BI/analytics league pick: Apache Superset replaces Metabase's slot
Status: ACCEPTED AND EXECUTED 2026-08-14: flagship: analytics is
applied. The body below is the 2026-08-06 worksheet, preserved as written.
Whether and when apps/enterprise/metabase is retired from the catalogue is
still a separate owner decision; this document records the replacement pick,
the evidence, and (in §0) what was actually applied.
Date: 2026-08-06. Candidates verified against upstream repos/docs/images on this date (every claim below is cited; image-level claims were verified by running the pinned digest, not by reading docs).
0. EXECUTION RECORD: 2026-08-14
Section titled “0. EXECUTION RECORD: 2026-08-14”This document IS the “separately commissioned re-evaluation” that
2026-08-06-flagship-tier.md §7’s analytics
row was holding the league open for (“assess the re-evaluation’s pick against
all eight bars when it lands”). It landed; the assessment was made; the label
is applied.
What was applied:
flagship: analytics→apps/flagship/superset/app.yaml, with the eight-bar assessment transcribed into the manifest beside the key and pinned to the 6.1.0 image, so a bump that changes an answer has to come back through it. The league token isanalytics, notbi. The tier doc’s naming doctrine is that a league names the FUNCTION, andanalyticsis already in §2’s league set.git mv apps/integrated/superset apps/flagship/superset. Not a promotion on the four bucket gates: Superset publishes no socket, so the integration-depth reading still saysintegrated. It moves because a labelled app must live inapps/flagship/(TestFlagshipLabelLivesInFlagshipBucket), the same one-way implication that holds openreplay and uptime-kuma there. Recorded as such inflagshipMembersand indocs/app-development/app-buckets.md§5.- The label surface stopped being decorative. A new hard fitness check,
flagship-league-roster, pins every ruled league (analytics→ superset,errors→ glitchtip,replay→ openreplay,uptime→ uptime-kuma, andstorage→ VACANT) as a set equality with the catalogue. Deleting a declaration, moving it to another app, claiming the vacant league, or inventing an unruled token each fail BY NAME. Before it, theflagship-contractcheck validated only the claims that existed and reported clean over zero of them (half-done-sweep F9).
The honest gap, restated rather than buried: §3’s scorecard already marks
Superset ❌ on an auto-redirect knob, and the
2026-08-14 auto-redirect ruling
§2 says an SSO-capable flagship that cannot be armed is “a package gap to fix,
not an accepted state.” It is therefore OWED, not waived: FAB renders a
provider chooser even with exactly one provider, so arming corp/prod needs an
edge-flavour path or a superset_config.py view override. The label stands
because requirement 1’s floor is a visible, unpaid SSO button and that is met,
but the gap is on the books.
Not changed, deliberately: metabase keeps its apps/enterprise/ placement
and its requires_paid_license declaration. Retiring it is the separate owner
decision this document’s header has always named.
1. Why the slot is open
Section titled “1. Why the slot is open”Metabase is disqualified from the SSO-everywhere posture
(sso_coverage_policy): the OSS build has no SSO: the SSO/OIDC code
lives in the enterprise source tree under a proprietary license, so it cannot
be fork-enabled without violating the license, and (pending a parallel
verdict) there is no auth-off mode that would let the forward-auth gate carry
authentication instead. The catalogue keeps v0.x OSS tags pinned
(v1.x tags are Enterprise), so no packaging trick reaches SSO.
2. The five bars
Section titled “2. The five bars”- License: open source permitting modification + redistribution (no SSPL/BSL).
- SSO without a paid tier: native OIDC strongly preferred; auto-redirect knob ideal.
- Zero-touch: OIDC configurable via env/config-file/API so the package wires it with no admin-console click-ops.
- Actively maintained upstream.
- Capability parity with Metabase’s role: self-serve exploration + dashboards + SQL editor + sane permissions.
3. Scoring
Section titled “3. Scoring”| Bar | Superset | Lightdash | Redash | Grafana |
|---|---|---|---|---|
| 1. License | ✅ Apache-2.0, ASF project, no EE tree | ⚠️ MIT with packages/backend/src/ee carve-out under a separate license | ✅ BSD-2-Clause | ✅ AGPL-3.0 (copyleft, but modification+redistribution are exactly what AGPL grants) |
| 2. SSO w/o paid tier | ✅ flask-appbuilder AUTH_OAUTH, free; native keycloak provider handler; ❌ no auto-redirect knob (FAB renders a provider chooser even with one provider) | ✅ generic OIDC env vars (AUTH_OIDC_*) documented for self-hosted, not flagged enterprise | ❌ no generic OIDC; Google OAuth, SAML, LDAP, trusted-header only | ✅ generic OAuth free in OSS with auto_login knob |
| 3. Zero-touch | ✅ everything in superset_config.py + env; no admin-console step, no settings API to converge | ✅ env-var driven | ⚠️ SAML env-driven, but the SSO shape itself misses the bar | ✅ GF_AUTH_GENERIC_OAUTH_* env |
| 4. Maintained | ✅ 6.1.0 (2026-05-13); 74k★; continuous release trains | ✅ very active (Renovate-driven, 21k+ commits) | ⚠️ alive but sparse: v26.3.0 (2025-03-02) is the last release, ~17 months ago | ✅ extremely active |
| 5. Metabase-role parity | ✅ no-code explore + dashboards + SQL Lab + semantic layer + RBAC (Admin/Alpha/Gamma + row-level security) | ⚠️ semantic-layer BI: explores come from authored YAML models (dbt project or standalone Lightdash YAML); no point-at-a-database self-serve; SQL runner exists | ✅ query-first dashboards; weaker no-code explore | ❌ monitoring-first: per-datasource query editors, no self-serve business-data explore, no semantic layer |
| Verdict | WINNER | ❌ parity + ee-carve-out risk | ❌ SSO bar + cadence | ❌ parity bar |
4. Evidence per candidate
Section titled “4. Evidence per candidate”Apache Superset: winner
Section titled “Apache Superset: winner”- License: Apache-2.0, ASF top-level project; no enterprise/EE split in the repo (github.com/apache/superset).
- Maintained: latest release 6.1.0; PyPI lists it as current; Docker Hub pushed the 6.1.0 tags 2026-05-13. Release ladder 4.1.x → 5.0.0 (2025-06) → 6.1.0 shows a steady train.
- SSO:
AUTH_TYPE = AUTH_OAUTHvia flask-appbuilder, OSS, no tier. The image ships FAB 5.0.2 (verified in-container), whoseget_oauth_user_infonatively handles a provider namedkeycloak: username/email/name from userinfo,role_keysfrom thegroupsclaim (verified in the v5.0.2 source).AUTH_ROLES_MAPPING+AUTH_ROLES_SYNC_AT_LOGINgive claim-driven role sync;AUTH_USER_REGISTRATION_ROLEis always granted alongside mapped roles (FAB_oauth_calculate_user_roles), so a missing claim demotes rather than locks out: no dependency-track-style zero-teams lockout class. No auto-redirect knob: FAB’sAuthOAuthView.loginrenders the provider chooser even with exactly one provider (verified in v5.2.2 source; same view in 5.0.x). A config-file override could force it later (superset_config.pyis arbitrary Python), but the package does not smuggle custom view code in. - Zero-touch: whole surface is
superset_config.py(env-driven, rendered as a compose config) +SUPERSET_OIDC_*env vars the oidc integration materialises. The plug has no jobs; contrast windmill’s settings-API merge/verify dance. - The honest costs (all verified by running
apache/superset:6.1.0@sha256:fb34…):- The lean image ships no authlib and no psycopg2 (psycopg2-binary is
the
[postgres]extra; authlib is in no extra), and the uv-built venv has no pip. The package provisions both wheels (exact pins:authlib==1.6.12,psycopg2-binary==2.9.12, upstream’s own extra pin) from a digest-pinnedpython:3.10.20-sliminit container into a shared volume consumed viaPYTHONPATH. Needs pypi egress on FIRST deploy only (idempotent skip after); precedent: the dtrack oidc plug’sapk add jq. - Operational weight: metadata postgres (via the estate’s standard
postgres plug / separate
<stack>-pgstack) + an in-stack valkey for caching and rate limiting. Celery worker/beat are deliberately not shipped: they serve only async SQL Lab queries, alerts/reports and thumbnails; sync queries are the default and the metabase slot never had alerts wired either. Adding a worker later is additive, not a redesign. - Two one-shot services (
superset-pylibs-init,superset-init) run per deploy: more moving parts than metabase’s single container, though the authored memory quota is comparable (2G web + 256M cache + 512M sidecar- 1G transient init vs metabase’s 4G + 512M sidecar).
