SWIP-14 Inspect API for Admin Server

Motivation

When operators author OAL / MAL / LAL rules and turn them on (statically or via runtime-rule hot-update from SWIP-13), the natural follow-up question is:

• "Which metrics did this rule actually produce?"
• "For metric service_xyz, which entities are emitting values right now?"
• "Once I have an entity, can I query it via MQE without guessing how to reconstruct the Entity shape?"

Today the only way to answer these is to dig through dashboards, write exploratory MQE expressions against guessed entity names, or read storage directly. There is no first-class "browse" interface.

The public GraphQL surface deliberately does not expose "enumerate entities with values for metric X in range R", because that scan is unbounded and would compete with user-facing query traffic for storage resources. Operators still need it — for debugging rules, validating new monitoring features, and onboarding new agents.

This SWIP proposes an Inspect API on the existing admin-server (SWIP-13's admin-host bundle, default port 17128) that exposes:

1. The metric catalog with type + scope + supported downsamplings.
2. For a given metric + time range, the set of entities (capped) that hold values, decoded into MQE-ready form.

The output of (2) is shaped so the operator can paste it directly into a follow-up MQE call against the existing MetricsExpressionQuery GraphQL mutation. This SWIP does not introduce a new MQE entry point — the public GraphQL path remains the single MQE surface.

Scope

In scope:

• COMMON_VALUE and LABELED_VALUE metric types — the two dataTypes that MQEVisitor.visitMetric actually dispatches today.
• Service, ServiceInstance, Endpoint, ServiceRelation, ServiceInstanceRelation, EndpointRelation scopes.
• All three storage backends: BanyanDB, Elasticsearch, JDBC.
• /inspect/metrics lists every registered metric regardless of type (including HEATMAP and SAMPLED_RECORD) so operators can see the full catalog. Only /inspect/entities is restricted to MQE-dispatchable types.
• Companion: relocate + dual-bind the Status API. The existing status-query-plugin (cluster nodes / alarm runtime status / TTL config / debug query trace) is replaced by a new status feature module under server-admin/. The module registers its handlers on both the public REST HTTPHandlerRegister (preserving today's binding — skywalking-ui consumes these endpoints on core.restPort) and, if admin-server is enabled, the admin-server HTTPHandlerRegister. The default state remains "status on the public REST port" so the UI keeps working unchanged. No new endpoint is added — clients that need to discover the OAP REST address parse the existing GET /debugging/config/dump, which already publishes the effective configuration including REST host and port.

Out of scope (deferred):

• HISTOGRAM / HEATMAP metrics from /inspect/entities. MQE has no dispatch path for histogram dataType — MQEVisitor.visitMetric only handles COMMON_VALUE, LABELED_VALUE, and SAMPLED_RECORD; histograms flow through the legacy MetricsQuery.readHeatMap GraphQL resolver, which is being phased out. The mqeEntity block this API produces would not be usable for them. They remain visible in /inspect/metrics (with type: HEATMAP) so operators can see they exist.
• SAMPLED_RECORD metrics from /inspect/entities. These live in Stream storage (logs, trace segments, browser errors) and are addressed via IRecordsQueryDAO / trace-relative queries, not the metrics DAO. A future Inspect-Records endpoint can mirror this design once the scoping question is settled. They remain visible in /inspect/metrics.
• Process and ProcessRelation scopes. ProcessID is a SHA-256 hash with no decoder; readable rendering would require a ProcessTraffic join, which we punt on for v1.
• Firing MQE from admin-server. Operators take the inspect response and call the existing public GraphQL execExpression mutation. Keeping MQE on one surface preserves auth/quotas/observability there.

Architecture

   ┌── operator / swctl / curl ──┐         ┌── skywalking-ui ──┐
   │  /inspect/* (admin-only)    │         │  /status/* (UI)   │
   │  /status/*  (admin mirror)  │         │                   │
   └─────────────┬───────────────┘         └─────────┬─────────┘
                 │ admin port 17128                  │ public REST port 12800
                 ▼                to kee                   ▼
   ┌─────────────────────────────────┐   ┌──────────────────────────────┐
   │      admin-server (Armeria)     │   │  public REST (sharing/core)  │
   │  ┌───────────────────────────┐  │   │   GraphQL / MQE / receivers  │
   │  │  inspect feature module   │  │   │                              │
   │  │   ├── /inspect/metrics    │  │   │   ┌────────────────────────┐ │
   │  │   └── /inspect/entities   │  │   │   │   status feature module │ │
   │  └───────────────────────────┘  │   │   │   ├── /status/cluster/.. │ │
   │                                 │   │   │   ├── /status/alarm/..  │ │
   │  ┌───────────────────────────┐  │   │   │   ├── /status/config/.. │ │
   │  │  status feature module    │◄─┼───┼───┤   └── /debugging/...    │ │
   │  │  (same handler instance,  │  │   │   │                          │ │
   │  │   dual-bound when admin   │  │   │   └────────────────────────┘ │
   │  │   is on; this side is     │  │   │                              │
   │  │   silent if admin off)    │  │   └──────────────────────────────┘
   │  └───────────────────────────┘  │
   └─────────────────────────────────┘
                     │                             │
                     ▼                             ▼
       ┌──────────────────────────┐  ┌──────────────────────────────┐
       │ ValueColumnMetadata      │  │  IMetricsQueryDAO            │
       │ StorageModels            │  │   .listEntityIdsInRange(...) │  ◄── new method
       │ MetadataQueryService     │  │  (BanyanDB / ES / JDBC impl) │
       │ DefaultScopeDefine       │  └──────────────────────────────┘
       └──────────────────────────┘
                     │
                     ▼
              decode entity_id via IDManager.*ID.analysisId(),
              enrich with layer(s) from MetadataQueryService cache,
              shape into MQE-ready Entity

  inspect-client flow (admin operator):
    1. (one-time) GET admin:17128/debugging/config/dump → parse REST host/port
       OR rely on a known deployment-time URL.
    2. GET admin:17128/inspect/metrics            → pick a metric
    3. GET admin:17128/inspect/entities?metric=…  → get entityId + mqeEntity
    4. POST <REST URL>/graphql execExpression(    → fire the actual query on
         entity = mqeEntity, expression = …)        the public GraphQL surface

  ui flow (skywalking-ui, unchanged):
    1. GET oap:12800/status/cluster, /status/alarm, /status/ttl, …
       (admin-server may be off — these still work via the public REST binding)

The inspect feature module is a sibling of dsl-debugging and runtime-rule under the existing admin-host. It registers its handlers in start() against the shared HTTPHandlerRegister. No new port, no new auth surface.

API

GET /inspect/metrics

Lists all registered metrics with type, scope catalog, and the downsampling levels at which each metric is materialized.

Query parameters:

Response (200 OK, application/json):

{
  "metrics": [
    {
      "name": "service_cpm",
      "type": "REGULAR_VALUE",
      "catalog": "SERVICE",
      "scopeId": 1,
      "scope": "Service",
      "valueColumnName": "value",
      "downsamplings": ["MINUTE", "HOUR", "DAY"]
    },
    {
      "name": "service_percentile",
      "type": "LABELED_VALUE",
      "catalog": "SERVICE",
      "scopeId": 1,
      "scope": "Service",
      "valueColumnName": "value",
      "downsamplings": ["MINUTE", "HOUR", "DAY"]
    },
    {
      "name": "endpoint_response_time",
      "type": "HEATMAP",
      "catalog": "ENDPOINT",
      "scopeId": 3,
      "scope": "Endpoint",
      "valueColumnName": "dataset",
      "downsamplings": ["MINUTE", "HOUR", "DAY"]
    }
  ]
}

Implementation:

• Source of metric list is ValueColumnMetadata.INSTANCE.getAllMetadata(), the same map already exposed via MetricsMetadataQueryService.listMetrics.
• Entries with Column.ValueDataType.NOT_VALUE are filtered out — these are persisted-but-not-queryable internal columns (topology lines, service entries) per Column.java's contract. They never appear in /inspect/metrics.
• Entries whose scope resolves to Scope.All are filtered out as well (Scope.All is deprecated and not routable through the inspect path).
• type is mapped from Column.ValueDataType exactly as MetricsMetadataQueryService.typeOfMetrics does today.
• scope is Scope.Finder.valueOf(scopeId).name().
• downsamplings is derived once at startup by scanning StorageModels.allModels(), grouping by metric name, collecting model.getDownsampling() values into a Set. Pure metadata, no storage I/O.
• This endpoint is safe to poll; results change only when models are registered or removed (boot, runtime-rule add/delete).

GET /inspect/entities

Lists entities that hold values for a metric within a time range, decoded into MQE-ready form.

