Applies the 11 recommendations from the codebase optimization review.
Module decomposition:
- Extract Towerops.Topology.InferenceEngine: pulls device-role inference
(capability rules, vendor matching, evidence merging) out of the topology
context. Topology now delegates merge_confidence/2,
infer_role_from_capabilities/1, and infer_device_role/1.
- Extract Towerops.Accounts.TOTP: TOTP secret/QR generation, multi-device
verification, recovery code lifecycle, and sudo MFA logic. Accounts
delegates the public TOTP API for backwards compatibility.
- Extract Towerops.Accounts.Sessions: browser session create/list/touch/
revoke/anonymize/expire. Accounts delegates the public session API.
- Extract Towerops.Snmp.Queries: time-series read paths
(sensor readings, interface stats, latest-per-id batches).
- Extract Towerops.Snmp.Monitoring: time-series write paths
(sensor/interface/processor reading inserts, batch inserts via
Repo.insert_all). Snmp.ex shrinks from 2985 to 2742 lines.
DRY refactoring:
- Add per-schema Query modules (Devices.DeviceQuery, Alerts.AlertQuery)
with composable for_organization/for_site/with_status/etc filters;
refactor list_site_devices, count_organization_devices,
count_site_devices, count_site_devices_down, count_organization_alerts
to compose them.
- Introduce create_discovered_checks/6 helper in Snmp; collapses 4 nearly
identical Enum.reduce blocks for sensors/interfaces/processors/storage.
- Unify MAC and ARP evidence collectors behind collect_lookup_evidence/3
(lldp/cdp were already factored).
Pattern matching:
- Replace cond block in infer_role_from_capabilities/1 with a declarative
@role_rules table consulted via Enum.find_value (Elixir 1.19 forbids `in`
with runtime lists in guards, so multi-clause functions weren't an option).
- Replace inline `if since` in get_sensor_readings/2 with a maybe_filter_since
pipe helper.
Performance:
- Replace Enum.each + Repo.insert per evidence row in upsert_link with a
single Repo.insert_all batch (truncates :observed_at to seconds).
Human-centric:
- Decompose build_device_lookup/1 into fetch_lookup_devices, map_ips_to_ids,
map_names_to_ids, map_macs_to_ids.
- Adopt Towerops.Result.map/2 in Snmp.Client.do_get_multiple_sequential
(kept narrow; `with` is preferred over Result.and_then for chained
operations and is already used widely).
CLAUDE.md updates:
- Replace stale main→staging / production-branch deployment notes with the
current flow: PRs deploy to Dokku staging, push to main runs the
ExUnit gate then builds an image; argocd-image-updater rolls production.
- Refresh data-model diagram and references from "Equipment" to "Device"
(post equipment→devices rename migration).
- Correct stale architecture facts: Hammer → in-tree Towerops.RateLimit,
Rust NIF → C NIF (libnetsnmp), gitlab-registry → forgejo-registry,
expand the Oban queue and cron lists, list actual implemented Ecto types.
All 10,228 tests pass.
Bumps vendored Oban Pro from 1.6.13 to 1.7.0, which transitively bumps
oban core to 2.22.1 (requires migrations v14) and db_connection to 2.10.0.
- mix.exs constraint: ~> 1.6 → ~> 1.7
- vendor/oban_pro/ replaced with v1.7.0 hex package
- migration 20260430142200: Oban core v12 → v14 (must run first)
- migration 20260430142241: Oban Pro 1.7.0 schema additions
Also corrects a pre-existing constraint name in AgentAssignment that was
left stale by the equipment→devices rename migration (20260117190134).
The unique index is named agent_assignments_device_id_index, but the
schema's unique_constraint/3 still pointed at the old equipment_id name,
causing the test suite to surface Ecto.ConstraintError instead of a
changeset error. Required to get the suite green for this upgrade.
With 1 replica, the old pod was torn down too quickly after the new pod
passed its first health check. Increase successThreshold to 3 (15s of
consecutive checks) and minReadySeconds to 30 to ensure Traefik has fully
propagated the new endpoint before the old pod is terminated.
Apply patterns from testingliveview.com — replace render_click/submit/change
event-name calls with element/form-targeted equivalents that validate the
actual DOM wiring, not just handler logic. Adds stable test selectors
(id="...") to interactive elements across 10 LiveView templates.
Surfaces (and fixes) several real issues that the bypass-style tests masked:
- Removes dead regenerate_token handler in agent_live/index.ex (no UI ever
triggered it) and its corresponding test.
