[webup]

Description

When the telemetry database is unavailable, plain /api/v1/health still returns 200. Add anonymous GET /api/v1/ready, backed by the configured telemetry store's read-only SELECT 1 with a one-second request-derived deadline. Failures return fixed HTTP 503 JSON without database details. OceanBase uses the portable store API rather than the unsupported legacy ClickHouse handle.

Add native healthcheck --ready for scratch images, preserving default local liveness. Keep both root and prefixed probe URLs working with global.external_url paths. Document the endpoint and its limits. Implementation and validation were assisted by Codex.

Issues closed by this PR

Dependency of AI Vision deployment PR #94, addressing the readiness requirement in issue #88. This PR does not close that issue.

Screenshots / Screen Recordings

No visible UI changes. HTTP/native command assertions and real database fault injection cover the behavior.

Additional Information

Earlier isolated validation on 2026-10-25 (readiness commit d2c84788dcb4ce0bf8e9ee781153fb985aeadb88):

• Go 1.25.7: go test ./cmd ./pkg/http/readiness ./pkg/apiserver/signozapiserver -count=1 passed against the actual packages.
• Focused handler, native-command and prefix regressions also passed with race detection and shuffled order. Regression runs reproduced the missing readiness mode and prefixed-root 404 before fixes.
• Full static Linux/arm64 community binary built with CGO_ENABLED=0 GOOS=linux GOARCH=arm64 go build -trimpath -tags timetzdata -ldflags '-s -w' ./cmd/community.
• Real Docker fault/recovery on OrbStack 29.4.0 Linux/aarch64, with pinned oceanbase/seekdb@sha256:500910bda197c0bc5f8bbada10c6ad727807308826818c0ea0391844dedc5cc2 (Linux/amd64 under emulation; SQL reports seekdb v1.4.0.0). Rebuilt binaries ran in cached scratch-derived runtimes, not newly published images.

State	Provider liveness	Provider SQL readiness	Gateway native readiness
Available	HTTP 200 / native exit 0	HTTP 200 / native exit 0	Exit 0
Only database paused	HTTP 200 / native exit 0	HTTP 503 in 1.004s / native exit 1	Exit 1 in 1.052s; gateway liveness stayed exit 0
Database recovered	HTTP 200 / native exit 0	HTTP 200 in 0.003s / native exit 0	Exit 0 in 0.121s

Provider and gateway IDs remained unchanged, restart counts stayed zero, and a retained SQL canary matched after recovery. All task resources were removed after verification; unrelated deployments were preserved.

Tested source was the worktree based on d560a6ad5459220af64c851899aae8f696d5b9fe, now committed as d2c84788dcb4ce0bf8e9ee781153fb985aeadb88. Runtime source hashes were checked against that final worktree. Provider binary SHA-256: 13fa73798b274f80ff9c449b6b352fcc7965f5101ded9f6f3379b6824792faad.

Additional native full-stack qualification — 2026-10-25

The integrated candidate cafe91a3021829b4ccae499bd70f0ff934130971 includes this readiness implementation unchanged plus the shared-SQL-pool shutdown fix in #14. Its full original native Linux/amd64 Dockerfile built successfully; actual provider config ID is sha256:4c5313ac782dee939c6ca5cbfc7ecf2cc22edf43e4366b9137ba14492f808823. AI Vision source was 68c9ebb65b20be6fac378ca60878f533701d6ad0 with gateway image sha256:a45adbf9e5d0a7129a65596f0d14ee3ca4f31638b8d078652e069cb7da8adf73. These source/dev images use explicitly recorded SDK/base bindings; they were not published.

All seven final configurations passed on the existing native Linux/amd64 evaluation VM: two CLI managed Docker cases, four managed/external CE/seekdb Compose cases and one real K3s 1.34.11+k3s1 fixed-Operator CE installation. See AI Vision #94 for the exact inputs, maintained entry points and full evidence.

Native managed DB fault	seekdb	OceanBase CE
Only DB paused	3.4281 s	3.4831 s
Provider HTTP/native liveness	200 / exit 0	200 / exit 0
Provider SQL HTTP readiness	503 in 1.0017 s	503 in 1.0017 s
Provider native readiness	Exit 1 in 1.1268 s	Exit 1 in 1.1363 s
Gateway native readiness	Exit 1 in 1.0524 s; live exit 0	Exit 1 in 1.0520 s; live exit 0
Recovery including Docker health polling	11.5169 s	10.7531 s

Database/provider/gateway IDs and restart counts were unchanged, no rescue timer fired, readiness recovered, and retained SQL/trace data matched. In real Kubernetes, a configured 110-second task-only SQL-network fault produced 19 strict HTTP/native samples and 11 Pod/EndpointSlice snapshots: SQL-dependent readiness failed while local liveness stayed healthy; all affected service endpoints became NotReady and recovered without consumer/observer UID or restart changes. Recast and real same-source FE replacement retained SQL, organization/Secret/PVC identities and three traces/nine exact spans. All 105 recorded Compose product stop checks, including 22 provider stops with #14, exited cleanly. Original source assertions and strict stop gates were retained; failed harness-only setup attempts remain separate. Test runtime resources were removed and original VM resources preserved.

Not qualified: real ClickHouse runtime, full repository suite, published release/image artifacts, all supported architectures/Kubernetes versions, full-component cross-release/code/schema upgrades, sustained/peak load, HA or request draining. No hosted CI success is claimed. These are functional bounded-fault observations, not benchmarks. SELECT 1 establishes connectivity; bootstrap and authenticated ingest/query acceptance are separate checks. Existing liveness behavior is unchanged; consumers still require a matching image before using the new readiness interface.