Restart-neutral VPP LB sync; deterministic AS ordering; maglevt cadence; v0.9.5

Three reliability fixes bundled with docs updates.

Restart-neutral VPP LB sync via a startup warmup window
(internal/vpp/warmup.go). Before this, a maglevd restart would
immediately issue SyncLBStateAll with every backend still in
StateUnknown — mapped through BackendEffectiveWeight to weight
0 — and VPP would black-hole all new flows until the checker's
rise counters caught up, several seconds later. The new warmup
tracker owns a process-wide state machine gated by two config
knobs: vpp.lb.startup-min-delay (default 5s) is an absolute
hands-off window during which neither the periodic sync loop
nor the per-transition reconciler touches VPP; vpp.lb.
startup-max-delay (default 30s) is the watchdog for a per-VIP
release phase that runs between the two, releasing each frontend
as soon as every backend it references reaches a non-Unknown
state. At max-delay a final SyncLBStateAll runs for any stragglers
still in Unknown. Config reload does not reset the clock. Both
delays can be set to 0 to disable the warmup entirely. The
reconciler's suppressed-during-warmup events log at DEBUG so
operators can still see them with --log-level debug. Unit tests
cover the tracker state machine, allBackendsKnown precondition,
and the zero-delay escape hatch.

Deterministic AS iteration in VPP LB sync. reconcileVIP and
recreateVIP now issue their lb_as_add_del / lb_as_set_weight
calls in numeric IP order (IPv4 before IPv6, ascending within
each family) via a new sortedIPKeys helper, instead of Go map
iteration order. VPP's LB plugin breaks per-bucket ties in the
Maglev lookup table by insertion position in its internal AS
vec, so without a stable call order two maglevd instances on
the same config could push identical AS sets into VPP in
different orders and produce divergent new-flow tables. Numeric
sort is used in preference to lexicographic so the sync log
stays human-readable: string order would place 10.0.0.10 before
10.0.0.2, and the same problem in v6. Unit tests cover empty,
single, v4/v6 numeric vs lexicographic, v4-before-v6 grouping,
a 1000-iteration stability loop against Go's randomised map
iteration, insertion-order invariance, and the desiredAS
call-site type.

maglevt interval fix. runProbeLoop used to sleep the full
jittered interval after every probe, so a 100ms --interval
with a 30ms probe actually produced a 130ms period. The sleep
now subtracts result.Duration so cadence matches the flag.
Probes that overrun clamp sleep to zero and fire the next
probe immediately without trying to catch up on missed cycles
— a slow backend doesn't get flooded with back-to-back probes
at the moment it's already struggling.

Docs. config-guide now documents flush-on-down and the new
startup-min-delay / startup-max-delay knobs; user-guide's
maglevd section explains the restart-neutrality property, the
three warmup phases, and the relevant slog lines operators
should watch for during a bounce.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

internal/config/config.go

@@ -59,6 +59,30 @@ type VPPLBConfig struct {
 	// removed from the table. Must be between 1 and 120 seconds inclusive.
 	// Defaults to 40s.
 	FlowTimeout time.Duration
+
+	// StartupMinDelay is the absolute hands-off window at the start of
+	// the maglevd process. For the first StartupMinDelay seconds after
+	// startup, the VPP LB sync path makes no calls to VPP at all —
+	// neither the periodic SyncLBStateAll loop nor the per-transition
+	// SyncLBStateVIP path from the reconciler. This gives a restarting
+	// maglevd a chance to complete its first few probes before any VPP
+	// state is touched, so a bounce does not black-hole traffic while
+	// the checker is still warming up. Default 5s. Set to 0 together
+	// with StartupMaxDelay to disable the warmup entirely and sync VPP
+	// immediately on startup.
+	StartupMinDelay time.Duration
+
+	// StartupMaxDelay is the watchdog for the per-VIP release phase.
+	// Between StartupMinDelay and StartupMaxDelay, each VIP is released
+	// (and one SyncLBStateVIP runs against it) as soon as every backend
+	// it references has reached a non-Unknown state. At StartupMaxDelay
+	// the warmup driver unconditionally runs SyncLBStateAll to handle
+	// any stragglers whose backends are still Unknown — those get
+	// programmed with whatever weight their current state maps to,
+	// which for a still-Unknown backend is 0. Must be >= StartupMinDelay.
+	// Default 30s. Set to 0 together with StartupMinDelay to disable
+	// the warmup.
+	StartupMaxDelay time.Duration
 }
 
 // HealthCheck describes how to probe a backend.
@@ -161,6 +185,8 @@ type rawVPPLBCfg struct {
 	IPv6SrcAddress       string  `yaml:"ipv6-src-address"`
 	StickyBucketsPerCore *uint32 `yaml:"sticky-buckets-per-core"` // default 65536
 	FlowTimeout          string  `yaml:"flow-timeout"`            // Go duration; default 40s, [1-120]s
+	StartupMinDelay      *string `yaml:"startup-min-delay"`       // Go duration; default 5s; 0 disables
+	StartupMaxDelay      *string `yaml:"startup-max-delay"`       // Go duration; default 30s; must be >= startup-min-delay
 }
 
 type rawHealthCheck struct {
@@ -403,6 +429,41 @@ func convertVPP(r *rawVPPCfg, cfg *VPPConfig) error {
 		cfg.LB.FlowTimeout = 40 * time.Second
 	}
 
