ServiceRadar v1.4.43

Makes device identity stop inventing devices it only overheard, identifies
hosts from mDNS without probing, moves netprobe's discovery onto the
generic add-on contract, and stops the topology graph asserting links
nothing has confirmed for months.

Whats New

1.4.43

• Identity: a passive netprobe fingerprint no longer resolves onto the
  collector that observed it. The same class of bug had already been
  fixed for the census; this closes the fingerprint path, plus the
  address debris it left behind on hosts a collector had absorbed.
  (PRs #3971, #3979)
• Identity: device metadata is merged in the database instead of being
  read, modified and written back whole, which was silently discarding
  concurrent writers' fields. (PR #3976)
• Identity: every device revival is recorded in an append-only audit
  table. A revival used to clear deleted_reason/deleted_by, leaving
  a resurrected device indistinguishable from one never deleted.
  (PR #3974)
• Identity: a link-local (fe80::) sighting can no longer overwrite a
  routable primary IP, and link-local addresses are no longer accepted
  as cross-device alias evidence. Addresses are now ranked by
  routability in both Elixir and SQL, with a parity test. Existing
  link-local alias rows are archived rather than marked stale, because
  stale alias debris in this system self-revives.
  (PRs #4007, #4009, #4027)
• Identity: interface MACs anchor a device through a per-device table
  rather than a global identifier, so two chassis sharing an interface
  MAC no longer collapse into one. (PRs #4007, #4027)
• Identity: the duplicate-device sweep now also merges a host split
  between a MAC column and a MAC identifier row. (PR #3883)
• Inventory: discovered_interfaces is keyed to current state instead
  of appending a row per poll. A four-device sample held 36,816 rows
  for 374 distinct states -- 98x amplification, growing 767 rows per
  device per day with no retention and no compression. (PR #4027)
• Inventory: vendor is derived from the IEEE OUI registry instead of
  waiting on an integration to fill it in. (PR #3880)
• Inventory: mDNS/DNS-SD announcements identify a host without a probe.
  Capture is in-kernel (no avahi), snapshots ride the same discovery
  path as census, and core classifies from the evidence.
  (PRs #3885, #3887, #3888, #3891, #3896, #3899, #3901)
• Inventory: a mapper's own interface address no longer mints a
  phantom device, an unroutable address mints nothing, IPv6 addresses
  come from ipAddressTable, and a switch no longer loses its role to
  one management address. (PRs #3907, #3929, #3933, #3947)
• Netprobe: census, mDNS, fingerprint, DPI and process observations now
  ride the generic add-on contract as DISCOVERY_V1 payloads named by a
  schema STRING, decoded in core against goldens captured from the Go
  translators before they were removed. Adding an observation type costs
  a registry entry, not an agent release.
  (PRs #3906, #3934, #3943, #3981, #3982, #3990, #3995, #3996)
• Netprobe: banner matching moved to AddonService.RunCommand, the last
  FUNCTIONAL request/response arm on the bespoke IPC socket. Misses are
  no longer padded with a zero-confidence sentinel the caller had to
  strip. The agent falls back to the legacy arm against an older
  netprobe, so a mixed fleet keeps working. (PR #4023)
• Netprobe: the collector's address is stamped from the live interface
  rather than the onboarding pin, so a re-IP'd host stops attributing
  traffic to its old address. (PRs #3991, #4004)
• Netprobe: a restart no longer leaks a TC classifier on the
  full-capture path, and ApplyConfig now reports the settings it used
  to drop silently. (PRs #3918, #3940)
• Topology: projected links age out on the discovery cadence and stale
  links are pruned by default. A deleted discovery job used to leave its
  seeds and UniFi controllers orphaned, freezing hundreds of edges as
  permanently "current". Deleting a job now cascades to its dependants.
  (PRs #4019, #4020)
• Topology: one authoritative compound ELK scene replaces the remaining
  hand-authored geometry -- node placement, endpoint grouping and edge
  routing. Overlapping, penetrating or crossing routes are rejected and
  the last valid scene is retained transactionally, so a bad layout can
  no longer render. Redundant device-level inferred-segment evidence is
  collapsed without disconnecting otherwise isolated devices.
  (PRs #3928, #3999)
• Web UI: Armis, NetBox, Proxmox and Ansible logos on device details;
  mDNS discovery sources render with correct acronym casing; interface
  addresses show in the IP Aliases panel, labelled.
  (PRs #3942, #3983, #4014, #4016)
• Web UI: the device Flows tab keeps its summary and sparkline, puts
  protocol breakdown in the four-card row, and inventory pagination
  survives a refresh. (PRs #3895, #3916, #3923)
• SRQL: device flow queries scope by the device's resolved addresses,
  not device_id, so a host whose IP is only on an alias still shows
  traffic. NetFlow exporters resolve the same way.
  (PRs #3897, #3902)
• Sync: a raw UUID binary no longer poisons the bulk device upsert, and
  an unencodable metadata value is dropped instead of failing the
  whole write. (PRs #3915, #3920)
• CNPG: the image pin is 18.4.0-sr4 at a digest that exists. The
  previous tag pointed at a digest Harbor never published, so a fresh
  cluster could not pull it. (PR #3889)
• CI: the OCI publish workflow runs on ARC runners again -- it was
  bootstrapping Forgejo tools that cannot install as non-root, and
  pointing its cache at a volume only BuildBuddy executors mount.
  (PRs #4015, #4017)
• CI: a test target that executes zero tests now fails instead of
  reporting success. (PR #3972)

