May builtin/traefik and builtin/keycloak drop securityopt? (2026-08-11)
Superseded by New-line runtime. The text below is the record as of its date.
Status: RULED 2026-09-25. Keep the guard, fix the built-ins, change nothing
on compose. None of A, B or C as written. The swarm render leaves the key out
and every other runtime keeps it: addons/keycloak and addons/traefik wrap
security_opt in {% if shc.runtime != "swarm" %}, where shc.runtime is the
same EffectiveRuntime(opts.Runtime) the swarm loader gate decides on (ce
shcTemplateNamespace). Swarm runs what it always ran, since it dropped the key
anyway. Compose keeps the waiver. addons/postgres got the same treatment for
the server’s restart: unless-stopped, which conflicted with its swarm-only
restart_policy: {condition: none}. The floor below does not move: the catalogue
sweep neutralises Jinja by keeping the first branch, so it still counts both
built-ins as carriers (re-measured at ruling time: ten packages, floor nine).
Two consequences of the ruling, recorded here because nothing else says them:
- keycloak’s tmpfs is runtime-split too. The long form listed under “What
is already done” below mounts the same thing but changes the compose
container config, so it would have recreated every live Keycloak on its
next update. Compose keeps
tmpfs: [/tmp]; the long-form volume is emitted only on swarm. The shc store’sTestAddonComposeRenderMatchesTheGoldencompares the whole non-swarm render of every built-in against goldens rendered from main (cemodules/app/testdata/addon-compose-renders), so a change like that fails by name. - A swarm restore of a built-in from an archive older than this ruling is
refused. Recover renders the archived app tree, which still carries
security_optunconditionally, and the swarm loader gate refuses it with X400073SWMXLT before anything is written. Accepted: the built-ins are the compose-runtime system stacks on both estates (not re-checked on the live engines at ruling time), so no such archive should exist, and the refusal names the key and the way out (--runtime compose). Before the guard the same restore deployed green without the exemption. Option C stays open as the way to actually give swarm traefik and keycloak the exemption.
Status: OPEN. Owner ruling required. It blocks a landing.
Decision. builtin/traefik and builtin/keycloak each declare
security_opt: [apparmor:unconfined]. docker stack deploy loads that key,
prints it back, and never converts it, so on a swarm-runtime estate the
exemption is silently absent. ce’s swarm guard now refuses rather than deploy
green-and-wrong, and that refusal is what stops both system built-ins from being
installable on swarm at all. Three ways out; all three change something the owner
owns. Ruling wanted on which.
Cost of delay: blocks-work: lane/20260811-swarm-silent-discards is the
last of nineteen lanes not landed, and this is the only thing still holding it
(everything else in that lane is fixed on fix/swarm-discards-deadlock, ce). It
is also latent-security: nothing is wrong on any estate today, because every
live deployment runs --runtime compose.
What is actually true (measured, not read off docs)
Section titled “What is actually true (measured, not read off docs)”1. docker stack deploy drops the key in silence, and the preflight cannot
see it. docker stack config returns exit 0, echoes security_opt back
verbatim, and writes zero bytes to stderr, because stack config never
calls loadComposeFile, which is where the “Ignoring unsupported options”
warning lives (docker/cli cli/command/stack/loader.go:42-45 vs
cli/command/stack/config.go). So the existing stack-config preflight
(X500062SWMCFG) is structurally blind to this whole class; more preflight would
not have found it.
2. The compose→swarm converter genuinely never reads it. security_opt is in
docker’s own composetypes.UnsupportedProperties
(cli/compose/types/types.go:17-34, sixteen entries at v29.4.1), and nothing in
cli/compose/convert/ reads service.SecurityOpt. Two independent derivations,
field-usage over ServiceConfig, and docker’s own declaration, agree.
