Close search's active-tenant write-routing gap with a polled allowlist

search/src/consumer.rs's write-routing (built last pass) had no active-
tenant check at all: IndexRegistry.resolve() would open-or-create an
index directory for any syntactically-valid tenant_id, active or not --
unlike ClickHouse's chwriter.Registry (an active-tenants-only snapshot
built at enterprise-ingest startup) or the read side (gated by
searchclient.TenantChecker, a direct rbacstore query). search is AGPL
core with no Postgres access and no enterprise/ import allowed, so it
needed a network boundary instead -- the same shape ingest's
TenantResolver already uses against enterprise-auth, just Rust calling
Go instead of Go calling Go.

New GET /internal/active-tenants endpoint on enterprise-auth
(rbacstore.ListActiveTenantIDs + authhandler.handleActiveTenants),
gated on a RoleService Bearer credential -- server-to-server auth, the
same shape alerting presents to api, minted via the already-generic
enterprise-auth -mint-service-token search. search/src/tenants.rs's
ActiveTenantTracker polls it every 60s, blocking startup on the first
fetch succeeding (fail-closed cold start -- a control-plane outage at
boot must not silently accept every tenant_id) and keeping the last-
known-good set on any later refresh failure (a transient blip shouldn't
stop every tenant's indexing, only prevent the allowlist from growing/
shrinking until connectivity resumes). consumer.rs refuses any tagged
record whose tenant isn't in the polled set, before ever calling
resolve() -- IndexRegistry itself stays policy-free, matching the same
mechanism/policy split clickhousewriter.Writer vs. chwriter.Registry
already draws on the ClickHouse side.

Off unless ENTERPRISE_AUTH_URL/ENTERPRISE_AUTH_SERVICE_TOKEN are both
set (search/src/config.rs rejects exactly one being set) -- every
existing deployment is unaffected.

Verified with real HTTP round trips in this environment: tenants.rs's
tests exercise real reqwest requests (actual Authorization: Bearer
header, actual JSON parsing) against a hand-rolled dependency-free TCP
test server, including both fail-closed paths (rejected first fetch,
unreachable server). authhandler's new tests cover the credential-kind
distinction this endpoint exists to enforce -- a real human session,
even for a genuine Owner, must not satisfy a check meant for a service
identity.

One asymmetry remains, disclosed rather than fixed: chwriter.Registry's
snapshot still never refreshes (stale until enterprise-ingest restarts),
while ActiveTenantTracker's 60s poll gives Tantivy a materially tighter
staleness window. Neither is a live per-write check -- that would mean
a database/HTTP round trip per record, a throughput cost neither
implementation accepts -- so both have some staleness window by design;
the gap between the two windows is what's disclosed, not a claim either
is fully live.
This commit is contained in:
2026-08-14 23:47:25 -07:00
parent 5a845f06ee
commit 088677643f
19 changed files with 1139 additions and 141 deletions
+15 -11
View File
@@ -171,17 +171,21 @@ escape hatch is opaque to any compiler-injected filter.
index. No "second binary" needed here, unlike ClickHouse — Tantivy has
no grant system to gate a commercially-licensed credential behind, so
`IndexRegistry` already lived directly in this AGPL-core binary, and
read/write just share it. One gap is disclosed rather than closed by
this change: unlike `chwriter.Registry` (an active-tenants-only
snapshot built at startup) and unlike the read side (gated by
`searchclient.TenantChecker`), this consumer's `resolve()` call has no
active-tenant check — the process has no Postgres access to check
against — so a still-valid-but-should-be-revoked ingest credential can
cause an index directory to be created for a tenant that's no longer
active. Narrow blast radius (an orphan, isolated, empty index, not
cross-tenant leakage, and only reachable with a real signed
credential), but real; see `search/src/registry.rs`'s doc comment on
`resolve`.
