Files

Pim van Pelt d5fbf5c640 Prometheus: add VPP, LB sync, and gRPC metrics; expand docs

New metrics plus the corresponding documentation for everything that's
accumulated since the last Prometheus pass.

internal/metrics/metrics.go
- New VPPSource interface (IsConnected, VPPInfo) plus a metrics-local
  VPPInfo struct that mirrors vpp.Info. Decoupling via interface +
  struct-mirror keeps the dependency direction one-way (vpp → metrics),
  so vpp can import metrics to update inline counters without a cycle.
- New Collector gauges scraped on demand: maglev_vpp_connected,
  maglev_vpp_uptime_seconds (from /sys/boottime), maglev_vpp_connected_seconds
  (time since maglevd connected), and maglev_vpp_info (static 1-gauge
  carrying version, build_date, and pid as labels).
- New inline counters:
  - maglev_vpp_api_total{msg, direction, result} — bumped from the
    loggedChannel wrapper on every VPP binary-API send/recv. Gives full
    visibility into what maglevd is doing with VPP, broken down by
    message name, direction (send/recv), and result (success/failure).
  - maglev_vpp_lbsync_total{scope, kind} — bumped from the reconciler
    at the end of each SyncLBStateAll/SyncLBStateVIP run. kind ∈
    {vip_added, vip_removed, as_added, as_removed, as_weight_updated};
    scope ∈ {all, vip}. Zero-valued kinds are not emitted so noise
    stays low.
- Register() signature now takes a VPPSource (may be nil) alongside
  the existing StateSource.

internal/vpp/client.go
- New VPPInfo() (metrics.VPPInfo, bool) shim method on *Client that
  satisfies metrics.VPPSource. Returns (_, false) when disconnected so
  the collector skips the vpp_* gauges cleanly.

internal/vpp/apilog.go
- The loggedChannel's SendRequest / SendMultiRequest / ReceiveReply
  paths now call metrics.VPPAPITotal.WithLabelValues(...).Inc() in
  addition to slog.Debug. Since every VPP API call in the codebase
  must go through loggedChannel (NewAPIChannel is unexported), this
  one instrumentation point catches everything.

internal/vpp/lbsync.go
- New recordSyncStats(scope, st) helper called once at the end of
  SyncLBStateAll and SyncLBStateVIP to bump maglev_vpp_lbsync_total.
  Zero-valued stats are skipped.

cmd/maglevd/main.go
- Added github.com/grpc-ecosystem/go-grpc-middleware/providers/prometheus
  for the standard gRPC server metrics (grpc_server_started_total,
  grpc_server_handled_total, grpc_server_handling_seconds, etc.,
  labelled by service/method/type/code).
- Constructs grpcprom.NewServerMetrics(WithServerHandlingTimeHistogram())
  before creating the grpc.Server, installs it as UnaryInterceptor +
  StreamInterceptor, then calls InitializeMetrics(srv) after service
  registration so every method appears at 0 on the first scrape
  instead of materialising lazily on first RPC.
- Passes the vppClient (or nil) as a metrics.VPPSource to
  metrics.Register so the vpp_* gauges are emitted when integration
  is enabled and silently omitted otherwise.

docs/user-guide.md
- New 'Prometheus metrics' section in the maglevd chapter,
  tabulating every metric family: backend state gauges, probe
  counters/histogram, transition counters, the new VPP gauges and
  counters, and the standard gRPC server metrics.
- 'show frontends <name>' description updated to mention the two
  weight columns ('weight' = configured from YAML, 'effective' =
  state-aware after pool-failover logic).
- Pause / disable descriptions clarified: transition history is
  preserved across these operator actions.

docs/healthchecks.md
- New 'Static (no-healthcheck) backends' section explaining that
  backends without a healthcheck use rise/fall=1, fire a synthetic
  passing probe immediately on startup (no 30s wait), and idle at
  30s between iterations thereafter.
- New 'Pool failover' section documenting the priority-tier model,
  the active-pool definition, when promotion happens, cascading to
  further tiers, and graceful drain on demotion. Points readers at
  'maglevc show frontends <name>' as the inspection interface.

docs/config-guide.md
- healthcheck field doc now describes static-backend behavior and
  cross-references healthchecks.md.
- pools field doc now explains failover semantics at a high level
  and cross-references the detailed healthchecks.md section.

2026-04-12 13:00:35 +02:00

8.7 KiB

Raw Blame History

Health Checking

maglevd probes each backend independently of how many frontends reference it. Every backend runs exactly one probe goroutine. State changes are broadcast as gRPC events to all connected WatchEvents subscribers.

States

State	Meaning
`unknown`	Initial state; also entered after a resume or enable.
`up`	Backend is healthy and eligible to receive traffic.
`down`	Backend has failed enough consecutive probes to be considered offline.
`paused`	Health checking stopped by an operator. No probes are sent.
`disabled`	Backend disabled by an operator. No probes are sent.
`removed`	Backend removed from configuration by a reload. No probes are sent.

Rise / fall counter

The state machine is driven by HAProxy's single-integer health counter.

counter ∈ [0, rise + fall − 1]   (called Max below)

backend is UP   when counter ≥ rise
backend is DOWN when counter < rise

On each probe:

pass — counter increments, ceiling at Max.
fail — counter decrements, floor at 0.

This gives hysteresis: a backend that is barely up (counter = rise) needs fall consecutive failures before it transitions to down. A backend that is fully down (counter = 0) needs rise consecutive passes to come back up. A backend that oscillates between passing and failing stays in the degraded range without bouncing between up and down.

Expedited unknown resolution

When a backend enters unknown state (new, restarted, resumed, or re-enabled) its counter is pre-loaded to rise − 1. This means a single probe result is enough to resolve the state:

1 pass → up
1 fail → down (also via the special unknown shortcut below)

In addition, any failure while state is unknown transitions immediately to down, regardless of the counter value.

Example: rise=2, fall=3 (Max=4)

counter:  0   1   2   3   4
state:   DOWN DOWN  UP  UP  UP
                ^
                rise boundary

A backend starting from unknown has counter=1 (rise−1). One pass → counter=2 → up. One fail while unknown → down immediately.

A backend that just became up sits at counter=2. It needs 3 failures to go down (2→1→0, crossing the rise boundary at 2→1).

A backend that has been fully healthy for a while sits at counter=4. It needs 3 failures to go down (4→3→2→1, crossing the rise boundary at 2→1).

Probe intervals

The interval used between probes depends on the backend's counter state:

Condition	Interval used
State is `unknown`	`fast-interval` (falls back to `interval`)
Counter = Max (fully healthy)	`interval`
Counter = 0 (fully down)	`down-interval` (falls back to `interval`)
Counter between 0 and Max (degraded)	`fast-interval` (falls back to `interval`)

Using fast-interval in degraded and unknown states means a flapping or recovering backend is re-evaluated quickly without waiting a full interval. Using down-interval for fully down backends reduces probe traffic to servers that are known to be offline.

Transition events

Every state change is logged as backend-transition and emitted as a gRPC BackendEvent to all active WatchEvents streams.

Backend added (config load or reload)

unknown → unknown  (code: start)

The counter is pre-loaded to rise − 1. The first probe fires after fast-interval (or interval if not configured). One pass produces unknown → up; one fail produces unknown → down.

If multiple backends start together they are staggered across the first interval to avoid probe bursts.

Probe pass

Counter increments.
If counter reaches rise from below: down → up (or unknown → up).
If already up: no transition. Next probe at fast-interval if degraded, interval if fully healthy.

Probe fail

Counter decrements.
If counter drops below rise from above: up → down.
If state is unknown: transition immediately to down regardless of counter.
Next probe at down-interval if fully down, fast-interval if degraded.