- Removes dead refresh_topology test in network_map_live_test (event was
referenced only in the test — no handler, no button, no JS anywhere).
- Corrects "shows notification tab with add device modal" — modal lives on
the security tab; original test mounted ?tab=notifications and only
worked because direct event calls bypassed tab gating.
- Corrects form input shape mismatches in agent_live/index_test that the
handler's defensive case clauses had been masking.
- Expands setup data for org/settings_live_test (snmp_community,
default_agent_token_id, multi-org membership) so the gated buttons
actually render.
11 direct render_click/submit/change calls remain — all intentional,
documented server-side guard tests (tampered POSTs, race conditions,
defensive idempotent handlers).
Full LiveView test suite: 1300 tests, 0 failures.
The HelmRelease/HelmRepository CRs targeted flux-system; with flux uninstalled
they were never being applied. The gitlab-agent workload is not deployed on
the cluster.
ArgoCD refuses to clone repos containing tracked symlinks that point outside
the working tree. The .pre-commit-config.yaml symlink resolved to /nix/store/...
on the local dev machine and broke remote checkouts. Generate it locally via
direnv/nix instead.
Performance:
- schedule_live: preload page once instead of get_schedule!/1 per row
- alert_live + alerts: DB-side status filter; Repo.aggregate counts replace length/Enum.count over 500-row fetches on every event
- dashboard_live: drop duplicate get_device_status_counts; cap active alerts to 20 + use count_active_alerts/1
- maintenance.active_windows_for_device: 3 round-trips collapsed into one query (per-site-id branch keeps Postgres parameter types unambiguous)
- sites.build_site_tree: O(N^2) -> O(N) via group_by(parent_site_id)
- accounts.sole_owner_organizations: single group_by + having instead of per-org Repo.aggregate loop
- agents + agent_live (org + admin): count_assigned_devices_batch/1 + count_agent_polling_targets/1 (no preloads)
- alert_digest_worker: list_alerts_by_ids/1 batches digest fetch
- gaiia: distinct: true at DB; limit 50 on bidirectional ilike
Indexes:
- maintenance_windows(organization_id, starts_at, ends_at) WHERE suppress_alerts = true
- alerts(check_id) WHERE resolved_at IS NULL
Antipatterns:
- agents.delete_agent_token: PubSub.broadcast moved outside Repo.transaction so a rollback no longer leaves subscribers acting on a non-existent deletion
- integrations_controller.to_atom_keys: replaced String.to_existing_atom on user-controlled JSON keys with explicit allowlist
- 9 Task.start callsites converted to Task.Supervisor.start_child(Towerops.TaskSupervisor, ...) so background DB writes survive shutdown (test-mode discovery shims left as-is for sandbox semantics)
Captures a high-level review of module decomposition, context-layer
boundaries, query patterns, error handling, and test coverage gaps to
inform future refactor work.
Three deferred followups from the chunk-4 plan, all together:
A. Restrictions editor
- Resolver: time_of_week support (weekday set + time-of-day window),
in addition to the existing time_of_day.
- New OnCall.update_layer_restrictions/2 + clear_layer_restrictions/1.
- schedule_live/show: per-layer Restrict on-call shifts form
(Always on call / Time of day / Time of week with Mon-Sun checkboxes),
changes patched live and persisted via the resolver.
B. Drag-and-drop reorder
- New OnCall.reorder_layers/2 + reorder_escalation_rules/2 (atomic
transaction; validates id count + membership).
- assets/js/sortable_list.ts: minimal HTML5 drag-and-drop hook (no
external dep; mirrors the existing DeviceListReorder pattern),
registered as SortableList in app.ts.
- schedule_live/show + escalation_policy_live/show: layer/level cards
now carry data-id + draggable + a visible drag handle. Up/down arrow
buttons remain as a fallback for users on assistive tech.
C. Unified new-schedule UX
- schedule_live/show pre-opens the Add Layer form when the schedule
has zero layers, so the new-schedule -> add-layers flow lands ready
to act on. Two existing tests updated to seed a layer first; two
new tests verify the auto-open behaviour both ways.
Full suite 10,231 / 0 failures. Format/dialyzer clean.
ScheduleLive.Index
- Name search input above the date nav (phx-debounce 300ms).
- Standard <.pagination> footer (10 per page) under the gantt cards.