+	// startup-min-delay: absolute hands-off window at process start.
+	// Default 5s. May be 0 (no gate) but must not be negative.
+	if r.LB.StartupMinDelay != nil {
+		d, err := time.ParseDuration(*r.LB.StartupMinDelay)
+		if err != nil {
+			return fmt.Errorf("vpp.lb.startup-min-delay: %w", err)
+		}
+		if d < 0 {
+			return fmt.Errorf("vpp.lb.startup-min-delay must be >= 0")
+		}
+		cfg.LB.StartupMinDelay = d
+	} else {
+		cfg.LB.StartupMinDelay = 5 * time.Second
+	}
+
+	// startup-max-delay: watchdog for the per-VIP release phase. Default
+	// 30s. May be 0 (no warmup at all, together with min-delay=0), but
+	// must be >= min-delay so the per-VIP release phase is well-formed.
+	if r.LB.StartupMaxDelay != nil {
+		d, err := time.ParseDuration(*r.LB.StartupMaxDelay)
+		if err != nil {
+			return fmt.Errorf("vpp.lb.startup-max-delay: %w", err)
+		}
+		if d < 0 {
+			return fmt.Errorf("vpp.lb.startup-max-delay must be >= 0")
+		}
+		cfg.LB.StartupMaxDelay = d
+	} else {
+		cfg.LB.StartupMaxDelay = 30 * time.Second
+	}
+	if cfg.LB.StartupMaxDelay < cfg.LB.StartupMinDelay {
+		return fmt.Errorf("vpp.lb.startup-max-delay (%s) must be >= startup-min-delay (%s)",
+			cfg.LB.StartupMaxDelay, cfg.LB.StartupMinDelay)
+	}
+
 	// A missing src address is a hard error: VPP's GRE encap needs a source,
 	// and every VIP we program uses GRE. Fail the config check so the
 	// operator cannot start maglevd with a broken setup.

internal/vpp/client.go

@@ -66,6 +66,12 @@ type Client struct {
 	// lbStatsLoop. Published as an immutable slice via atomic.Pointer so
 	// Prometheus scrapes (metrics.Collector.Collect) don't take any lock.
 	lbStatsSnap atomic.Pointer[[]metrics.VIPStatEntry]
+
+	// warmup gates every VPP LB sync call (both periodic and event-
+	// driven) during the first StartupMaxDelay seconds after Client
+	// construction. See warmup.go for the state machine. Process-wide,
+	// not per-connection: reconnects do not re-enter warmup.
+	warmup *warmupTracker
 }
 
 // SetStateSource attaches a live config + health state source. When set, the
@@ -86,9 +92,18 @@ func (c *Client) getStateSource() StateSource {
 	return c.stateSrc
 }
 
-// New creates a Client for the given socket paths.
+// New creates a Client for the given socket paths. The warmup tracker's
+// clock starts here — the restart-neutrality window is measured from the
+// moment the Client is constructed, which in practice is a few tens of
+// milliseconds after process start (see cmd/maglevd/main.go startup
+// sequence). If main.go ever grows a long-running initialisation step
+// before vpp.New(), the warmup clock should be moved accordingly.
 func New(apiAddr, statsAddr string) *Client {
-	return &Client{apiAddr: apiAddr, statsAddr: statsAddr}
+	return &Client{
+		apiAddr:   apiAddr,
+		statsAddr: statsAddr,
+		warmup:    newWarmupTracker(),
+	}
 }
 
 // Run connects to VPP and maintains the connection until ctx is cancelled.
@@ -165,19 +180,48 @@ func (c *Client) Run(ctx context.Context) {
 	}
 }
 
-// lbSyncLoop periodically runs SyncLBStateAll to catch drift between the
-// maglev config and the VPP dataplane. The first run happens immediately
-// on loop start (VPP has just connected, so any pre-existing state needs
-// reconciliation). Subsequent runs fire every cfg.VPP.LB.SyncInterval.
-// Exits when ctx is cancelled.
+// lbSyncLoop drives the periodic VPP LB sync. On first entry (after the
+// first successful VPP connect) it runs the warmup phase via runWarmup,
+// which enforces the restart-neutrality window and handles the first full
+// sync itself. Subsequent reconnect entries find warmup.allDone == true
+// and skip straight to the periodic ticker. Exits when ctx is cancelled.
+//
+// The warmup phase is intentionally run from inside this loop rather
+// than from Run: it needs the state source registered (which happens
+// only after SetStateSource) and it wants to be cancelled by the same
+// connCtx that cancels the stats loop on disconnect, so a VPP drop
+// during warmup doesn't leak a goroutine.
 func (c *Client) lbSyncLoop(ctx context.Context) {
 	src := c.getStateSource()
 	if src == nil {
 		return // no state source registered; nothing to sync
 	}
 
-	// next-run timestamp starts at "now" so the first tick is immediate.
-	next := time.Now()
+	// Warmup phase: runs once per process. On the first successful
+	// VPP connect, runWarmup handles the entire window (min-delay
+	// hands-off, per-VIP release phase, final SyncLBStateAll at
+	// max-delay) and calls finishAll before returning. On any
+	// reconnect after that, the gate is already open and we skip
+	// straight to the periodic ticker. A VPP drop mid-warmup
+	// returns from runWarmup via ctx.Done without closing the gate;
+	// the next reconnect re-enters runWarmup, which re-reads the
+	// process-relative clock and picks up wherever it left off.
+	if !c.warmup.isAllDone() {
+		c.runWarmup(ctx)
+		if ctx.Err() != nil {
+			return
+		}
+	}
+
+	cfg := src.Config()
+	if cfg == nil {
+		return
+	}
+	interval := cfg.VPP.LB.SyncInterval
+	if interval <= 0 {
+		interval = defaultLBSyncInterval
+	}
+	next := time.Now().Add(interval)
 	for {
 		wait := time.Until(next)
 		if wait < 0 {
@@ -189,12 +233,12 @@ func (c *Client) lbSyncLoop(ctx context.Context) {
 		case <-time.After(wait):
 		}
 