read/write just share it. The active-tenant gap this design left open
is now closed too: `search/src/tenants.rs`'s `ActiveTenantTracker`
polls a new `GET /internal/active-tenants` endpoint on
`enterprise-auth``search` has no Postgres access, so unlike
`chwriter.Registry`'s direct `rbacstore` query or the read side's
`searchclient.TenantChecker`, this needed a network call instead (the
same "network boundary, not import boundary" shape `ingest`'s
`TenantResolver` already uses against the same service, authenticated
with a RoleService credential like `alerting``api`) — and
`consumer.rs` refuses any tagged record whose tenant isn't in the
polled allowlist. Off unless configured, fail-closed on the first
fetch, last-known-good on later refresh failures; see
`search/src/tenants.rs`'s doc comment for the full design and
`search/src/registry.rs`'s for how mechanism (index lifecycle) and
policy (the gate) responsibilities split.
- `deploy/operator`'s `Tenant` CRD and `enterprise-api -provision-tenant`
are now unified, deliberately lightweight: `-provision-tenant` stays
the sole real actor (ClickHouse + `rbacstore`), and now also syncs its
+68 -29
View File
@@ -39,14 +39,18 @@ of what was already run and passed. Two genuine exceptions:
OIDC: not yet tried against a real external IdP or a running
`enterprise-auth` container.
- Tantivy tenant isolation, both directions -- `search/src/registry.rs`'s
cross-tenant read isolation (§9) and, since this pass,
`search/src/consumer.rs`'s per-tenant write-routing (§14) -- verified
live, no disclaimer needed, because Tantivy is an embedded library with
no Docker/broker dependency: real indices, real documents, real
commits, run in this environment. The one thing about it that's still
unverified is not Tantivy itself but the upstream credential/header
plumbing feeding it (ingest's `TenantResolver`, `enterprise-auth`'s
`/internal/authorize-ingest`) against a real running stack.
cross-tenant read isolation (§9), `search/src/consumer.rs`'s per-tenant
write-routing (§14), and `search/src/tenants.rs`'s active-tenant gate
(§14) -- verified live, no disclaimer needed, because Tantivy is an
embedded library with no Docker/broker dependency, and the active-
tenant gate's only external dependency (`enterprise-auth`'s HTTP API)
was exercised against a real hand-rolled test server, not a live
container: real indices, real documents, real commits, real HTTP
requests, all run in this environment. The one thing about it that's
still unverified is not Tantivy itself but the upstream credential/
header plumbing feeding it (ingest's `TenantResolver`, `enterprise-auth`'s
`/internal/authorize-ingest` and `/internal/active-tenants`) against a
real running stack.
- The tenant-picker frontend page (§12) -- the first frontend-only piece
in this phase exercised in a real browser rather than only
type-checked: `web/src/routes/select-tenant`'s cross-origin
@@ -746,21 +750,54 @@ cargo test --quiet
# ingest/cmd/ingest's own guard test does on the Go side.
```
**What's still not built, for either engine**: a live active-tenant
recheck at write time. `chwriter.Registry`'s per-tenant writer map is a
snapshot built once at `enterprise-ingest` startup from
`rbacstore.ListProvisionedDataSources` (active tenants only) -- an
unrecognized `tenant_id` is refused outright, but a tenant deprovisioned
*after* startup keeps writing successfully until the next restart.
`IndexRegistry.resolve()` has no allowlist at all on either the read or
write side -- `search` has no Postgres access to check tenant status
against -- so a still-valid-but-should-be-revoked ingest credential can
cause an orphan Tantivy index directory to be created for a tenant
that's no longer active. Narrow blast radius either way (isolated, not
cross-tenant leakage, reachable only with a real signed credential), but
real and disclosed, not silently accepted -- see
`search/src/registry.rs`'s doc comment on `resolve` and
`/docs/security/threat-model.md`'s "Read this first".