- Both reflected in URL params (?search= and ?page=) for shareable
links; defaults are stripped from the URL.
- Refactored: load_schedules_with_timeline/3 -> hydrate_schedules/3
so filter/paginate/hydrate are clearly separate steps.
Tests
- 2 new ExUnit integration tests (search filter narrows results +
search input renders).
- e2e/tests/schedules.spec.ts: switched from table-row selectors to
card-link selectors (schedules tab is now gantt cards, not a table),
added smoke tests for date nav controls + search input + REPEAT
footer + Services sidebar, renamed "Add Rule" -> "Add Level".
Full suite 10,216 / 0 failures. Format/dialyzer clean.
OnCall context
- move_layer/2: up/down position swap for schedule layers, mirror of
move_escalation_rule/2 (tested for both directions + boundary no-op).
ScheduleLive.Show
- New move_layer event handler.
- Up/down arrow buttons on each layer card; the top arrow is disabled on
the first layer, the bottom on the last (matches the escalation policy
level reorder UX shipped in chunk 1).
- Replaced two per-hour walks (build_layer_spans + build_final_spans -
~672 resolve calls per render at 14 days) with Resolver.resolve_range/3
+ a small segments_to_spans/3 mapper. Per-layer view wraps each layer
in a temporary single-layer %Schedule{} so the resolver code path is
shared and overrides don't bleed into the per-layer strip.
Tests: 3 new context tests for move_layer, 1 new integration test for
the layer reorder UI. Full suite 10,214 / 0 failures. Format/dialyzer
clean.
Resolver
- Resolver.resolve_range/3: returns the contiguous on-call segments for a
schedule across [start_at, end_at). Walks layer handoffs + override
boundaries, drops gap segments, and merges adjacent same-user segments.
UserColor module
- New Towerops.OnCall.UserColor with deterministic id -> color mapping
(Tailwind class + matching #RRGGBB hex). Refactored ScheduleLive.Show
to use it instead of its own per-page index palette so the same person
gets the same color across the schedules list and detail pages.
ScheduleLive.Index
- Date navigation: Today / prev / next + 1/2/4-week range selector.
- URL-driven: ?start=YYYY-MM-DD&range=N for shareable views; falls back
to defaults on malformed values.
- Per-row card: schedule name link, on-call-now avatar swatch, day-of-week
header strip, gantt strip rendered as a CSS grid with one column per day,
Today marker overlay.
Tests: 5 new resolver tests, 9 new UserColor tests, 5 new ScheduleLive
integration tests. Full suite: 10,210 / 0 failures. Format/dialyzer clean.
Also: cleaned up two stale inline comments in k8s/deployment.yaml that
disagreed with the values they sat next to (replicas/maxSurge).
- k8s/deployment.yaml: bump startup-probe failureThreshold from 12 to 24
(70s -> 130s max startup time) to give the app more headroom on cold start.
- lib/towerops/profiles/yaml_profiles.ex: defer YAML profile loading from
init/1 to a handle_info(:load_profiles, ...) so the supervisor (and Bandit)
come up immediately and the profiles populate in the background within ~60s.
- e2e/Makefile: convenience target wrappers around the existing npm scripts.
Chunk 1 of the on-call/escalation redesign plan
(docs/plans/2026-04-28-on-call-pagerduty-style-redesign.md):
Schema
- on_call.ex: list_devices_using_policy/2 (direct + org-default inheritance)
- on_call.ex: move_escalation_rule/2 for up/down level reordering
- list_escalation_policies/1 now preloads :rules
Show page
- Levels rendered as numbered cards with up/down/delete controls
- "Notify the following users immediately and escalate after N minutes."
- REPEAT footer surfaces repeat_count inline (no longer a standalone column)
- Right-side Services sidebar lists devices using the policy
- Header line shows the handoff notifications mode (when in use by a service /
always / never)
Edit form
- Removed standalone Repeat Count input
- Added handoff notifications mode select
- Added inline "If no one acknowledges, repeat this policy N time(s)" block
Index views (policies tab + standalone)
- Replaced Repeat Count column with a Levels (rule count) column
Other
- rate_limit.ex: bind unmatched :ets.new/2 + schedule_clean/1 returns to fix
dialyzer :unmatched_returns warning under the project's strict flags.
All 10,191 tests pass; mix format clean; mix compile --warnings-as-errors clean;
mix dialyzer passes.