- The lean image ships no authlib and no psycopg2 (psycopg2-binary is
the
Lightdash
Section titled “Lightdash”- License: MIT with a carve-out:
packages/backend/src/eeis under a separate license (verified in the repo LICENSE text). “Enterprise features are available on self-hosted deployments with a license key.” - SSO: self-hosted docs list Okta/Google/OneLogin/AzureAD and generic
OIDC (
AUTH_OIDC_CLIENT_ID,AUTH_OIDC_METADATA_DOCUMENT_URL, …), not flagged enterprise-only. - Why it loses, parity: the README’s own framing: “Use dbt projects or standalone Lightdash YAML pointed at your warehouse.” Every explore is an authored semantic-layer model; there is no Metabase-style point-at-a-DB self-serve browsing. Excellent tool, different league. The ee/ tree also makes “which parts are paid” a moving boundary that Apache-2.0 simply does not have.
Redash
Section titled “Redash”- License: BSD-2-Clause.
- SSO: settings source (master) carries Google OAuth, SAML, LDAP and
trusted-header (
REMOTE_USER_LOGIN_ENABLED); no generic OIDC. SAML against Keycloak would work, but the bar says native OIDC strongly preferred, and header-auth would force the forward-auth-gated posture that native-SSO apps deliberately avoid. - Cadence: last release v26.3.0 on 2025-03-02 (~17 months before this writing); the post-revival cadence is a few releases per year at best.
Grafana
Section titled “Grafana”- License: AGPL-3.0, genuinely open source; modification and redistribution permitted (copyleft obligations are acceptable for catalogue packaging: SHC redistributes unmodified upstream images).
- SSO: generic OAuth in OSS, env-driven (
GF_AUTH_GENERIC_OAUTH_*), includingauto_login, the only candidate with the ideal knob. - Why it loses, parity: Grafana is monitoring-first. No self-serve business-data exploration, no semantic layer, no pivot-style explore for non-SQL users; the metabase slot is business BI, and the estate already holds the observability lane elsewhere.
5. What the package does (summary)
Section titled “5. What the package does (summary)”apps/superset/: digest-pinned apache/superset:6.1.0; postgres via the
standard plug (separate <stack>-pg stack, dtrack/windmill pattern);
in-stack valkey (superset-cache, FQDN-addressed per the paperless NOAUTH
lesson); superset-pylibs-init wheel provisioner; superset-init runs
db upgrade + fab create-admin (convergence-checked, exit codes are
unreliable upstream) + superset init before the web tier starts;
oidc.plug.yaml mints a confidential client with callback
/oauth-authorized/keycloak and no jobs (env-only integration);
stack-scoped backup.yaml (metadata lives in the separate postgres stack;
group backup via --all-connected); native-SSO access posture: not
forward-auth gated (huly posture).
Default SSO role posture mirrors dependency-track: auto-provisioned SSO users
get Admin (vars.keycloak.registrationRole) because the corp realm mints
no groups claim; least-privilege operators set Alpha/Gamma or wire
vars.keycloak.rolesMapping + a hand-minted groups mapper.
6. Verified vs assumed
Section titled “6. Verified vs assumed”Verified (fetched or executed 2026-08-06):
- Superset: Apache-2.0; 6.1.0 latest (PyPI + Docker Hub, digest recorded);
in-image facts (python 3.10.20, FAB 5.0.2, no authlib, no psycopg2, no
pip, curl present,
AUTH_OAUTHimportable, gunicorn defaults, port 8088) by running the pinned digest. - FAB keycloak handler +
role_keys←groups(v5.0.2 and v5.2.2 sources); no single-provider auto-redirect (v5.2.2views.py); callback route/oauth-authorized/<provider>. - Lightdash: MIT + ee carve-out (LICENSE); OIDC env vars in self-hosted docs; “dbt projects or standalone Lightdash YAML” positioning (README).
- Redash: BSD-2; release dates (GitHub releases); auth options incl. absence
of generic OIDC (master
redash/settings/__init__.py). - Grafana: generic OAuth availability in OSS +
auto_login(official docs). - authlib 1.6.12 latest (PyPI); psycopg2-binary 2.9.12 = superset’s own
[postgres]extra pin (pyproject.toml).
Assumed (explicitly not proven in this lane):
- Lightdash SSO’s freedom from the ee license was read off the docs’ silence, not the ee tree’s file list.
- Metabase’s “no auth-off mode” is pending the parallel verdict named in the tasking; nothing here depends on it.
- The end-to-end Keycloak login round-trip for the packaged app was verified at render/source level (env → config → FAB handler → callback route), not against a live keycloak; the bats suite deploys with postgres only, per catalogue convention.
- Grafana’s AGPL was not re-fetched this date (stable, widely known); all other license claims were fetched.