3. ★ BUT SWARM ITSELF CAN EXPRESS IT. This corrects the framing the lane was
filed under (“swarm has NO security_opt equivalent at all, so there is no
translation”). That is true of docker stack deploy; it is not true of
swarm. swarm.ContainerSpec.Privileges carries both:
type Privileges struct { CredentialSpec *CredentialSpec SELinuxContext *SELinuxContext Seccomp *SeccompOpts `json:",omitempty"` AppArmor *AppArmorOpts `json:",omitempty"` NoNewPrivileges bool}const ( AppArmorModeDefault AppArmorMode = "default" AppArmorModeDisabled AppArmorMode = "disabled")(github.com/docker/[email protected] api/types/swarm/container.go:44-84.)
AppArmorMode: "disabled" is the swarm spelling of apparmor:unconfined.
Custom AppArmor profiles are not supported, but neither built-in wants one;
both want the profile off.
What is missing is only the WIRE-UP: docker stack deploy will not put it there,
and docker service create --help at 29.4.1 offers no --security-opt,
--apparmor or --seccomp flag. The field is reachable only from the daemon
API.
4. Both exemptions are load-bearing, and they are load-bearing for different reasons.
builtin/traefik/compose.yaml:499-501: “Traefik needs unrestricted apparmor for Docker socket access and dynamic config reloads”. It mounts/var/run/docker.sock:/var/run/docker.sock:ro(:590). Dropped, the task runs under docker’s default apparmor profile and the failure surfaces as an unexplained permission denial on the socket, i.e. ingress comes up and does not work.builtin/keycloak/compose.yaml:39-41: “Keycloak JVM uses native memory access (Quarkus/Netty direct buffers)”. Dropped, the failure is a JVM native-memory error at some later point, i.e. SSO comes up and does not work.
5. This host has AppArmor enabled (/sys/module/apparmor/parameters/enabled
→ Y), so the key is not inert here. On a host with no AppArmor LSM the key is
a no-op and dropping it costs nothing; which estates are which is not
something this write-up measured, and it is worth one command per estate before
ruling (docker info | grep -i apparmor).
6. Removing the key changes Compose-runtime behaviour on every existing
estate. 24 deployments run on the compose runtime across two estates. compose
reads security_opt; swarm is what ignores it. So “drop it for swarm’s benefit”
is not a swarm-only edit; it re-confines traefik and keycloak on every live
compose deployment, today.
Why this reached the owner rather than being decided in the lane
Section titled “Why this reached the owner rather than being decided in the lane”The guard is not the problem, and neither is the built-in. Two landed guards state opposite post-conditions about the same two files:
TestBuiltinsCarryNoUntranslatableSwarmConstruct(ced04fb18): every rendered built-in must carry NOTHINGdocker stack deploydiscards. Its stake: “a built-in that swarm cannot load is a swarm estate with no ingress, no system store, or both.”- the silent-discard guard:
security_optis discarded, so it must be refused.
Cleaning the built-ins satisfies the first and (before this branch) broke the
second; leaving them satisfies the second and breaks the first. Every exit that
does not involve a ruling is a weakening of one of the two, and weakening
either reintroduces exactly the defect it was written to catch. The other three
collisions in that deadlock were mechanical and are fixed on
fix/swarm-discards-deadlock (see “What is already done”, below). This one is
not mechanical: it is a question about what the platform supports.
Options
Section titled “Options”A. Rule that swarm-loadable wins: delete security_opt from both built-ins
Section titled “A. Rule that swarm-loadable wins: delete security_opt from both built-ins”What it costs. Traefik loses unrestricted apparmor on compose too, on all 24 live deployments, against a mounted Docker socket. Keycloak’s Quarkus JVM loses it for native memory access. Both regressions are runtime and silent until exercised, exactly the class this whole seam exists to stop shipping. Neither is currently measured: nobody has run either image under docker’s default apparmor profile on these estates.
What it buys. Both guards go green, the lane lands, no new code.
Verdict: not recommended without measurement. This trades a latent swarm defect for a live compose regression, on the ingress and the SSO, to unblock a runtime no estate is on yet. If it is taken, it must be taken after running both images unconfined-free on a scratch estate, not before.
