Skip to content

armgit / Pulse

Public Imported from GitHub
0

Real-time monitoring dashboard for Proxmox VE, PBS, Docker, Kubernetes, TrueNAS and vSphere. Self-hosted, with alerts for failed backups, full disks, unhealthy containers and offline nodes.

Deliver retained PBS and scoped lazy recovery on current main

pulse-triage[bot]3541511d9cbf
diff --cc docs/release-control/v6/internal/subsystems/storage-recovery.md
index c48f132ce3,c967f1fb79..ee4e4c9a05
--- a/docs/release-control/v6/internal/subsystems/storage-recovery.md
+++ b/docs/release-control/v6/internal/subsystems/storage-recovery.md
@@@ -21,59 -21,27 +21,80 @@@
## Purpose
+### Backup-server memory does not invent empty or current readings
+
+The Backup servers table keeps a valid selected memory percentage when a partial
+metric reports capacity without used bytes; it never substitutes zero used bytes.
+A displayed byte pair needs finite non-negative used bytes and positive capacity,
+with used not exceeding capacity. Valid reported pairs still determine the
+percentage, including an explicit measured zero. An independent finite percentage
+in the range 0–100 remains usable when the byte pair is absent or invalid.
+
+Server reachability does not renew memory evidence. The table reuses the resource
+drawer's memory-observation qualification: retained and unknown readings say
+`Last known` or `Freshness unknown`; explicitly unavailable readings show no
+number. A stopped linked agent qualifies its own reading without invalidating an
+independent current source. Offline suppression, datastore risk, backup counts,
+identity, History targets and original observation clocks remain unchanged.
+
+Normal-suite `ProxmoxBackupServersTable.memory.test.tsx` covers partial and invalid
+metrics, measured zero, retained/unavailable/future observations, stopped-agent
+source isolation, live recovery and multi-datastore/offline controls. Browser
+proof must consume the resource producer through the actual Backups surface and
+its existing drawer at desktop and narrow widths. This improves truthful
+inspection, not installed memory relief, a successful backup or guest recovery.
+
+### Unread PVE task history cannot establish absence or completion
+
+The paginated backup-history collector admits only a complete, bounded HTTP200
+response with an explicit task array and unambiguous, scoped task identities and
+times. Missing/null inventories, duplicate or case-conflicting fields, invalid
+rows, contradictory errors, trailing data and unfinished bodies are unread,
+not empty history. Audit authority likewise requires a whole unambiguous
+response granting `Sys.Audit` on the requested node; a valid prefix cannot
+qualify collection completeness. Valid empty arrays, independent accepted pages,
+provider metadata, multi-guest jobs and existing pagination bounds remain valid.
+
+Live tasks on a transiently unread/offline node retain only their last accepted
+connection-scoped observations. Accepted IDs win, other nodes still advance, and
+carried rows neither renew `ObservedAt` nor re-enter durable ingestion. Accepted
+empty reads and removed nodes withdraw their live rows. Explicit access denial
+withdraws that node's live rows and accepted page prefix; a later failed read
+cannot resurrect them. History coverage remains incomplete until its existing
+collection/authority checks pass. Retained running intent continues to keep its
+writer's partial archive from claiming completion. No new endpoint, field, token
+scope, policy, store/schema migration or guest/backup operation is introduced.
+
+Normal-suite `backup_history_response_test.go` covers adverse envelopes, rows,
+body completion, audit evidence and later-page failures; pagination controls
+remain required. `backup_task_inventory_lifecycle_test.go` exercises the actual
+HTTP client and poller through unread/empty/recovery/denial, connection isolation,
+fresh-page precedence and VM/LXC partial-archive ingestion. Desktop/narrow browser
+proof consumes those actual task frames in the production Backups table. These
+are source continuity and attribution proofs, not restore availability, original
+guest recovery or the cause of a reported freeze or stale date.
+
+ ### PBS snapshot group contradictions cannot renew recovery evidence
+
+ Every returned PBS snapshot must belong to its exact requested kind/opaque ID,
+ and each native snapshot time may occur only once within that group response.
+ The PBS client rejects an ambiguous response as a whole, including invalid rows
+ beyond the live retention cap. Existing malformed-read handling retains accepted
+ raw/history evidence without cache or point renewal, marks collection partial
+ rather than complete or access-denied, and never reconciles absence from that
+ read. Independent readable namespaces remain usable. Only ordinary valid reads
+ restore complete evidence; explicit empty arrays still reconcile confirmed
+ absence. No guest, backup, restore, retention policy, access or schema changes.
+
+ The normal-suite client `backup_snapshot_identity_test.go` and collector
+ `pbs_snapshot_identity_test.go` cover exact subjects, opaque host IDs, duplicates,
+ cold/warm VM/CT reads, historical preservation, cache/retry, independent scope,
+ valid recovery and authoritative emptiness. Desktop/narrow browser proof must
+ consume actual collector outputs through Backups inventory/posture clients,
+ inspect retained evidence, uncertainty, recurrence and recovery with keyboard
+ and touch. This proves truthful inspection, not original guest/native recovery.
+
+
### Fleet protection reads bound their request fan-out
Coverage reads keep the existing 200-subject batch size and issue at most four