linuser/zabbix-network-topology v5.4.0

v5.4.0 — parallel links

on GitHub

Updating from 5.3.2 — nothing to re-import

No action was added or renamed, so no rescan: replace the module directory,
chown, reload php-fpm. Reload the page once with a cache bypass — the bundle
changed, and this release changes what the map draws. No template changed.
Layouts, manual links, pins, notes and presets stay where they are.

What will look different immediately: a device pair connected by several
cables is now several lines. On a meshed core that is a visible change, and
View → All parallel links: off puts it back to one line per pair, marked
×N. Nothing is hidden either way — the count is on the line.

Expect one burst of change notifications, once. The baseline the topology
diff compares against knows one key per device pair; from this release on
it knows one per cable. On the first refresh after the upgrade the old entry
is carried over to one member of each bundle, and the remaining members are
formally new — a 4×10G LAG therefore announces another cable … — 4 parallel
now three times. Once, per user, then it is quiet. Nothing is wrong, and
nothing was rediscovered: the map simply counts cables now where it used to
count pairs.

Thanks

@christos-diamantis reported
#20 and then
built it: #21 is
his, including the tests that pin the matching rules. His default — fan out
when zoomed in, bundle in the overview — replaced the one the roadmap had
sketched, and it is the better answer.

A reporter who wrote by email sent the screenshots behind four of the
fixes below: the 4.3 Gb/s that is a counter ceiling, red links on a switch
whose ports are merely unused, a port named on one side only, and a map where
30 hosts and 127 unmonitored neighbours piled up on each other.

Added

• Parallel links (LAG, bonding, several cables) are drawn as what they are.
  Until now every pair of devices had exactly one edge: a second LLDP/CDP
  report for the same pair was merged into the first, first-wins. A 2×10G LAG
  showed one member's port and one member's counters, and a failed member was
  invisible. Now every cable is its own edge, with its own ports, traffic,
  utilisation and state.

  • Fanned out (zoomed in): one curve per member, side by side, each in its
    own weathermap colour; port labels at both ends stack into a small list.
    A member that stops being reported while its siblings still are is drawn
    red and dashed among the live ones.
  • Bundled (zoomed out): one line marked ×N (only slightly thicker —
    the label carries the count), total traffic against total capacity of
    the live members, amber glow while a member is down.
  • View → All parallel links: on / off. On (default) fans out from 60 %
    zoom upwards and bundles below. Off always draws one line per device
    pair with ×N — for meshed cores with six or seven cables between
    every pair, where a fan turns into a thicket at any zoom.
  • Clicking a bundle lists its members in the edge panel; a row opens that
    member. Path highlighting marks the whole bundle.
  • The topology diff reports a lost or added member on its own, instead of a
    false "port moved". A link that gains a second cable is not reported as
    removed and re-added.

• The dashboard widget keeps one line per device pair — the tile is too small
  for a fan — but it counts the cables: ×4 on the line, ×4 (1 down)
  when a member is dead, red when all of them are. Without the count the tile
  claimed the same thing for a 4×10G bundle as for a single cable, and a
  failed member was invisible there — which is the very thing this release
  set out to fix.

• "Connected to" — the port a device hangs on, as a column. The hosts table
  carries device and port of the far end, and sorts and filters by them. That
  turns it into a patch list: which access point sits on which switch port,
  printable, and out through the new CSV button. Infrastructure sorts first,
  an ageing link is marked with a clock (for a device that is already gone,
  the port it hung on last is exactly the question), several cables stay one
  line with "+2 more". The table search knows port:Gi1/0/8 as a field.

• CSV export for the hosts table. It existed only in Items mode. Exports
  what is on screen — filtered, sorted, visible rows — with device and port in
  two separate columns, because a spreadsheet wants to group by switch, and
  "Last seen here" as a value one can filter on.

• The port in the connection list of the host panel. One click on the
  device, and the list says eth0 → Gi1/0/8 per connection. It survives the
  device going away, because edges age instead of vanishing.

• A change notification says which cable. A lost bundle member used to be
  announced as "link A ↔ B disappeared", which is false while the other
  members carry it. Now: cable core Gi1/0/1 ↔ acc Te1/1/1 gone — 3 of 4 still
  up, and when the last one dies, the link is down. A single line keeps the
  shorter wording.

• The legend knows the bundle: the ×N line and the amber glow of a
  bundle with a dead cable, drawn with the same symbols the map uses.