Upgrade Notes

• discovered_interfaces is re-keyed to current state (migration
  20260825030000). Historical per-poll rows are discarded. They were
  never a deliberate history -- per-poll provenance made every row
  byte-distinct, so there is no "true row" to keep -- but this IS data
  loss and is called out in the migration itself. timestamp keeps
  meaning "last observed" and is still bumped every poll, so liveness
  queries and the device Interfaces tab are unaffected.
• Link-local identity aliases are archived (migration
  20260825031000). Archived rather than deleted or marked stale:
  stale alias rows in this system self-revive.
• Six migrations ship in this release; migrations.expectedVersion is
  20260825031000:
  • 20260824020000 device revival audit
  • 20260825000000 sr_address_rank function
  • 20260825010000 per-device interface MAC table
  • 20260825020000 cascade mapper job dependants
  • 20260825030000 rekey discovered_interfaces to current state
  • 20260825031000 archive link-local identity aliases
• Netprobe: publish the add-on at 0.2.52. mDNS identification, the
  generic add-on contract, live collector-IP stamping and banner
  RunCommand are not on the fleet until that bundle rolls out. An
  agent older than 1.4.43 still talks the legacy IPC socket, so a
  mixed fleet keeps working during the rollout.
• Topology: stale projected links are pruned by default
  (mapper_topology_prune_stale_projected_links_enabled, default
  true). The first mapper run after upgrade will drop edges whose
  last_observed_at is older than the discovery cadence. Unseen-link
  pruning stays off unless you set
  mapper_topology_prune_unseen_projected_links_enabled.
• Mapper: deleting a discovery job now deletes its seeds and UniFi
  controllers with it (migration 20260825020000). Intended -- those
  rows were what froze topology edges as permanently current.
• CNPG: if you pin cnpg.imageTag / cnpg.imageName yourself, move
  to 18.4.0-sr4 (or the chart's digest). The previous pin did not
  exist in Harbor.
• Known, not fixed: about thirty Helm sub-feature flags remain inert,
  including cnpg.enabled=false and core.migrations.enabled=false,
  which still render their resources (issue #3859). On the device
  flows tab the Services facet stays empty and the traffic chart
  renders no data. A change-only interface history table is not in
  this release (issue #4021).