* UISP diff_snapshots test: backdate first snapshot via UPDATE instead of
Process.sleep(1_100) to get distinct second-precision snapshot_at values.
* PagerDuty.Client.send_event_with_retry/2: check max_retries BEFORE sleeping
on 429. Previously a 429 response always slept once (~1s) before checking
the retry cap, so even with @max_retries = 0 in test env every 429 test
paid that cost.
Suite time: 65.3s → 33.9s (-48%). No test is >400ms now.
Req's default `retry: :safe_transient` waits ~7s across exponential backoff
when the stubbed response is a 5xx or connection error. Tests that asserted
error-status behavior paid that price on every run.
Disable retries in test env for every direct Req caller:
* `Towerops.HTTP` (covers weather/netbox/geocoding/http_executor/pagerduty)
* `CnMaestro.Client`, `Gaiia.Client`, `Preseem.Client`, `Sonar.Client`,
`Splynx.Client`, `Agents.ReleaseChecker`, `Billing.StripeClient`
Use `Keyword.put_new(:retry, false)` so any test explicitly opting into
retry behavior is respected.
Suite time: 82.8s → 65.3s. Slowest test dropped from 7050ms to 1447ms.
Promoted pure presentation and utility helpers from `defp` to `def @doc false`
across ~20 LiveViews, Oban workers, and sync modules so they're reachable from
unit tests. Refactored several `cond` blocks into idiomatic function heads with
guards. Added ~250 new test cases in new files under test/towerops and
test/towerops_web, including DB-backed tests for CnMaestro.Sync and
AlertNotificationWorker, and removed dead LiveView tab components and
CapacityLive (no callers anywhere in lib/test).
Configured mix.exs test_coverage.ignore_modules to exclude vendored third-party
code (SnmpKit, protobuf-generated Towerops.Agent.*, Absinthe GraphQL types,
Phoenix HTML modules, Inspect protocol impls) from coverage calculations —
these are not our project code.
Coverage: 66.93% → 70.09%. Full suite: 10,127 tests, 0 failures.
The remaining @dialyzer {:no_opaque, ...} / {:nowarn_function, ...}
directives in accounts.ex, devices.ex, organizations.ex all masked the
same upstream issue: Ecto.Multi.new/0 produces a struct containing a
literal %MapSet{map: %{}} whose concrete shape violates MapSet's
@opaque t at every subsequent Ecto.Multi.* call boundary.
Tried first: public helper function with @spec Ecto.Multi.t() — confirmed
dialyzer uses success typing of the body, not the spec. Doesn't help.
Real fix: drop Ecto.Multi entirely for these transactions. Each was
performing a small, sequential set of Repo operations that map cleanly
to a plain `Repo.transaction(fn -> ... end)` block with Repo.rollback
on errors. No Multi machinery is actually needed:
- accounts.ex:
- register_user/1: insert + 0-2 consent grants
- register_user_with_organization/1: insert + create_org + consents
- confirm_user_transaction/1: update + delete_all tokens
- organizations.ex:
- do_create_organization/4: lock user → limit check → insert org + membership
- set_default_organization/2: two update_alls
- accept_invitation/2: update invitation + insert membership
- devices.ex:
- create_device/2: maybe lock org → maybe quota check → insert
Caught a real consent-failure rollback bug along the way: initial
refactor discarded the return of grant_consents_sequential/2, which
meant FK violations on consent insert would abort the transaction at
the Postgres level but the caller still saw {:ok, user} — later
operations in the transaction silently failed with 25P02. Fixed by
propagating the {:error, changeset} up so Repo.rollback is called.
Verified by the AccountsDefensiveCoverageTest regression tests.
Behavior is preserved — same transactional guarantees, same success
/ error tuple shapes. Net -86 LOC and no suppressions.
mix dialyzer: `done (passed successfully)` with zero @dialyzer
directives in lib/.
Changes to eliminate @dialyzer suppressions by fixing underlying causes:
NIF stubs (towerops_native.ex, mib_translator.ex):
- Change stubs to :erlang.nif_error(:nif_not_loaded) (no_return type).
Real NIF replaces stubs at load time; calls to unloaded stubs now fail
loudly instead of returning fake data. Lets dialyzer trust @spec.
- Remove @dialyzer :nowarn_function on three NIFs and on translate/1.
