Skip to content

Rack the shelf switch in U4 — buy it with the UPS pack #110

Closed

Description

@Gerrrt

A 1U vented shelf (~$25–45) in U4, the one free slot, to carry the unmanaged
TP-Link that prometheus and oracle hang off.

Why it is a monitoring issue, not a tidiness one

Both shelf machines are laptops with their own batteries. On a mains cut they
stay running and go deaf — the switch between them and the network has no
battery at all. A monitoring host that survives the event it exists to observe
but cannot report it is a strange way to lose an evidence trail.

That is the same class of fault as the ones found while verifying #12: a
component that is up, reports itself healthy, and produces nothing. This one just
has a physical cause.

Buy it with the UPS pack, not after

Coupled to #93. The APCRBC115 cartridge on its own leaves the monitoring path
half-protected: the hosts ride out the cut and the switch does not. Ordering them
together is the whole point — two purchases, one outcome.

The spec is settled

Measured at the rack on 2026-08-21:

  • 4-post frame, square mounting holes. Depth is not a constraint.
  • So: a full 1U shelf with rear support, not a cantilever.
  • Cage nuts plus screws — most shelves include them, not all. Check the
    listing rather than discovering it on the day.

U4 is empty only because nothing had been chosen for it, so there is no service
clearance to preserve.

The purchase cannot be avoided — this was checked

morpheus sits at U5 in a purpose-built mount that holds one ProDesk Mini —
a single-device bracket, not a shelf with spare room. The switch therefore needs
its own U, and the shelf is the way to give it one.

The 1U shelf also has room beside the switch for a second small box later, which
the U5 bracket by design does not.

Rack the cold spare here too, powered off

The spare ProDesk from #92 goes on this shelf beside the switch — racked, cabled
for its cold-spare role, and left powered off. A spare that is plugged in and
on the network is exposed to whatever took the primary.


Was the last item in docs/roadmap.md
still carrying its own checkbox; filed so the roadmap holds no state of its own.

Activity

  1. reopened this on Sep 8, 2026
  2. Gerrrt commented on Sep 8, 2026

    @Gerrrt
    OwnerAuthor

    Reopened. This was closed by accident, not by a rack visit.

    How it closed: commit 7147f0e (2026-08-27, docs(runbooks): fit the UPS
    battery and rack the shelf switch
    ) wrote the procedure before the boxes were
    open. Its message says "the middle step is the one that closes #110" — and
    GitHub read closes #110 as a close keyword and closed this issue when the
    commit reached main. The same commit's own last paragraph says "Nothing here
    claims a fitted battery."
    It did not claim a racked shelf either.

    What every document says since, and they agree:

    • docs/hardware.md — the shelf is on hand … not in the rack table above
      because it is not yet in the rack
      .
    • docs/runbooks/fit-the-ups-battery.md, status banner of 2026-09-03 — the
      shelf and the switch move (step 2, items 1–4) were not done, so Rack the shelf switch in U4 — buy it with the UPS pack #110 is
      untouched and prometheus and oracle still go deaf on a mains cut.
    • docs/roadmap.md — under Infrastructure, open: the shelf is on hand; what
      is left is the rack visit.
    • README.md — the switch between the monitoring host and the network still
      has no battery at all.

    Confirmed on 2026-09-08: still not racked. The roadmap entry is correct and
    stays where it is; the thing that was wrong was this issue's state.

    What closes it is unchanged from the runbook's step 2: shelf into U4 with
    the rear support engaged, the TP-Link onto it, and — the step that actually
    closes the gap — its power onto a UPS-fed outlet. The spare ProDesk from #92
    (ordered 2026-09-08) racks beside it, powered off, so the two visits can be
    one. Then hardware.md gains the U4 row, network.md's TP-Link note changes,
    and the roadmap entry moves to Done.

    🤖 Generated with Claude Code

  3. added a commit that references this issue on Sep 8, 2026
  4. reopened this on Sep 8, 2026
  5. Gerrrt commented on Sep 8, 2026

    @Gerrrt
    OwnerAuthor

    Reopened a second time. Closed at 15:40 UTC by the merge of #402 — the PR
    that documents the first accidental close.

    The commit message quoted the original phrase, "is the one that closes
    #110"
    , with a line break between closes and #110. GitHub matches close
    keywords across whitespace, so the quotation did the same thing the original
    sentence did. I checked the message for the phrase before merging and read it
    as safe because of the line break; it was not.

    Rule for anything that touches this issue from here: do not put the word
    closes (or fixes, resolves) in front of #110 in a commit message or PR
    body, quoted or not, until the shelf is racked. Say "the step that finishes the
    issue" or reference it as Refs #110. The roadmap paragraph merged in #402
    carries the quoted phrase inside a file, which GitHub does not read, and stays.

    State of the work is unchanged: shelf on hand, not racked, switch on a wall
    socket.

    🤖 Generated with Claude Code

  6. Gerrrt commented on Sep 8, 2026

    @Gerrrt
    OwnerAuthor

    Done, 2026-09-08. The shelf is in U4, the TP-Link is on it with its uplink back
    on port 3 of neo, and its power is on a UPS-fed outlet — the step this issue
    was actually about.

    Seen from the monitoring host during the move

    oracle's link out 18:33:04 UTC
    oracle's link back 19:01:07 UTC — 28 minutes
    neo port 3 up, ifOperStatus=1
    Scrape targets 22 of 22 up within a minute of the link returning; oracle pushing again
    Silence 45 minutes on `severity=~"warning
    Paged nothing

    The 28 minutes crossed every for window that matters — the SNMP and gateway
    checks at 10 minutes, oracle's push staleness at 15 — which is what the silence
    was for. Two alerts outlived it briefly and cleared on their own: the wiki probe
    on oracle, and Alertmanager's own delivery failures from the disconnect.

    What the monitoring host cannot verify: that the outlet is UPS-fed. The UPS
    reports its output source, not which device draws from it. Attested by the
    person at the rack, and recorded in the docs on that basis.

    Docs in #406: the U4 row in
    hardware.md, the TP-Link note in network.md, the UPS runbook's banner, the
    roadmap entry to Done, and the README sentence. That PR closes this issue on
    merge, on purpose — the two earlier closes were a commit message quoting the
    phrase, and are recorded in the roadmap entry so the next reader knows which
    close was the real one.

    🤖 Generated with Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    featureA new capability or a change

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions