Rework Ceph collector as a Prometheus complement - #23357

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/collect_pools.go, line 81:

<comment>A pool named `other` can reuse a prior synthetic overflow state (or vice versa) across polls, leaving it with the wrong chart set. Track overflow/full state or replace the existing entity charts when that role changes; otherwise a newly selected real `other` emits dimensions with no charts, while a new overflow may retain full charts.</comment>

<file context>
@@ -3,56 +3,157 @@
+	seen := make(map[string]bool, len(selected))
+	for _, sample := range selected {
+		seen[sample.key] = true
+		if _, ok := c.seenPools[sample.key]; !ok {
+			c.seenPools[sample.key] = &entityState{}
+			c.addPoolCharts(sample.key, !sample.overflow)
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/charts.go, line 54:

<comment>A Dashboard outage on a fresh collector startup produces no `ceph.dashboard_api_status` chart, so the default API-unavailable alert cannot fire until one successful identity probe has already occurred. Consider registering an unlabeled/status chart before the identity probe, or otherwise creating it on the failure path without consuming the later FSID-labeled cluster-chart initialization.</comment>

<file context>
@@ -48,6 +51,7 @@ const (
 )
 
 var clusterCharts = collectorapi.Charts{
+	dashboardAPIStatusChart.Copy(),
 	clusterStatusChart.Copy(),
 	clusterHostsCountChart.Copy(),
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/collect_osd.go, line 116:

<comment>A Dashboard response that omits `X-Total-Count` and ignores pagination can make an OSD collection run forever, preventing subsequent collection cycles. Consider detecting repeated page identities (or otherwise bounding headerless pagination) and return an error before continuing.</comment>

<file context>
@@ -3,64 +3,285 @@
+func (c *Collector) fetchAllOSDs(ctx context.Context) ([]apiOsdResponse, error) {
+	var all []apiOsdResponse
+	total := -1
+	for offset := 0; ; offset += osdPageSize {
+		query := url.Values{
+			"offset": {strconv.Itoa(offset)},
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/cephfunc/router.go, line 28:

<comment>Disabled RGW Functions remain published in the Function UI and fail only after invocation with 404. Bind each method's availability to the job configuration (or avoid publishing disabled methods) so opt-in methods are not advertised before enabled.</comment>

<file context>
@@ -0,0 +1,428 @@
+	return &router{deps: deps, config: config}
+}
+
+func Methods() []funcapi.FunctionConfig {
+	return []funcapi.FunctionConfig{
+		{ID: MethodHealth, Name: "Ceph Health", UpdateEvery: 10, Help: "Detailed Ceph health checks and RCA messages."},
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/func_deps.go, line 163:

<comment>OSD Function results can represent a mixed inventory snapshot when OSD membership changes during pagination: a later `X-Total-Count` overwrites the earlier value and rows may have been skipped or shifted. Preserve the first advertised total and reject/retry when a later page differs, matching `fetchAllOSDs`.</comment>

<file context>
@@ -0,0 +1,962 @@
+			if err != nil || parsed < 0 {
+				return cephfunc.OSDResult{}, errors.New("OSD inventory returned an invalid X-Total-Count")
+			}
+			total = parsed
+		}
+		pageLength := len(page)
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/metadata.yaml, line 844:

<comment>The Function UI columns document OSD and RGW quota `utilization` with `unit: percent`, but the runtime `cephfunc` schema registers these columns with unit `%`. Keep the metadata and the emitted Function schema consistent so the documented column type/unit matches what operators actually see in the Function table.</comment>

<file context>
@@ -211,218 +595,394 @@ modules:
+              - {name: total_bytes, type: integer, unit: bytes, visibility: hidden, description: Total OSD capacity.}
+              - {name: used_bytes, type: integer, unit: bytes, description: Used OSD capacity.}
+              - {name: available_bytes, type: integer, unit: bytes, description: Available OSD capacity.}
+              - {name: utilization, type: float, unit: percent, description: OSD utilization.}
+              - {name: read_bytes_per_sec, type: float, unit: bytes/s, visibility: hidden, description: Current read rate.}
+              - {name: write_bytes_per_sec, type: float, unit: bytes/s, visibility: hidden, description: Current write rate.}
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/testdata/v20.2.2/dashboard_contract.json, line 244:

<comment>The RGW account fixture's fields don't line up with what the collector extracts for an account quota row: the fixture provides `id`/`name`, but lookupRGWQuota reads `account_id`/`account` for the Account attribution and `stats` for usage. As a result the contract test's account target never exercises those fields (Account comes out empty, no usage), so the fixture doesn't validate the account contract it's meant to guard. Consider aligning the fixture's key names/usage block with the fields the parser reads, or the mismatch stays invisible to the test.</comment>

<file context>
@@ -0,0 +1,348 @@
+        "max_objects": 100,
+        "max_size": 1000
+      },
+      "id": "RGW00000000000000001",
+      "name": "account-a",
+      "tenant": "tenant"
</file context>

Check if this issue is valid — if so, understand the root cause and fix it. At src/go/plugin/go.d/collector/ceph/client.go, line 1:

<comment>The set of headers that the Ceph client forbids (authorization, cookie, host) is now enforced in two separate places: newCephClient in client.go and validateConfig in init.go. Keeping the same rule duplicated in two layers makes it easy for the two copies to drift (e.g., adding a new managed header in one place but not the other), and one of them is functionally redundant since newCephClient is always constructed after validation. Consider keeping the single source of truth in validateConfig (or in the client) and having the other path rely on it.</comment>

<file context>
@@ -0,0 +1,528 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+package ceph
</file context>