How members are matched

Both ends of one cable report it, often with labels that do not compare
("10101" on one side, "GigabitEthernet1/0/1" on the other), and LLDP and CDP
can number the same port differently. A report therefore only opens a new
member when the same device already reported every existing member
over the same protocol on other ports. The number of lines is exact; what
can be wrong with incomparable labels is which far-end port is paired with
which near-end port — never how many cables are drawn.

Changed

• The ghost filter step is called what it does. "Network gear only" became
  "without endpoints", and when nothing can be dropped — because not one
  neighbour reports its capabilities, which is normal for CDP — the map says
  so once per session instead of leaving a step that looks broken.

• The cluster toggle dims itself when the map has fewer than two host
  groups, instead of silently doing nothing.

• The upper limit for LLDP/CDP edges is 6000 instead of 8000. It exists to
  prevent a PHP fatal — a white page mid-render — not clutter, and a member
  edge costs more than the merged one did: measured 5.9 KB against 4.6 KB,
  46 MB against 36 MB at 8000 edges. 6000 restores the headroom the number was
  chosen for. It is still far past any readable map, and when it does bite,
  the map says so.

Fixed

• ifSpeed 4294967295 is not a link speed. It is 2³²−1, the ceiling of a
  32-bit counter: a 10G port reports it and puts the truth in ifHighSpeed.
  The panel said "0.0 % of 4.3 Gb/s", and every utilisation percentage on such
  a link was wrong with it. The value is discarded; without ifHighSpeed the
  speed stays unknown rather than invented.

• A red link meant the host, not the link. The colour came from the ratio
  of down interfaces across both endpoints, so a switch with 17 of 31 ports
  unused but administratively up painted every one of its links red — next to
  a panel showing 0.00 errors for the actual port. The operational state of
  the port the edge hangs on now decides where it is known; the host-wide
  ratio survives only as the fallback for edges without a port mapping.

• A port named on one side only. One end showed 29, the other Port 9 —
  the reporter's raw ifIndex against the text the neighbour announces. A
  resolved neighbour port now takes that interface's own ifName, and a
  host's own label wins over what the far side claims about it.

• Layout "auto" stopped reusing an arrangement that covers a third of the
  map. With 30 hosts and 127 unmonitored neighbours, the saved positions
  covered 30 of 157 drawn nodes and the rest landed in rings around their
  reporting switch. Where ghosts outnumber hosts more than two to one, auto
  computes a fresh layout and says once why.

• A ghost node is no longer treated as a host in the tooltip, the detail
  panel and on its edges: no severity pill, no history request for an id that
  belongs to no host.

• A replugged bundle member is reported as moved, not as one cable gone
  and another arrived — and no ghost line is left next to the cable it is.

• nt:uplink finds its member by port, not by spelling: Gi1/0/2 in the
  tag matches GigabitEthernet1/0/2 on the edge.

• A bundle counts a cable whose port is down, not only one that stopped
  being reported — on the line, in the member list, and in the trunk's colour.
  The two counts used to disagree.

Widgets

Their versions move independently of the module: NT Topology 3.2.0 (the
cable count above), NT Health 2.1.2, NT Table 2.1.2, NT KPI 1.0.2
(the generated data access below). NT Items 1.0.0 is unchanged.

And they are not 7.4-only any more. The note that said so came from a
test on 7.0.28 in July, where they registered but stayed on "Loading…".
Re-tested on 7.0.30 with all five on one dashboard: topology, KPI, health,
table and items render, no console errors. What changed in between is not
established — so if they do hang on a 7.0 of yours, that is worth a report
rather than a shrug.

For contributors

• tools/devnet/ — a compose stack with Zabbix 7.0, Postgres and
  snmpsim: simulated SNMP devices, one .snmprec file per device, all
  values invented. Everything interesting in this module hangs off SNMP
  values, and a bug report can be translated into a file and reproduced in
  minutes. npm run ci:snmprec checks the format, because snmpsim answers a
  file it does not understand with "No Such Instance" and no error at all.

• The shared data access in the widgets is generated, not maintained.
  window.NtWidgetData sat byte-identical in four widget files, because the
  jsLoader knows no ES modules and a widget cannot import the main module.
  ci:parity guarded the copies against each other — but a change was still
  a change in four places, and four identically wrong copies would have
  passed. The source is tools/widget-shared.js now, npm run build writes
  it into the four files, and the gate compares them against it.

• npm run ci:frontend — a gate that reads what the interface says,
  without a browser: detail panel, tooltip, ghost filter, layout decision,
  bundle label, notification text, legend, and the CSV export. Three bug
  reports in September 2026 lived exactly there, and no existing test went red
  for any of them, because none had ever looked at the output.