Query parameters:

coldStage is intentionally not exposed. The inspect API targets the operationally-active hot path; cold-stage scans bypass the same TTL window the operator would be debugging within.

The limit is applied as LIMIT N at the storage layer — it bounds the total rows scanned (N = 300 ≅ 10 buckets × 30 entities), not 300 entities. Backends dedup on entity_id before returning.

step is parsed via DownSampling.valueOf(step) after upper-casing and must appear in the metric's supportedDownsamplings (else 400). The (start, end) pair is converted to time buckets through the same Duration.getStartTimeBucket() / getEndTimeBucket() path MQE uses, so range semantics match exactly.

Response (200 OK, application/json) — example: service_cpm over the last 10 minutes, where service payment is registered under both MESH and GENERAL layers:

{
  "metric": "service_cpm",
  "scope": "Service",
  "step": "MINUTE",
  "start": "2026-05-10 1230",
  "end":   "2026-05-10 1240",
  "rows": [
    {
      "entityId": "cGF5bWVudA==.1",
      "decoded": { "serviceName": "payment", "isReal": true },
      "layer": "MESH",
      "mqeEntity": {
        "scope": "Service",
        "serviceName": "payment",
        "normal": true
      }
    },
    {
      "entityId": "cGF5bWVudA==.1",
      "decoded": { "serviceName": "payment", "isReal": true },
      "layer": "GENERAL",
      "mqeEntity": {
        "scope": "Service",
        "serviceName": "payment",
        "normal": true
      }
    },
    {
      "entityId": "Y2hlY2tvdXQ=.1",
      "decoded": { "serviceName": "checkout", "isReal": true },
      "layer": "GENERAL",
      "mqeEntity": {
        "scope": "Service",
        "serviceName": "checkout",
        "normal": true
      }
    },
    {
      "entityId": "bXlzcWw=.0",
      "decoded": { "serviceName": "mysql", "isReal": false },
      "layer": "VIRTUAL_DATABASE",
      "mqeEntity": {
        "scope": "Service",
        "serviceName": "mysql",
        "normal": false
      }
    }
  ]
}

An empty result is not an error — { "metric": …, "rows": [] } simply means no entity wrote a value in the requested window.

Multi-layer rule: one logical entity emits one row per layer it is registered under. payment appears twice above because ServiceTraffic has two rows (payment/MESH and payment/GENERAL), and the inspect API does not collapse them — the operator may want to investigate one layer at a time, and Layer is meaningful for downstream calls such as listInstances(serviceId). The mqeEntity block, however, is identical across layer rows because MQE itself does not take a layer parameter — including it lets the operator copy-paste without thinking.

ServiceInstance example (service_instance_cpm):

{
  "rows": [
    {
      "entityId": "cGF5bWVudA==.1_cG9kLTAx",
      "decoded": {
        "serviceName": "payment",
        "isReal": true,
        "serviceInstanceName": "pod-01"
      },
      "layer": "GENERAL",
      "mqeEntity": {
        "scope": "ServiceInstance",
        "serviceName": "payment",
        "serviceInstanceName": "pod-01",
        "normal": true
      }
    }
  ]
}

ServiceRelation example (service_relation_client_cpm):

{
  "rows": [
    {
      "entityId": "Y2hlY2tvdXQ=.1-cGF5bWVudA==.1",
      "decoded": {
        "source":      { "serviceName": "checkout", "isReal": true },
        "destination": { "serviceName": "payment",  "isReal": true }
      },
      "layer": "GENERAL",
      "mqeEntity": {
        "scope": "ServiceRelation",
        "serviceName": "checkout",
        "normal": true,
        "destServiceName": "payment",
        "destNormal": true
      }
    }
  ]
}

ServiceInstanceRelation example (service_instance_relation_client_cpm, e.g., a consumer instance calling a provider instance in the java-agent provider/consumer showcase):

{
  "rows": [
    {
      "entityId": "Y29uc3VtZXI=.1_cG9kLWE=-cHJvdmlkZXI=.1_cG9kLWI=",
      "decoded": {
        "source": {
          "serviceName": "consumer",
          "isReal": true,
          "serviceInstanceName": "pod-a"
        },
        "destination": {
          "serviceName": "provider",
          "isReal": true,
          "serviceInstanceName": "pod-b"
        }
      },
      "layer": "GENERAL",
      "mqeEntity": {
        "scope": "ServiceInstanceRelation",
        "serviceName": "consumer",
        "normal": true,
        "serviceInstanceName": "pod-a",
        "destServiceName": "provider",
        "destNormal": true,
        "destServiceInstanceName": "pod-b"
      }
    }
  ]
}