-		cfg := src.Config()
+		cfg = src.Config()
 		if cfg == nil {
 			next = time.Now().Add(defaultLBSyncInterval)
 			continue
 		}
-		interval := cfg.VPP.LB.SyncInterval
+		interval = cfg.VPP.LB.SyncInterval
 		if interval <= 0 {
 			interval = defaultLBSyncInterval
 		}

internal/vpp/lbsync.go

@@ -3,11 +3,13 @@
 package vpp
 
 import (
+	"bytes"
 	"errors"
 	"fmt"
 	"log/slog"
 	"net"
 	"regexp"
+	"sort"
 	"strconv"
 	"strings"
 
@@ -21,6 +23,80 @@ import (
 	lb_types "git.ipng.ch/ipng/vpp-maglev/internal/vpp/binapi/lb_types"
 )
 
+// sortedIPKeys returns the keys of m in ascending numeric IP order. Used
+// by the AS iteration sites in reconcileVIP and recreateVIP to make the
+// sequence of lb_as_add_del calls deterministic across maglevd instances
+// on the same config.
+//
+// Why this matters for Maglev: VPP's LB plugin stores ASes in a vec in
+// insertion order and rebuilds the new-flow lookup table by walking that
+// vec. The per-AS permutation is a pure function of the AS address (so it
+// matches across instances), but when two ASes want the same bucket on
+// the same pass the tie is broken by whichever one comes first in the
+// vec. If maglevd issues its add calls in Go map iteration order — which
+// is randomised on every run — two independent maglevd instances with
+// bit-identical configs can push the same ASes into VPP in different
+// orders, and their lookup tables end up differing in tie-broken buckets.
+// Sorting here is the first half of the fix; a matching sort inside VPP's
+// lb_vip_update_new_flow_table closes the flap-history hole (where a
+// remove+re-add after steady state reuses freed slots in the as_pool in
+// locally-visited order) and is landing in a separate commit.
+//
+// The keys at every call site are IP literals from net.IP.String(). Sort
+// order is numeric, not lexicographic: lexicographic would put 10.0.0.10
+// before 10.0.0.2 (and 2001:db8::10 before 2001:db8::2), which is
+// correct for determinism but confusing to operators reading the sync
+// log. We parse each key back into a net.IP once, compare the canonical
+// 16-byte form, and group IPv4 before IPv6 so mixed-family frontends
+// read as "v4 block, v6 block" top-to-bottom. ParseIP always succeeds
+// here because the caller built the map key via net.IP.String() in the
+// first place.
+func sortedIPKeys[V any](m map[string]V) []string {
+	type kv struct {
+		key string
+		ip  net.IP
+	}
+	pairs := make([]kv, 0, len(m))
+	for k := range m {
+		pairs = append(pairs, kv{k, net.ParseIP(k)})
+	}
+	sort.Slice(pairs, func(i, j int) bool {
+		return compareIPNumeric(pairs[i].ip, pairs[j].ip) < 0
+	})
+	out := make([]string, len(pairs))
+	for i, p := range pairs {
+		out[i] = p.key
+	}
+	return out
+}
+
+// compareIPNumeric returns <0, 0, >0 in the three-way convention. IPv4
+// sorts before IPv6. Within each family the comparison runs against the
+// canonical fixed-width byte form (4 bytes for v4, 16 bytes for v6),
+// which makes byte ordering match numeric ordering. A nil input — which
+// should not happen given sortedIPKeys's call contract — sorts before
+// any non-nil address to keep the comparator total.
+func compareIPNumeric(a, b net.IP) int {
+	a4 := a.To4()
+	b4 := b.To4()
+	switch {
+	case a == nil && b == nil:
+		return 0
+	case a == nil:
+		return -1
+	case b == nil:
+		return 1
+	case a4 != nil && b4 == nil:
+		return -1
+	case a4 == nil && b4 != nil:
+		return 1
+	case a4 != nil: // both v4
+		return bytes.Compare(a4, b4)
+	default: // both v6
+		return bytes.Compare(a.To16(), b.To16())
+	}
+}
+
 // ErrFrontendNotFound is returned by SyncLBStateVIP when the caller asks for
 // a frontend name that does not exist in the config.
 var ErrFrontendNotFound = errors.New("frontend not found in config")
@@ -241,8 +317,8 @@ func reconcileVIP(ch *loggedChannel, d desiredVIP, cur *LBVIP, curSticky bool, f
 			return err
 		}
 		st.vipAdd++
-		for _, as := range d.ASes {
-			if err := addAS(ch, d.Prefix, d.Protocol, d.Port, as); err != nil {
+		for _, addr := range sortedIPKeys(d.ASes) {
+			if err := addAS(ch, d.Prefix, d.Protocol, d.Port, d.ASes[addr]); err != nil {
 				return err
 			}
 			st.asAdd++
@@ -279,7 +355,8 @@ func reconcileVIP(ch *loggedChannel, d desiredVIP, cur *LBVIP, curSticky bool, f
 	}
 
 	// Remove ASes that are in VPP but not desired.
-	for addr, a := range curASes {
+	for _, addr := range sortedIPKeys(curASes) {
+		a := curASes[addr]
 		if _, keep := d.ASes[addr]; keep {
 			continue
 		}
@@ -290,7 +367,8 @@ func reconcileVIP(ch *loggedChannel, d desiredVIP, cur *LBVIP, curSticky bool, f
 	}
 