Pause

<any> → paused  (operator action)

The counter is reset to 0. The probe goroutine is cancelled — no further probes are sent and no traffic reaches the backend while it is paused. The backend stays paused until explicitly resumed.

Resume

paused → unknown  (operator action)

The counter is reset to rise − 1. A fresh probe goroutine is started, which fires its first probe after fast-interval (or interval if not configured). One pass produces unknown → up; one fail produces unknown → down.

Disable

<any> → disabled  (operator action)

The probe goroutine is cancelled and the backend is marked enabled: false. No further probes are sent. The backend remains visible via the gRPC API (state disabled) and can be re-enabled without a config reload.

Enable

disabled → unknown  (operator action, via fresh goroutine)

A new probe goroutine is started and the backend re-enters unknown with the counter pre-loaded to rise − 1. The enabled flag is set back to true. The first probe fires after fast-interval and resolves state as described under Backend added.

Backend removed (config reload)

<any> → removed  (code: removed)

The probe goroutine stops. No further state changes occur. The removed event is emitted using the frontend map from before the reload so that consumers can correlate it to the correct frontend.

Backend healthcheck config changed (config reload)

The old probe goroutine is stopped (<any> → removed) and a new one started (unknown → unknown, code: start). The new goroutine resolves state on the first probe as described under Backend added above.

Backend metadata changed without healthcheck change (config reload)

Weight, enabled flag, and similar fields are updated in place. The probe goroutine is not restarted and no transition event is emitted.

Static (no-healthcheck) backends

A backend with no healthcheck field in YAML skips the probe loop entirely. Instead of actually probing, maglevd synthesises a single passing result on startup. Specifically:

The worker's rise/fall counters are forced to 1/1, so a single synthetic pass is enough to reach StateUp.
The first "probe" fires immediately (zero sleep). Subsequent iterations idle at 30 seconds — there is nothing to do.
The backend reaches up within milliseconds of startup.

Static backends are useful for administrative VIPs where the caller knows the backend is always available, or for test configurations where deterministic state is more valuable than real health signals.

Pool failover

Every frontend has one or more pools. The pools are priority tiers: pool[0] is the primary, pool[1] is the first fallback, pool[2] the next, and so on. At any moment, maglevd computes an active pool — the first pool that contains at least one backend in StateUp:

As long as pool[0] has any up backend, it stays active. Its up backends receive traffic at their configured weights; backends in lower-priority pools stay on standby with effective weight 0.
When pool[0] has zero up backends (all down, paused, disabled, or still unknown), pool[1] is promoted: its up backends get their configured weights, and pool[0] backends stay at 0 until at least one recovers.
The same rule cascades to pool[2], pool[3], etc., for further fallback tiers.
When no pool has any up backend, every backend's effective weight is 0 and the VIP serves nothing.

Failover is evaluated on every backend state transition and also on the periodic VPP drift reconciliation (every maglev.vpp.lb.sync-interval). The resulting effective weight for each backend can be inspected via maglevc show frontends <name> — each pool backend row shows both the configured weight and the effective weight after failover.

Demotion on recovery (e.g. pool[1] → standby when pool[0] comes back up) drains gracefully: the demoted backends have their weight set to 0 but existing flows in the VPP flow table are left to drain naturally. The only state that forces immediate flow-table flushing is operator disable.

Log lines

All state changes produce a structured log line at INFO level:

{"level":"INFO","msg":"backend-transition","backend":"nginx0-ams","from":"up","to":"paused"}
{"level":"INFO","msg":"backend-transition","backend":"nginx0-ams","from":"paused","to":"unknown"}
{"level":"INFO","msg":"backend-transition","backend":"nginx0-ams","from":"unknown","to":"up","code":"L7OK","detail":""}

Probe-driven transitions also carry code and detail fields from the probe result (e.g. L4CON, L7STS, connection refused). Operator-driven transitions (pause, resume, disable, enable) carry empty code and detail.

8.7 KiB Raw Blame History Unescape Escape