EndpointRelation example (endpoint_relation_cpm, where consumer's /order calls provider's /charge):

{
  "rows": [
    {
      "entityId": "Y29uc3VtZXI=.1-L29yZGVy-cHJvdmlkZXI=.1-L2NoYXJnZQ==",
      "decoded": {
        "source": {
          "serviceName": "consumer",
          "isReal": true,
          "endpointName": "/order"
        },
        "destination": {
          "serviceName": "provider",
          "isReal": true,
          "endpointName": "/charge"
        }
      },
      "layer": "GENERAL",
      "mqeEntity": {
        "scope": "EndpointRelation",
        "serviceName": "consumer",
        "normal": true,
        "endpointName": "/order",
        "destServiceName": "provider",
        "destNormal": true,
        "destEndpointName": "/charge"
      }
    }
  ]
}

Error responses:

Status API relocation

The status-query-plugin module currently registers its HTTP handlers (AlarmStatusQueryHandler, ClusterStatusQueryHandler, TTLConfigQueryHandler, DebuggingHTTPHandler for query trace) against the public HTTPHandlerRegister exposed by the core / sharing-server module — the same HTTP surface as GraphQL, on the public REST port (core.restPort, default 12800, or the sharing-server port if configured). skywalking-ui calls these endpoints on the public REST port today, so the binding cannot move; it can only be supplemented.

This SWIP replaces status-query-plugin with a new feature module status under server-admin/ that dual-binds:

1. Always: registers every handler on the public REST HTTPHandlerRegister resolved through CoreModule — the same surface status-query-plugin uses today, with the same URIs and payloads. UI compatibility is preserved.
2. Conditionally: if AdminServerModule is loaded (SW_ADMIN_SERVER=default), also registers every handler on the admin-server HTTPHandlerRegister. Operators that gateway-protect the admin port can drive status from the same private surface as /inspect/*, /dsl-debugging/*, /runtime/rule/*.

The module is gated by SW_STATUS=default and is enabled by default (matching today's status-query-plugin posture so existing UI deployments keep working without an opt-in step). It does not require admin-server: when admin-server is off, status still binds on the public REST port — that is the entire point of the dual-bind design.

Routes hosted by the status feature module — every existing route preserved verbatim from status-query-plugin. No new endpoint is added by this SWIP; clients that need to discover the OAP REST URL parse the existing GET /debugging/config/dump response, which already publishes the effective configuration including REST host and port.

Every route is published on both binding surfaces (public REST + admin when admin-server is enabled). URI paths, query parameters, and response payloads do not change — only the host:port the routes bind on.

Storage layer

A new abstract method on IMetricsQueryDAO:

List<String> listEntityIdsInRange(String metricName,
                                  String valueColumnName,
                                  Duration duration,
                                  int limit) throws IOException;

Abstract on purpose so any 3rd party storage backend that implements IMetricsQueryDAO MUST provide this override — a default body would let a missing override slip through compilation and surface as silent "no entities" or 500 the first time the inspect API hit that backend.

Returns distinct entity_ids, ordered by most recent timestamp descending, capped at limit. Each backend must enforce that ordering explicitly — without an explicit sort directive, scoring / index-internal order can drop a hot entity that ingested late before the cap is reached. Implementations:

• BanyanDB (BanyanDBMetricsQueryDAO): MeasureQuery against the resolved measure (downsampling chosen via MetadataRegistry.formatName(metricName, downSampling)), timestamp range filter, tag projection limited to ENTITY_ID, value-column field projection, setOrderBy(new AbstractQuery.OrderBy(Sort.DESC)) for timestamp-DESC ordering, .limit(limit). Client-side LinkedHashSet dedup preserves the server-side ordering.
• Elasticsearch (MetricsQueryEsDAO): range filter on time_bucket, search.sort(Metrics.TIME_BUCKET, Sort.Order.DESC) for timestamp-DESC ordering, Search.size(limit). The downsampling-suffixed index name is resolved through the existing IndexController / MetricsModelInstaller path (the same name the existing readMetricsValues uses for the requested step). Client-side LinkedHashSet dedup on the returned hits.
• JDBC (JDBCMetricsQueryDAO): getTablesForRead(metricName, startTimeBucket, endTimeBucket) resolves the day-partitioned tables. Per table: SELECT entity_id, MAX(time_bucket) AS latest_time_bucket FROM <table> WHERE table_column = ? AND time_bucket BETWEEN ? AND ? GROUP BY entity_id ORDER BY latest_time_bucket DESC LIMIT ?. (PostgreSQL rejects ORDER BY time_bucket when it isn't in a SELECT DISTINCT list, so the GROUP BY shape is the portable one across H2 / MySQL / PostgreSQL.) Per-table results are merged into a Map<entityId, maxTimeBucket> with Math::max, then sorted by the value DESC and capped at the global limit — so a multi-day range cannot fill the result with stale entities from older partitions before newer partitions are queried.