 	// Add new ASes, update weights on existing ones.
-	for addr, a := range d.ASes {
+	for _, addr := range sortedIPKeys(d.ASes) {
+		a := d.ASes[addr]
 		c, hit := curASes[addr]
 		if !hit {
 			if err := addAS(ch, d.Prefix, d.Protocol, d.Port, a); err != nil {
@@ -362,8 +440,8 @@ func recreateVIP(ch *loggedChannel, d desiredVIP, cur LBVIP, st *syncStats, reas
 		return err
 	}
 	st.vipAdd++
-	for _, as := range d.ASes {
-		if err := addAS(ch, d.Prefix, d.Protocol, d.Port, as); err != nil {
+	for _, addr := range sortedIPKeys(d.ASes) {
+		if err := addAS(ch, d.Prefix, d.Protocol, d.Port, d.ASes[addr]); err != nil {
 			return err
 		}
 		st.asAdd++

internal/vpp/lbsync_test.go

@@ -438,3 +438,207 @@ func TestDesiredFromFrontendSharedBackend(t *testing.T) {
 		})
 	}
 }
+
+// TestSortedIPKeysDeterministic pins the iteration-order helper that
+// reconcileVIP and recreateVIP use to sequence their lb_as_add_del
+// calls. The Maglev lookup table in VPP's LB plugin breaks per-bucket
+// ties by the order ASes sit in its internal vec, which is just the
+// order maglevd issued add calls — so if this helper ever stops
+// returning a total, stable ordering, two independent maglevd
+// instances on the same config can silently program different
+// new-flow tables.
+//
+// Sort order is numeric (by the parsed net.IP), not lexicographic.
+// The specific cases that a string sort would get wrong and a
+// numeric sort must get right:
+//
+//   - 10.0.0.2 < 10.0.0.10  (string sort puts "10" before "2")
+//   - 2001:db8::2 < 2001:db8::10 (same issue in v6)
+//   - all IPv4 before all IPv6 (operator-friendly grouping)
+func TestSortedIPKeysDeterministic(t *testing.T) {
+	t.Run("empty", func(t *testing.T) {
+		got := sortedIPKeys(map[string]int{})
+		if len(got) != 0 {
+			t.Errorf("empty map: got %v, want []", got)
+		}
+	})
+
+	t.Run("single entry", func(t *testing.T) {
+		got := sortedIPKeys(map[string]int{"10.0.0.1": 1})
+		if len(got) != 1 || got[0] != "10.0.0.1" {
+			t.Errorf("got %v, want [10.0.0.1]", got)
+		}
+	})
+
+	t.Run("v4 numeric order beats string order", func(t *testing.T) {
+		// The headline bug: "10.0.0.10" < "10.0.0.2" lexicographically
+		// because '1' < '2'. Numeric sort must place 2 before 10.
+		m := map[string]int{
+			"10.0.0.10": 1,
+			"10.0.0.2":  2,
+			"10.0.0.1":  3,
+			"10.0.0.11": 4,
+		}
+		got := sortedIPKeys(m)
+		want := []string{"10.0.0.1", "10.0.0.2", "10.0.0.10", "10.0.0.11"}
+		if len(got) != len(want) {
+			t.Fatalf("got %v, want %v", got, want)
+		}
+		for i := range want {
+			if got[i] != want[i] {
+				t.Errorf("pos %d: got %q, want %q", i, got[i], want[i])
+			}
+		}
+	})
+
+	t.Run("v6 numeric order beats string order", func(t *testing.T) {
+		// Same bug in v6: "2001:db8::10" < "2001:db8::2" lexicographically.
+		// The To16() canonical byte form handles both compressed and
+		// expanded forms correctly.
+		m := map[string]int{
+			"2001:db8::10": 1,
+			"2001:db8::2":  2,
+			"2001:db8::1":  3,
+		}
+		got := sortedIPKeys(m)
+		want := []string{"2001:db8::1", "2001:db8::2", "2001:db8::10"}
+		for i := range want {
+			if got[i] != want[i] {
+				t.Errorf("pos %d: got %q, want %q", i, got[i], want[i])
+			}
+		}
+	})
+
+	t.Run("v4 before v6", func(t *testing.T) {
+		// Mixed-family frontends: the operator-friendly order is
+		// the v4 block before the v6 block, each sorted numerically
+		// within its family.
+		m := map[string]int{
+			"2001:db8::1": 1,
+			"10.0.0.2":    2,
+			"10.0.0.1":    3,
+			"fe80::1":     4,
+			"192.168.0.1": 5,
+		}
+		got := sortedIPKeys(m)
+		want := []string{
+			"10.0.0.1", "10.0.0.2", "192.168.0.1",
+			"2001:db8::1", "fe80::1",
+		}
+		if len(got) != len(want) {
+			t.Fatalf("got %v, want %v", got, want)
+		}
+		for i := range want {
+			if got[i] != want[i] {
+				t.Errorf("pos %d: got %q, want %q", i, got[i], want[i])
+			}
+		}
+	})
+
+	t.Run("repeated calls produce identical sequence", func(t *testing.T) {
+		// Core determinism property: Go's map iteration is randomised,
+		// but sortedIPKeys must normalise it. Run the helper many
+		// times and compare every result to the first — if the
+		// normalisation ever breaks we'll see a divergence well within
+		// the loop count.
+		m := map[string]int{
+			"10.0.0.5": 1, "10.0.0.3": 2, "10.0.0.11": 3,
+			"10.0.0.2": 4, "10.0.0.4": 5, "10.0.0.20": 6,
+		}
+		first := sortedIPKeys(m)
+		for i := 0; i < 1000; i++ {
+			got := sortedIPKeys(m)
+			if len(got) != len(first) {
+				t.Fatalf("iter %d: length drift: got %v, first %v", i, got, first)
+			}
+			for j := range first {
+				if got[j] != first[j] {
+					t.Fatalf("iter %d pos %d: got %q, first %q", i, j, got[j], first[j])
+				}
+			}
+		}
+	})
+
+	t.Run("insertion order does not matter", func(t *testing.T) {
+		// A map built by inserting keys in ascending order must
+		// produce the same result as one built in descending order.
+		// Both go through the same normalisation.
+		asc := map[string]int{}
+		for _, k := range []string{"10.0.0.1", "10.0.0.2", "10.0.0.3", "10.0.0.10", "10.0.0.11"} {
+			asc[k] = 0
+		}
+		desc := map[string]int{}
+		for _, k := range []string{"10.0.0.11", "10.0.0.10", "10.0.0.3", "10.0.0.2", "10.0.0.1"} {
+			desc[k] = 0
+		}
+		gotAsc := sortedIPKeys(asc)
+		gotDesc := sortedIPKeys(desc)
+		if len(gotAsc) != len(gotDesc) {
+			t.Fatalf("length mismatch: asc %v, desc %v", gotAsc, gotDesc)
+		}
+		for i := range gotAsc {
+			if gotAsc[i] != gotDesc[i] {
+				t.Errorf("pos %d: asc %q, desc %q", i, gotAsc[i], gotDesc[i])
+			}
+		}
+	})
+
+	t.Run("desiredAS map", func(t *testing.T) {
+		// Exercise the actual call-site type: map[string]desiredAS.
+		// If the generic helper ever loses its type parameterisation
+		// this catches it at compile time (the call would fail).
+		m := map[string]desiredAS{
+			"10.0.0.9":  {Address: net.ParseIP("10.0.0.9"), Weight: 100},
+			"10.0.0.11": {Address: net.ParseIP("10.0.0.11"), Weight: 100},
+			"10.0.0.5":  {Address: net.ParseIP("10.0.0.5"), Weight: 50},
+			"10.0.0.1":  {Address: net.ParseIP("10.0.0.1"), Weight: 25},
+		}
+		got := sortedIPKeys(m)
+		want := []string{"10.0.0.1", "10.0.0.5", "10.0.0.9", "10.0.0.11"}
+		for i := range want {
+			if got[i] != want[i] {
+				t.Errorf("pos %d: got %q, want %q", i, got[i], want[i])
+			}
+		}
+	})
+}
+
+// TestCompareIPNumeric pins the ordering comparator that sortedIPKeys
+// delegates to. Split out so the v4/v6 boundary and nil-safety logic
+// have named failure modes rather than being buried in the map-based
+// subtests.
+func TestCompareIPNumeric(t *testing.T) {
+	cases := []struct {
+		name string
+		a, b net.IP
+		want int // -1, 0, +1 (sign of compareIPNumeric)
+	}{
+		{"v4 numeric asc", net.ParseIP("10.0.0.2"), net.ParseIP("10.0.0.10"), -1},
+		{"v4 numeric desc", net.ParseIP("10.0.0.10"), net.ParseIP("10.0.0.2"), 1},
+		{"v4 equal", net.ParseIP("10.0.0.1"), net.ParseIP("10.0.0.1"), 0},
+		{"v6 numeric asc", net.ParseIP("2001:db8::2"), net.ParseIP("2001:db8::10"), -1},
+		{"v6 numeric desc", net.ParseIP("2001:db8::10"), net.ParseIP("2001:db8::2"), 1},
+		{"v4 before v6", net.ParseIP("192.168.0.1"), net.ParseIP("2001:db8::1"), -1},
+		{"v6 after v4", net.ParseIP("2001:db8::1"), net.ParseIP("192.168.0.1"), 1},
+		{"nil before v4", nil, net.ParseIP("10.0.0.1"), -1},
+		{"v4 after nil", net.ParseIP("10.0.0.1"), nil, 1},
+		{"nil equal nil", nil, nil, 0},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := compareIPNumeric(tc.a, tc.b)
+			sign := func(x int) int {
+				switch {
+				case x < 0:
+					return -1
+				case x > 0:
+					return 1
+				}
+				return 0
+			}
+			if sign(got) != tc.want {
+				t.Errorf("got %d (sign %d), want sign %d", got, sign(got), tc.want)
+			}
+		})
+	}
+}