**Tantivy's write path is now active-tenant-gated too.**
`search/src/tenants.rs`'s `ActiveTenantTracker` polls a new
`GET /internal/active-tenants` endpoint on `enterprise-auth` (Go side:
`rbacstore.ListActiveTenantIDs` + `authhandler.handleActiveTenants`,
RoleService-credentialed -- mint one with
`enterprise-auth -mint-service-token search`, the same generic flag
`alerting` already uses, just a different subject name) and
`consumer.rs` refuses any tagged record whose tenant isn't in the polled
set. Off unless `ENTERPRISE_AUTH_URL`/`ENTERPRISE_AUTH_SERVICE_TOKEN` are
both set (`search/src/config.rs` rejects exactly one being set); startup
blocks on the first fetch succeeding (fail-closed cold start), and a
later refresh failure keeps serving the last-known-good set rather than
clearing it. Genuinely verified in this environment, no live
enterprise-auth needed:
```sh
cd search
cargo test --quiet tenants::
# start_fetches_and_serves_the_initial_list / start_sends_the_bearer_token
# run against a real (hand-rolled, dependency-free) TCP server -- actual
# reqwest request construction (the Authorization: Bearer header, the
# /internal/active-tenants path) and actual JSON response parsing, not a
# fake HTTP client. start_fails_closed_when_the_first_fetch_fails and
# start_fails_closed_when_the_server_is_unreachable prove the cold-start
# refusal: ActiveTenantTracker::start returns Err rather than falling
# back to an empty (accept-nothing, silently-safe-looking-but-wrong-for-
# operators) or permissive (accept-anything, the exact bug being closed)
# default.
go test ./internal/authhandler/... -run TestActiveTenants -v
# GET /internal/active-tenants requires a RoleService credential -- a
# valid human session (even for a real Owner) is rejected the same way
# an invalid/missing token is, the regression test for this endpoint's
# whole reason to distinguish token kinds.
```
**One asymmetry remains, disclosed rather than fixed**: `chwriter.
Registry`'s per-tenant writer map is a snapshot built once at
`enterprise-ingest` startup, never refreshed -- a tenant deprovisioned
after startup keeps writing successfully to ClickHouse until the next
restart, a real staleness window `ActiveTenantTracker`'s 60-second
refresh doesn't have on the Tantivy side. Neither is a live per-write
check (a database/HTTP round trip on every record would be a real
throughput cost neither implementation accepts), so both have *some*
staleness window by design -- the gap between the two windows is what's
disclosed as inconsistent here, not a claim that either is fully live.
Closing it would mean adding a periodic refresh to `chwriter.Registry`
too; not done in this pass. See `/docs/security/threat-model.md`'s "Read
this first".
**Also not built**: Helm/`docker-compose.yml` do gate *whether*
`enterprise-ingest` runs at all (`ingest.requireTenantCredential`, same
@@ -811,12 +848,14 @@ Full accounting: `/docs/security/threat-model.md`. Headline items:
see §14). `search/src/consumer.rs` reads the same header and routes
each record into its own tenant's Tantivy index -- genuinely verified
in this environment, unlike the ClickHouse side, since Tantivy needs
no Docker to exercise real logic. Both engines share one remaining,
disclosed gap: neither write path rechecks tenant-active status live
(ClickHouse: a startup-time snapshot, stale until restart; Tantivy: no
allowlist at all, since `search` has no Postgres access) -- narrow
blast radius, not cross-tenant leakage, but real; see §14 and
`/docs/security/threat-model.md`'s "Read this first". A
no Docker to exercise real logic. Both engines now also gate writes on
an active-tenant check: ClickHouse's is a startup-time snapshot with
no refresh (stale until the next `enterprise-ingest` restart), while
Tantivy's `ActiveTenantTracker` polls `enterprise-auth` every 60
seconds, a tighter staleness bound the ClickHouse side wasn't updated
to match -- a real, disclosed asymmetry between the two, not a gap in
either alone; see §14 and `/docs/security/threat-model.md`'s "Read
this first". A
newly-provisioned tenant's ClickHouse database and Tantivy index are
both now real, isolated, and actually populated by write-routed
traffic (the ClickHouse claim pending live confirmation, the Tantivy
+49 -33
View File
@@ -9,21 +9,27 @@ for the full design rationale behind the controls described here.