This shape mirrors the existing readLabeledMetricsValuesWithoutEntity(metricName, valueColumnName, labels, duration) (limit hard-coded to 10, projects value as well) — we reuse the range-scan and downsampling-resolution scaffolding from those implementations.

Decode & enrich

Per row, the handler:

1. Calls Scope.Finder.valueOf(scopeId) to resolve the scope from the metric's metadata.
2. Switches on the scope to call the correct IDManager.{ServiceID,ServiceInstanceID,EndpointID}.analysisId(...) / analysisRelationId(...).
3. For Service-bearing rows, reads MetadataQueryService.getService(serviceId) from the in-memory cache and emits one row per layer in service.getLayers().
4. Builds the mqeEntity block by inverting the decode — same fields Entity.buildId() would consume to reconstruct the same entity_id.

If a service is missing from the metadata cache (e.g., it was just deleted but a metric row still exists in the storage TTL window), the row is still emitted with layer: null and the decoded names; the operator can still call MQE.

Module layout

Two new feature modules under the existing admin-server bundle:

oap-server/server-admin/inspect/
├── pom.xml
├── src/main/java/org/apache/skywalking/oap/server/admin/inspect/
│   ├── InspectModule.java               # ModuleDefine
│   ├── InspectModuleProvider.java       # ModuleProvider
│   ├── InspectModuleConfig.java
│   ├── handler/InspectRestHandler.java  # Armeria handler
│   ├── decoder/EntityDecoder.java       # scope → IDManager dispatch
│   └── response/{MetricRow, EntityRow, MqeEntity}.java
└── src/main/resources/META-INF/services/
    └── org.apache.skywalking.oap.server.library.module.ModuleDefine

oap-server/server-admin/status/        ◄── relocated from
                                           server-query-plugin/status-query-plugin
├── pom.xml
├── src/main/java/org/apache/skywalking/oap/server/admin/status/
│   ├── StatusModule.java
│   ├── StatusModuleProvider.java
│   ├── StatusModuleConfig.java
│   ├── handler/
│   │   ├── ClusterStatusHandler.java    # moved
│   │   ├── AlarmStatusHandler.java      # moved
│   │   ├── TTLConfigHandler.java        # moved
│   │   └── DebuggingHTTPHandler.java    # moved (query trace)
│   └── service/{AlarmStatusQueryService, …}.java   # moved
└── src/main/resources/META-INF/services/
    └── org.apache.skywalking.oap.server.library.module.ModuleDefine

requiredModules():

• inspect: CoreModule.NAME, StorageModule.NAME, AdminServerModule.NAME. Inspect is admin-only; it has no public-REST mirror and therefore does require admin-server.
• status: CoreModule.NAME only. AdminServerModule.NAME is consulted reflectively via moduleManager.has(...) for the conditional second handler registration. Declaring it as required would force every status-using deployment to also enable admin-server, which contradicts the "UI keeps working with admin off" goal.

Handler registration in status's start():

// Always — public REST surface, where the UI calls today.
HTTPHandlerRegister publicRegister = manager.find(CoreModule.NAME)
    .provider().getService(HTTPHandlerRegister.class);
registerAll(publicRegister);

// Additionally, when admin-server is enabled.
if (manager.has(AdminServerModule.NAME)) {
    HTTPHandlerRegister adminRegister = manager.find(AdminServerModule.NAME)
        .provider().getService(HTTPHandlerRegister.class);
    registerAll(adminRegister);
}

Inspect's start() registers only on the admin-server register (mirroring dsl-debugging and runtime-rule). The old status-query-plugin module is removed.

