINERACTIONSlike the other container write routes and returns 403 when container actions are disabled. - The webhook credential check now treats
/UPDATE/xand/Watchthe same as the lowercase paths. Before, with per-endpoint webhook credentials set, the shared webhook credential still worked on a differently cased path because Express routes case-insensitively and the credential selection did not. - Running an action trigger through
POST /api/v1/containers/:id/triggers/:type/:nameor the trigger routes now returns 403 whenDD_SERVER_FEATURE_CONTAINERACTIONSisfalse, like the other update paths. Notification triggers can still be test-fired. - The MQTT trigger now closes its connection to the previous broker when it re-initializes (for example after the broker URL changes or the broker is replaced) and when it is deregistered. Before, the old client kept reconnecting in the background next to the new one.
- The container detail security tab no longer shows a previously selected container's vulnerabilities or SBOM when responses arrive out of order. Switching containers, clearing the selection or changing the SBOM format now discards stale responses, and only the latest request updates results, errors and loading state.
- HTTP trigger proxy credentials in the proxy URL are now sent to the proxy. Previously only the host and port were used, so an authenticated proxy rejected the request.
Security
- The same-origin guard on state-changing API requests now also covers requests that carry no session cookie. On installs with anonymous access enabled, a page on another site could send a body-less POST such as
/api/v1/config/reloador/api/v1/containers/watchbecause the browser sent no cookie and the guard skipped its checks. Cross-site Fetch Metadata, and any Origin or Referer that doesn't match, are now refused with 403 whether or not a cookie is present. CLI and integration clients that send none of those headers are unaffected. - Self-updates and infrastructure updates now run the same security gates as ordinary updates. Before, they verified the signature against the mutable tag, skipped the configured vulnerability scan entirely, and then pulled the tag again, so a registry could serve different content at the check than at the pull. Drydock now pulls the image first, pins its digest, verifies the signature and runs the scan gate against that digest, and creates the replacement from the pinned image, all before any hook, rename or helper container. A failed check aborts the update and leaves the running container alone.
- HTTP trigger proxy URLs that carry credentials (for example
http://user:pass@proxy.local:3128) no longer appear in registration logs, trigger API responses, debug dumps or the configuration endpoint. The userinfo is replaced with[REDACTED]where the proxy URL is shown, and the lower-scope trigger views redact the whole value. A value that cannot be parsed as a URL is redacted entirely. - Classify four Alpine OpenSSL 3.5.9-r0 advisories (CVE-2026-107496, CVE-2026-88823, CVE-2026-95691, CVE-2026-95692) as
not_affectedthrough an OpenVEX document. The affected CMS encryption, AES-GCM-SIV and QUIC client code is not reachable in Drydock, and Node bundles its own OpenSSL. This stays in place until Alpine ships a fixed openssl.
Documentation
- The getting started guide now says a fresh install starts in
manualupdate mode and that action triggers default toAUTO=oninclude, so containers need add.action.includelabel before they auto-update. The combined example notes the socket proxy needsPOST=1andNETWORKS=1, and the quickstart tells you to change the defaultadmin/passwordafter your first login.