internal/vpp/reconciler.go

@@ -70,6 +70,15 @@ func (r *Reconciler) Run(ctx context.Context) {
 // Frontend-transition events are observational only — the dataplane work
 // they would imply has already been done by the backend-transition event
 // that triggered them.
+//
+// The handler consults the VPP client's warmup tracker before doing any
+// dataplane work. During the startup warmup window the reconciler is
+// either fully suppressed (inside min-delay) or per-VIP gated (the
+// frontend must have been released before events for it pass through).
+// When a transition fires for a VIP that isn't yet released but whose
+// backends have now all settled, the handler opportunistically releases
+// it here so the per-VIP release fires on the event rather than waiting
+// for the next warmup poll tick.
 func (r *Reconciler) handle(ev checker.Event) {
 	if ev.FrontendTransition != nil {
 		return // frontend-only event; no dataplane work
@@ -81,20 +90,66 @@ func (r *Reconciler) handle(ev checker.Event) {
 	if cfg == nil {
 		return
 	}
+
+	w := r.client.warmup
+	feName := ev.FrontendName
+	if !w.isReleased(feName) {
+		// Warmup is still gating this frontend. Decide whether to
+		// release it now, or defer until a later event / the warmup
+		// poll / the final max-delay SyncLBStateAll.
+		if w.inMinDelay() {
+			slog.Debug("vpp-reconciler-suppressed-min-delay",
+				"frontend", feName,
+				"backend", ev.BackendName,
+				"from", ev.Transition.From.String(),
+				"to", ev.Transition.To.String(),
+				"elapsed", w.elapsed(),
+				"reason", "inside vpp.lb.startup-min-delay window")
+			return
+		}
+		fe, ok := cfg.Frontends[feName]
+		if !ok {
+			return
+		}
+		if !allBackendsKnown(fe, r.stateSrc) {
+			slog.Debug("vpp-reconciler-suppressed-warmup",
+				"frontend", feName,
+				"backend", ev.BackendName,
+				"from", ev.Transition.From.String(),
+				"to", ev.Transition.To.String(),
+				"elapsed", w.elapsed(),
+				"reason", "frontend has backends still in StateUnknown; "+
+					"waiting for all to settle or for max-delay watchdog")
+			return
+		}
+		if !w.tryRelease(feName) {
+			// Lost a race with finishAll or another release caller.
+			// Either way the next isReleased call will return true, but
+			// for this event we've already done the right thing by
+			// letting the next few lines re-check and proceed.
+			return
+		}
+		slog.Info("vpp-lb-warmup-release",
+			"frontend", feName,
+			"trigger", "reconciler-event",
+			"backend", ev.BackendName,
+			"elapsed", w.elapsed())
+	}
+
 	slog.Debug("vpp-reconciler-event",
-		"frontend", ev.FrontendName,
+		"frontend", feName,
 		"backend", ev.BackendName,
 		"from", ev.Transition.From.String(),
 		"to", ev.Transition.To.String())
 