Maven module / pom changes

One module dropped, two added, three existing poms edited. Net: same number of distribution jars, just relocated under server-admin.

Cost & safety

• /inspect/metrics is in-process metadata only — O(N) over registered metrics, no I/O.
• /inspect/entities issues exactly one storage call per request bounded by LIMIT 300. The 300 ceiling is enforced server-side and cannot be raised via parameter. This is the same shape as today's readLabeledMetricsValuesWithoutEntity (limit 10) just with a higher cap; no new query class is introduced.
• The endpoint is bound only on the admin port (default 17128), which is not the agent or query port; existing operator-network controls apply.
• No caching layer in v1. If poll-storms become a problem we can add a short-TTL in-process cache keyed on (metric,start,end,step,limit), but given operators are humans and the cap is 300 rows, premature.

Compatibility

The dual-bind design keeps the Status API backwards compatible for the UI and for any scripted consumer of http://oap:12800/status/....

The admin-port binding for status is intentionally additional, not a replacement, so that admin-server may stay off (the default) without breaking the UI. A future SWIP can drop the public-REST binding once skywalking-ui has migrated to the admin-port path; until then, both surfaces co-exist.

Test plan

The inspect surface is e2e-validated against the existing java-agent e2e (test/e2e-v2/cases/simple/...) — that suite already runs across every supported storage backend (BanyanDB, Elasticsearch, JDBC), so piggy-backing keeps the matrix free.

The agent showcase boots a provider and a consumer Java service; this gives us live metrics across every scope inspect supports. The flow uses curl + jq directly (no swctl wiring) so the test is portable across the existing e2e runners:

1. Wait for the showcase to ingest at least one minute window.
2. (one-time) curl http://oap:17128/debugging/config/dump and parse the response to learn the OAP REST host:port the harness will fire MQE against. Existing endpoint; no new discovery primitive needed.
3. curl http://oap:17128/inspect/metrics?regex=service_cpm → assert service_cpm appears with scope=Service, type=REGULAR_VALUE, downsamplings includes MINUTE. Repeat for instance_jvm_memory_* (ServiceInstance scope) and instance_jvm_class_loaded_count (ServiceInstance, REGULAR_VALUE).
4. curl http://oap:17128/inspect/entities?metric=service_cpm&start=…&end=…&step=MINUTE → expect rows for both provider and consumer. jq out one row's mqeEntity block.
5. POST MQE expression service_cpm against <discovered MQE URL> with the extracted mqeEntity → assert non-empty time-series result.
6. Repeat (4)–(5) with instance_jvm_memory_heap_used (ServiceInstance scope) — confirm both provider and consumer instances surface, and MQE for the picked instance returns data.
7. Relation coverage: /inspect/entities?metric=service_relation_client_cpm → assert a consumer→provider row, then MQE service_relation_client_cpm against its mqeEntity returns data. Same pattern for service_instance_relation_client_cpm and endpoint_relation_cpm (when endpoint relations are present in the showcase).
8. Negative cases: HEATMAP metric (endpoint_response_time) → expect 400 from /inspect/entities; step=DAY against a metric registered only at MINUTE → expect 400 with the supported-downsamplings list.

Module-level unit tests live alongside each new source: EntityDecoder for every supported scope (covering isReal=true and isReal=false service IDs, multi-layer service merge, missing-from-cache fallback), the /inspect/metrics filter (NOT_VALUE / Scope.All exclusion), and the new DAO method per backend mocked at the storage interface.

Deliverables

Implementation order

The status-relocation work and the inspect-API work share zero code paths and can be staged independently:

1. Phase 1 — status relocation. Create server-admin/status module, move sources, set up dual-bind, delete status-query-plugin. UI and existing operator scripts continue to work; URI paths and payloads are unchanged.
2. Phase 2 — inspect. Add IMetricsQueryDAO.listEntityIdsInRange across the three backends, then server-admin/inspect with handlers and decoder. Independent of Phase 1 — inspect clients that need to discover the OAP REST URL parse /debugging/config/dump, an already-existing endpoint preserved by Phase 1.

Each phase is a self-contained PR.

Future work

• Inspect-Records (/inspect/records) for SAMPLED_RECORD metrics, backed by IRecordsQueryDAO.
• Inspect-Process for Process / ProcessRelation, contingent on a decoder for ProcessID (likely a ProcessTraffic join).
• Optional ?attribute=k=v filter pushdown to leverage the same attribute conditions TopNCondition accepts.