Discovery sync_* functions (snmp/discovery.ex, channels/agent_channel.ex):
- agent_channel passes %{device_id: _, interfaces: _} and %{id: _} maps
into Discovery.sync_ip_addresses/sync_processors/sync_storage, which
@spec'd only %Device{}. Add narrow map-type unions (ip_sync_device,
snmp_device_ref) reflecting what the functions actually access.
- Remove @dialyzer :nowarn_function on three agent_channel helpers.
remote_ip.ex — real bug caught and fixed:
- `:ranch.get_addr(socket.transport_pid)` was always raising since
Bandit uses ThousandIsland, not Ranch; the rescue _ -> nil silently
returned nil every time. Switched to Phoenix's documented
:peer_data connect_info (already enabled in endpoint.ex) via
socket.assigns; remote IP now actually works.
- Remove remote_ip.ex entry from .dialyzer_ignore.exs.
Accounts / Organizations (Ecto.Multi opacity):
- Add @specs to Multi-building helpers, refactor into pipe chains.
- 6 @dialyzer :nowarn_function → 0, but 7 :no_opaque remain. Root
cause is upstream: Ecto.Multi.new/0 returns a struct with a literal
%MapSet{} whose @opaque internal representation trips dialyzer on
every subsequent Multi.* call. Unfixable without an Ecto patch or
bypassing Multi entirely. Comments document the specific upstream
issue rather than a vague "Ecto.Multi opacity" claim.
Devices.ex:
- Adding @specs made it worse (call_without_opaque → contract_with_
opaque); inlining the Multi didn't help either — same MapSet root
cause. Suppression kept with a sharper comment.
Categories addressed:
- pattern_match / pattern_match_cov (32): remove dead case/with clauses
that dialyzer proved unreachable from the caller types.
- contract_supertype / extra_range / invalid_contract /
contract_with_opaque (25): narrow @spec declarations to match actual
success typings.
- call / call_without_opaque (18): fix bad calls, narrow User.t to
allow nil for in-memory changeset structs, suppress Ecto.Multi
opaque-type false positives with targeted @dialyzer directives.
- guard_fail / no_return / unused_fun / unknown_function (13): remove
dead || fallbacks, simplify always-true params, cascade-resolve
no_returns via the underlying pattern_match and call fixes.
Real production bug fixed: StormDetector.handle_cast/2 had swapped
`:queue.in` args (`queue |> :queue.in(ts)` which desugars to
`:queue.in(queue, ts)` — wrong argument order). Alert timestamps
were never being enqueued, so storm detection would fail at runtime.
Corrected to `ts |> :queue.in(queue)`.
.dialyzer_ignore.exs: suppress two genuine dep-PLT gaps
(:ranch.get_addr/1 false positive from Bandit's transitive ranch,
and the Cloak.Vault GenServer callback_info on the CI build path).
`mix dialyzer` now: Total errors: 114, Skipped: 114 — passes clean.
Warnings: 88 → 0.
Prefix fire-and-forget calls (PubSub.broadcast, Oban enqueue, Logger,
update_*, etc.) with `_ =` so dialyzer knows the return value is
intentionally discarded. Wrap a few if/case expressions whose
branches produce mixed types the same way.
No behavior changes — only explicit acknowledgement of discarded
returns.
Warnings: 242 → 88.
- Add @type t :: %__MODULE__{} to schemas referenced by callers:
ApiToken, Monitoring.Check, OnCall.Schedule, SiteOutage,
NotificationDigest, Snmp.WirelessClient, Topology.DeviceLink,
Agent.AgentJob, Agent.MikrotikResult.
- Proto.Types: qualify forward references to nested modules with
Types. prefix, and fix Metric typo (should be union type metric()).
- data_loaders.ex / chart_builders.ex: Snmp.SNMPDevice.t was a ghost
spec — module is Snmp.Device.
Warnings: 328 → 242.
Ignore file was silencing every warning under lib/towerops/**. Root
cause was plt_add_deps: :apps_direct missing Plug/Phoenix/Oban/Decimal
/Redix/ssl/public_key — hundreds of false `unknown_function` warnings
were being papered over.
Expand plt_add_apps to cover the common deps, narrow the ignore file
to real dep-PLT gaps only, and fix two concrete bugs uncovered by the
change:
- ScopedResource.fetch_preload/4: spec was atom() | [atom()] but
callers pass keyword lists (e.g. [rules: :targets]).
- ToweropsNative: tag NIF stubs with @dialyzer :nowarn_function so
dialyzer trusts the @spec instead of the fallback body.
Remaining 328 real warnings surface for follow-up.