-	if err := r.client.SyncLBStateVIP(cfg, ev.FrontendName, ""); err != nil {
+	if err := r.client.SyncLBStateVIP(cfg, feName, ""); err != nil {
 		if errors.Is(err, ErrFrontendNotFound) {
 			// Frontend was removed between the event being emitted and
 			// us handling it; a periodic SyncLBStateAll will clean it up.
 			return
 		}
 		slog.Warn("vpp-reconciler-error",
-			"frontend", ev.FrontendName,
+			"frontend", feName,
 			"backend", ev.BackendName,
 			"from", ev.Transition.From.String(),
 			"to", ev.Transition.To.String(),

internal/vpp/warmup.go

@@ -0,0 +1,408 @@
+// Copyright (c) 2026, Pim van Pelt <pim@ipng.ch>
+
+package vpp
+
+import (
+	"context"
+	"log/slog"
+	"sync"
+	"time"
+
+	"git.ipng.ch/ipng/vpp-maglev/internal/config"
+	"git.ipng.ch/ipng/vpp-maglev/internal/health"
+)
+
+// warmupPollInterval is how often runWarmup re-checks per-VIP backend
+// state during the [minDelay, maxDelay) per-VIP release phase. 250ms
+// is fast enough that a VIP whose last backend just completed rise
+// probes gets released within a quarter-second of settling, and slow
+// enough that the polling cost is negligible compared to probe work
+// the checker is doing on the same core at the same time.
+const warmupPollInterval = 250 * time.Millisecond
+
+// warmupTracker is the per-process gate for the VPP LB sync path
+// during the first StartupMaxDelay seconds of maglevd's life. It
+// exists to keep a maglevd restart dataplane-neutral: without it,
+// the first SyncLBStateAll would fire before any probes had
+// completed, every backend would still be in StateUnknown, and
+// BackendEffectiveWeight would reduce every AS to weight 0 — which
+// on VPP's side means the new-flow table empties and every new
+// connection hits the "no server" drop counter until the rise
+// counters catch up.
+//
+// The tracker expresses three states and the transitions between
+// them:
+//
+//  1. inside [0, minDelay)  — "min-delay window". No sync of any
+//     kind is allowed to touch VPP, neither the periodic SyncLBStateAll
+//     loop nor the per-transition SyncLBStateVIP path from the
+//     reconciler. This is the absolute hands-off window the operator
+//     configures with vpp.lb.startup-min-delay.
+//
+//  2. inside [minDelay, maxDelay), per-VIP gating — "release phase".
+//     Each frontend is released (and one SyncLBStateVIP runs against
+//     it) as soon as every backend it references has reached a
+//     non-Unknown state. Both the warmup driver goroutine (which
+//     polls at warmupPollInterval) and the reconciler event path
+//     (which checks on every received transition) attempt to release
+//     VIPs; tryRelease arbitrates.
+//
+//  3. allDone — warmup is complete. Either every VIP has been
+//     individually released during the release phase, or the
+//     maxDelay watchdog expired and the warmup driver ran a final
+//     SyncLBStateAll for any stragglers. After allDone every gate
+//     is open, the reconciler runs normally on every transition,
+//     and the periodic lbSyncLoop's ticker starts.
+//
+// The clock is process-relative: startAt is set in Client.New()
+// and does not reset across VPP reconnects. If VPP drops at t=8s
+// while the release phase is mid-run and reconnects at t=12s, the
+// warmup driver re-enters the release phase knowing that 12s of
+// the 30s maxDelay have already been consumed. If VPP stays down
+// past maxDelay, the first connect after that jumps straight to
+// the final SyncLBStateAll and marks allDone.
+type warmupTracker struct {
+	startAt  time.Time
+	minDelay time.Duration
+	maxDelay time.Duration
+
+	mu       sync.Mutex
+	released map[string]bool // frontend name → released-for-sync
+	allDone  bool
+	doneCh   chan struct{} // closed when allDone is first set
+}
+
+// newWarmupTracker constructs a tracker with startAt pinned to time.Now().
+// Delay values are not read at construction time — they come from the
+// config via runWarmup's call to getStateSource().Config() — so main.go
+// can construct the Client before the config has been fully propagated.
+func newWarmupTracker() *warmupTracker {
+	return &warmupTracker{
+		startAt:  time.Now(),
+		released: make(map[string]bool),
+		doneCh:   make(chan struct{}),
+	}
+}
+
+// configure latches the min/max delay values onto the tracker. Idempotent
+// if called with the same values; separate from the constructor so the
+// tracker exists before we've parsed the config, and so runWarmup can
+// read a consistent pair of values even if the config is reloaded mid-
+// warmup (per design decision, config reload does not reset the warmup
+// clock, and the delay values latched at first configure() are authoritative
+// for the lifetime of the warmup phase).
+func (w *warmupTracker) configure(minDelay, maxDelay time.Duration) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	// Only latch once. Subsequent calls are no-ops so a config reload
+	// doesn't re-run warmup against new (possibly shorter) delays.
+	if w.minDelay != 0 || w.maxDelay != 0 {
+		return
+	}
+	w.minDelay = minDelay
+	w.maxDelay = maxDelay
+}
+
+// inMinDelay reports whether the absolute hands-off window is still active.
+func (w *warmupTracker) inMinDelay() bool {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.allDone {
+		return false
+	}
+	return time.Since(w.startAt) < w.minDelay
+}
+