B. Rule that the refusal stands: keep security_opt, accept that the two system built-ins are compose-only
Section titled “B. Rule that the refusal stands: keep security_opt, accept that the two system built-ins are compose-only”What it costs. A swarm-runtime estate cannot install its own ingress or its
own SSO. That is the exact sentence
TestBuiltinsCarryNoUntranslatableSwarmConstruct exists to prevent, so that
guard has to be told about the exception explicitly (an allow-list of “these two
services may carry security_opt, and here is the ruling that says so”), or it
stays red forever.
★ An allow-list is a weakening of a fitness rule, and it is only honest if it
is (a) keyed to these two services by name, (b) carrying the ruling’s identifier
in the failure text, and (c) accompanied by the refusal still firing at install
time so an operator attempting --runtime swarm is told why, not merely
allowed to proceed into a broken estate. Written any looser, it is the
“allowlist a test to reach green” move this repository bans.
What it buys. No behaviour change anywhere. Honest about the limitation. Swarm support for the two system packages is deferred, not faked.
C. ★ Wire the property swarm actually has: post-deploy ServiceUpdate setting Privileges.AppArmor.Mode: disabled
Section titled “C. ★ Wire the property swarm actually has: post-deploy ServiceUpdate setting Privileges.AppArmor.Mode: disabled”The finding in (3) makes this a real option rather than a wish.
docker stack deploy cannot set it and docker service create has no flag, but
the daemon API field exists and means precisely what apparmor:unconfined means.
SHC already drives the daemon directly elsewhere.
What it costs. New code on the swarm install path, and a genuinely awkward ordering problem: the stack deploys with the default profile and is patched immediately after, so there is a window in which traefik is running confined against the socket. Whether that window is acceptable (traefik would fail its socket reads and retry) needs measuring. It is also the only option that adds a maintenance surface, a second way of expressing a security exemption, which must not drift from the compose one.
What it buys. The guarantee is kept on both runtimes rather than dropped on one or refused on both. It is the only option under which a swarm estate gets a working ingress AND a working SSO AND no compose regression.
Verdict: recommended to cost, at least. If the ServiceUpdate window is
tolerable, C is strictly better than A and B. If it is not, the fallback is B,
because B defers a capability and A regresses a live one.
What is already done, and what is not
Section titled “What is already done, and what is not”On ce branch fix/swarm-discards-deadlock (which merges
lane/20260811-swarm-silent-discards):
- Restart-equivalence carve-out.
restart:beside adeploy.restart_policyis only reported when the two keys say different things.no↔none,always↔anyandon-failure↔on-failure(bounds compared) are forgiven;unless-stoppedand every mismatch still fire. This was detector over-fire: traefik-init and postgres-guard declare both keys deliberately, and the ordering lane documented them as equivalent. Measured through docker’s own converter. - keycloak’s
tmpfsmoved fromtmpfs: [/tmp]tovolumes: [{type: tmpfs, target: /tmp}]. Docker’s converter producesmounts=[]for the first andmounts=[{"Type":"tmpfs","Target":"/tmp"}]for the second; under compose both produce a byte-identical writable tmpfs mount. One offence removed, no behaviour change. - The corpus guard’s by-name teeth re-sourced from a committed fixture
(
modules/app/testdata/swarm-discards/) instead of from the live built-ins, so the detector is proven without holding the real tree hostage. Empty-corpus and deny-list-completeness tripwires kept and strengthened.
Still red, and only this: service keycloak: security_opt: and
service traefik: security_opt:, on scenarios 0-2 of each, in
TestBuiltinsCarryNoUntranslatableSwarmConstruct and
TestBuiltinsSurviveTheSwarmInstallPipeline. Nothing else.
One thing the ruling must carry with it if it is A or C:
TestCatalogueSecurityOptCarriersAreRefusedUnderSwarm’s
securityOptPackageFloor = 9 is a measurement of the assembled catalogue, and
two of the nine are these built-ins. It has to be re-measured in the same change,
not lowered ahead of the ruling.