## Read this first: the single most important open finding
**Updated a fourth time.** This section originally read "log data
queried through `POST /query` is not tenant-isolated at all," then
"ClickHouse is isolated but Tantivy isn't," then "ingest tags records
with a tenant identity but nothing routes the write," then "ClickHouse
write-routing is built but Tantivy's isn't." Both storage engines are
now isolated on both the read and write paths. What's left is narrower:
**whether a given deployment actually runs the isolated binaries**
(deployment-time, not code-level), and two disclosed write-side gaps,
different in kind: ClickHouse's write registry is a startup-time
snapshot of active tenants with no live recheck (a *deprovisioned*
tenant can keep writing successfully until the next `enterprise-ingest`
restart), while Tantivy's write path has no active-tenant allowlist at
all — it opens an index for *any* syntactically-valid `tenant_id` a
record carries, active, deprovisioned, or never real. See below for
both, in full.
**Updated a fifth time.** This section originally read "log data queried
through `POST /query` is not tenant-isolated at all," then "ClickHouse
is isolated but Tantivy isn't," then "ingest tags records with a tenant
identity but nothing routes the write," then "ClickHouse write-routing
is built but Tantivy's isn't," then "both are write-routed but neither
rechecks tenant-active status live." Both storage engines are now
isolated on both the read and write paths, **and both now gate writes
on an active-tenant check** — ClickHouse's (`chwriter.Registry`) and
Tantivy's (`search/src/tenants.rs`'s `ActiveTenantTracker`, new) differ
in staleness bound, not in whether the gate exists at all: ClickHouse's
is a startup-time snapshot with no refresh (a *deprovisioned* tenant can
keep writing successfully until the next `enterprise-ingest` restart);
Tantivy's polls `enterprise-auth` every 60 seconds and keeps serving the
last-known-good set through a transient refresh failure, a materially
tighter staleness window with no code changed on the ClickHouse side to
match it — a real, disclosed asymmetry between the two, not a claim
they're now identical. What's left is narrower still: **whether a given
deployment actually runs the isolated binaries** (deployment-time, not
code-level), and closing ClickHouse's staleness gap to match Tantivy's
if that inconsistency matters for a given deployment. See below for
both engines' write-routing, in full.
**ClickHouse (the SQL path) is built.** `enterprise/internal/
tenantprovision` (real `CREATE DATABASE`/`CREATE USER`/`GRANT`) and
@@ -117,25 +123,35 @@ library with no Docker dependency. No "second binary" was needed here,
unlike ClickHouse: Tantivy has no grant system to gate a
commercially-licensed credential behind, so `IndexRegistry` already
lived directly in this AGPL-core `search` binary, and read/write simply
share it.
share it. This write path is now also active-tenant-gated:
`search/src/tenants.rs`'s `ActiveTenantTracker` polls a new
`GET /internal/active-tenants` endpoint on `enterprise-auth` every 60
seconds (RoleService-credentialed, the same auth shape `alerting` uses
against `api`) and `consumer.rs` refuses any tagged record whose tenant
isn't in the polled allowlist — fail-closed on the first fetch (startup
blocks until it succeeds), last-known-good on any later refresh failure.
Off unless `ENTERPRISE_AUTH_URL`/`ENTERPRISE_AUTH_SERVICE_TOKEN` are
both set. Genuinely verified here too: real HTTP round trips (the Bearer
header actually sent, JSON parsing, both fail-closed paths) against a
hand-rolled TCP test server, no live enterprise-auth needed since
`reqwest` doesn't care that the other end is real.