+// isReleased reports whether the given frontend may be synced. True if
+// warmup is fully done or this specific frontend has been individually
+// released during the release phase. Fast path for the reconciler event
+// handler: a cheap check before it considers attempting a release.
+func (w *warmupTracker) isReleased(feName string) bool {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	return w.allDone || w.released[feName]
+}
+
+// tryRelease atomically decides whether to release feName. Returns true
+// if the frontend is (now) eligible for sync:
+//
+//   - already released (by a previous caller, including allDone) → true
+//   - inside minDelay window → false
+//   - past minDelay and caller has verified allKnown externally → true,
+//     and the tracker is mutated to remember the release
+//
+// tryRelease does NOT check the allKnown precondition itself — the
+// caller is responsible for evaluating backend states before calling.
+// Separating the checks this way lets two independent release drivers
+// (the warmup poll goroutine and the reconciler event handler) share
+// the same gating state without exposing a mid-check race.
+//
+// Returns true for the "already released" case so callers have a
+// single branch: if tryRelease(fe) is true, proceed to sync.
+func (w *warmupTracker) tryRelease(feName string) bool {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.allDone || w.released[feName] {
+		return true
+	}
+	if time.Since(w.startAt) < w.minDelay {
+		return false
+	}
+	w.released[feName] = true
+	return true
+}
+
+// finishAll marks warmup fully complete. Called once by runWarmup when
+// either every frontend has been released via the per-VIP path or the
+// maxDelay watchdog has expired. Idempotent: repeat calls are no-ops.
+// Closes doneCh on the first call so waiters in lbSyncLoop unblock.
+func (w *warmupTracker) finishAll() {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.allDone {
+		return
+	}
+	w.allDone = true
+	close(w.doneCh)
+}
+
+// doneChan returns a channel that is closed when finishAll is called.
+// Waiters block on this to defer periodic sync work until after the
+// warmup phase has completed (or been skipped entirely).
+func (w *warmupTracker) doneChan() <-chan struct{} {
+	return w.doneCh
+}
+
+// isAllDone is the non-blocking companion to doneChan: true iff
+// finishAll has been called. Used by lbSyncLoop to decide whether
+// to re-enter runWarmup on each VPP reconnect.
+func (w *warmupTracker) isAllDone() bool {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	return w.allDone
+}
+
+// elapsed returns how long the tracker has been running, formatted
+// as a human-readable Go duration string (e.g. "959ms", "5.2s") for
+// use in slog attributes. Returning the string form directly —
+// rather than a time.Duration — is deliberate: slog's default JSON
+// handler renders time.Duration as a raw nanosecond int64 which is
+// unreadable in a log viewer, while a pre-formatted string lands
+// as "5.2s" and matches how the config values are written in YAML.
+func (w *warmupTracker) elapsed() string {
+	return time.Since(w.startAt).Round(time.Millisecond).String()
+}
+
+// allBackendsKnown reports whether every backend referenced by fe is
+// in a non-Unknown state. This is the precondition for releasing a
+// frontend during the per-VIP release phase: desiredFromFrontend can
+// only produce correct weights for a frontend whose backends have
+// all been probed at least once through the health checker's rise
+// counter (unknown → up/down).
+//
+// "Known" here is the literal reading: StateUnknown disqualifies,
+// everything else qualifies. That means a legitimately-down backend
+// counts as known and contributes its weight=0 to the desired set,
+// which is the correct restart behaviour — a backend that was down
+// before the restart stays down across the restart without waiting
+// for it to come back up.
+func allBackendsKnown(fe config.Frontend, src StateSource) bool {
+	for _, pool := range fe.Pools {
+		for bName := range pool.Backends {
+			s, ok := src.BackendState(bName)
+			if !ok || s == health.StateUnknown {
+				return false
+			}
+		}
+	}
+	return true
+}
+
+// runWarmup drives the warmup state machine. Called from lbSyncLoop
+// on first entry (subsequent reconnect entries find allDone == true
+// and skip straight to the periodic ticker).
+//
+// Phases:
+//
+//  1. Latch delay values from the current config.
+//  2. If maxDelay == 0 (warmup disabled): run SyncLBStateAll
+//     immediately, mark allDone, return.
+//  3. Sleep until minDelay has elapsed (absolute hands-off).
+//  4. Poll every warmupPollInterval, releasing any frontend whose
+//     backends are all known. Each release fires a single-VIP sync.
+//     Exit the poll when all frontends are released OR maxDelay
+//     elapses.
+//  5. Run SyncLBStateAll for any stragglers and mark allDone.
+//
+// Exits early if ctx is cancelled at any point.
+func (c *Client) runWarmup(ctx context.Context) {
+	src := c.getStateSource()
+	if src == nil {
+		// No state source ever registered; nothing meaningful to do.
+		// Close the gate so lbSyncLoop doesn't hang.
+		c.warmup.finishAll()
+		return
+	}
+	cfg := src.Config()
+	if cfg == nil {
+		c.warmup.finishAll()
+		return
+	}
+	c.warmup.configure(cfg.VPP.LB.StartupMinDelay, cfg.VPP.LB.StartupMaxDelay)
+
+	w := c.warmup
+