**Both engines share one open question** — deployment topology, covered
above — and Tantivy specifically has one gap ClickHouse's design doesn't:
`chwriter.Registry`'s map is built once at startup from an
active-tenants-only query, so an unrecognized `tenant_id` is refused
outright; `IndexRegistry.resolve()` (used for both read and write) has
no equivalent allowlist at all, because `search` has no Postgres access
to check tenant status against — a syntactically-valid `tenant_id` on a
still-valid-but-should-be-revoked ingest credential can cause an orphan
index directory to be created for a tenant that's no longer active.
Narrow blast radius (isolated, empty except for that traffic, not
cross-tenant leakage, and reachable only with a real signed credential),
but real, and not closed by this change; see
`search/src/registry.rs`'s doc comment on `resolve`. A newly-provisioned
tenant's ClickHouse database and Tantivy index are both now real,
isolated, and actually populated by write-routed agent traffic (the
ClickHouse claim pending live confirmation, the Tantivy claim already
verified).
above — and now differ only in staleness bound, not in whether an
active-tenant gate exists at all: `chwriter.Registry`'s map is built
once at `enterprise-ingest` startup from an active-tenants-only query
and never refreshed, so a tenant deprovisioned after startup keeps
writing successfully until the next restart; `ActiveTenantTracker`
refreshes every 60 seconds, a materially tighter window, with no
corresponding change made to the ClickHouse side to match it. Neither
is a live per-write check (that would mean a database round trip on
every record, a real throughput cost neither implementation accepts),
so both have *some* staleness window by design — the asymmetry between
the two windows is the one thing disclosed as inconsistent, not fixed,
here. A newly-provisioned tenant's ClickHouse database and Tantivy index
are both now real, isolated, and actually populated by write-routed
agent traffic (the ClickHouse claim pending live confirmation, the
Tantivy claim already verified).
## System overview
@@ -494,7 +510,7 @@ terms:
| Tantivy tenant_id resolution (`enterprise/internal/searchclient`) | **Enforced, verified live** — real gRPC wire-level test |
| Ingest tenant *identity* (credential validation, tagging) | **Built and tested** — fail-closed `TenantResolver`, `tenant_id` Kafka header attached per record |
| Ingest tenant *write-routing*, ClickHouse | **Built, not yet confirmed against a real ClickHouse**`enterprise-ingest`/`chwriter.Registry` route each tagged batch to its tenant's own database, fail-closed on an untagged/unprovisioned tenant; Docker-free tests pass, live-database tests are skip-gated. Startup-time active-tenant snapshot, no live recheck — a deprovisioned tenant can keep writing until the next restart |
| Ingest tenant *write-routing*, Tantivy | **Built and genuinely verified**`search/src/consumer.rs` routes each record into its own tenant's index via `IndexRegistry`, same registry the (already-verified) read side uses; no Docker needed, real tests pass. No active-tenant allowlist at all on write (`search` has no Postgres access) — a syntactically-valid `tenant_id` on a still-valid credential can create an orphan index for a no-longer-active tenant; narrow, disclosed, not cross-tenant leakage |
| Ingest tenant *write-routing*, Tantivy | **Built and genuinely verified**`search/src/consumer.rs` routes each record into its own tenant's index via `IndexRegistry`, same registry the (already-verified) read side uses; no Docker needed, real tests pass. Active-tenant-gated too: `tenants::ActiveTenantTracker` polls `enterprise-auth` every 60s (off unless configured), refusing any tenant not in the polled allowlist — tighter staleness bound than ClickHouse's startup-only snapshot, an asymmetry disclosed above, not a gap on Tantivy's side specifically |
| Deployment actually routing traffic to `enterprise-api` (Helm) | **Enforced**`api`/`enterprise-api` are mutually exclusive, same flag as RBAC/audit/SSO |
| Deployment actually routing traffic to `enterprise-api` (docker-compose) | **Enforced**`api`/`enterprise-api` are mutually exclusive via `COMPOSE_PROFILES`, same flag choice as Helm's `enterprise.enabled`; verified via `docker compose config`, not an actual `docker compose up` in this environment |
| Human SSO login — OIDC | **Built, verified with a real fake IdP** (not yet tried against a real external IdP) |