+	// maxDelay == 0 is the "no warmup" escape hatch: sync immediately
+	// and mark the gate open. Operators pick this for tests and dev
+	// setups where a few seconds of startup black-hole on bounce is
+	// acceptable in exchange for not having to wait out the warmup.
+	if w.maxDelay == 0 {
+		slog.Info("vpp-lb-warmup-skipped",
+			"impact", "VPP LB update skipped")
+		if err := c.SyncLBStateAll(cfg); err != nil {
+			slog.Warn("vpp-lb-sync-error", "err", err)
+		}
+		w.finishAll()
+		return
+	}
+
+	slog.Info("vpp-lb-warmup-start",
+		"min-delay", w.minDelay.String(),
+		"max-delay", w.maxDelay.String(),
+		"impact", "Gating all VPP LB updates")
+
+	// Phase 3: wait out the min-delay absolute hands-off window.
+	minDeadline := w.startAt.Add(w.minDelay)
+	if wait := time.Until(minDeadline); wait > 0 {
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(wait):
+		}
+	}
+	slog.Info("vpp-lb-warmup-min-delay-elapsed",
+		"elapsed", w.elapsed(),
+		"impact", "Ungating VPP LB updates for VIPs with known backend state")
+
+	// Phase 4: poll for per-VIP release until maxDelay expires.
+	// happyPath is set to true if we exit the poll loop because
+	// every frontend has been released individually via SyncLBStateVIP.
+	// In that case Phase 5 below skips the SyncLBStateAll entirely —
+	// running it would be a redundant full reconcile over a dataplane
+	// that's already in the desired state, and the log would misreport
+	// the warmup-complete event as a "max-delay-final" stragglers sweep.
+	maxDeadline := w.startAt.Add(w.maxDelay)
+	happyPath := false
+	for time.Now().Before(maxDeadline) {
+		// Re-read the state source every tick: the config may not
+		// have been available at loop entry (e.g. first connect
+		// beat the config load race) but could be present now.
+		src = c.getStateSource()
+		if src != nil {
+			cfg = src.Config()
+		}
+		if cfg != nil {
+			// Release any frontend whose backends have all settled.
+			allReleased := true
+			for feName, fe := range cfg.Frontends {
+				if w.isReleased(feName) {
+					continue
+				}
+				if !allBackendsKnown(fe, src) {
+					allReleased = false
+					continue
+				}
+				if !w.tryRelease(feName) {
+					// Still inside minDelay — shouldn't happen here
+					// because we waited above, but guard anyway.
+					allReleased = false
+					continue
+				}
+				slog.Info("vpp-lb-warmup-release",
+					"frontend", feName,
+					"trigger", "poll",
+					"elapsed", w.elapsed())
+				if err := c.SyncLBStateVIP(cfg, feName, ""); err != nil {
+					slog.Warn("vpp-lb-warmup-release-error",
+						"frontend", feName,
+						"err", err)
+				}
+			}
+			if allReleased {
+				// Everything settled before maxDelay — fast path
+				// out of the poll so we don't sit idle for the
+				// remainder of the watchdog window.
+				happyPath = true
+				break
+			}
+		}
+
+		// Sleep warmupPollInterval or until maxDelay, whichever
+		// is shorter, before trying again.
+		wait := warmupPollInterval
+		if rem := time.Until(maxDeadline); rem < wait {
+			wait = rem
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(wait):
+		}
+	}
+
+	// Phase 5: close out warmup. Two paths, but both emit
+	// vpp-lb-warmup-max-delay-elapsed at the max-delay boundary so
+	// the log timeline (start → min-delay-elapsed → (releases
+	// happen) → max-delay-elapsed) is consistent regardless of
+	// whether the warmup ended early or the watchdog tripped.
+	//
+	//   - happyPath: every frontend was released individually during
+	//     Phase 4 and each one's SyncLBStateVIP already ran. VPP is
+	//     in the desired state; finishAll is called immediately so
+	//     the periodic sync loop can start drift-correction without
+	//     waiting out the rest of max-delay. The warmup driver then
+	//     sleeps until max-delay and emits -max-delay-elapsed as a
+	//     gratuitous timeline marker — the gate is already open,
+	//     but the line completes the warmup picture for an operator
+	//     reading the log and keeps the event sequence symmetric
+	//     with the watchdog path.
+	//
+	//   - watchdog: max-delay elapsed with stragglers remaining. At
+	//     least one frontend never made it through allBackendsKnown,
+	//     so its effective weight computation still treats some
+	//     backends as StateUnknown and will program weight=0 for
+	//     them. Emit -max-delay-elapsed at the boundary, run
+	//     SyncLBStateAll to sweep stragglers, then finishAll.
+	if happyPath {
+		slog.Info("vpp-lb-warmup-complete",
+			"elapsed", w.elapsed(),
+			"impact", "Ungating VPP LB updates, all frontends released")
+		w.finishAll()
+
+		if wait := time.Until(maxDeadline); wait > 0 {
+			select {
+			case <-ctx.Done():
+				return
+			case <-time.After(wait):
+			}
+		}
+		slog.Info("vpp-lb-warmup-max-delay-elapsed",
+			"elapsed", w.elapsed(),
+			"impact", "Ungating all VPP LB updates")
+		return
+	}
+
+	slog.Info("vpp-lb-warmup-max-delay-elapsed",
+		"elapsed", w.elapsed(),
+		"impact", "Ungating all VPP LB updates")
+	src = c.getStateSource()
+	if src != nil {
+		cfg = src.Config()
+	}
+	if cfg != nil {
+		if err := c.SyncLBStateAll(cfg); err != nil {
+			slog.Warn("vpp-lb-sync-error", "err", err)
+		}
+	}
+	w.finishAll()
+}

internal/vpp/warmup_test.go

@@ -0,0 +1,232 @@
+// Copyright (c) 2026, Pim van Pelt <pim@ipng.ch>
+
+package vpp
+
+import (
+	"net"
+	"testing"
+	"time"
+
+	"git.ipng.ch/ipng/vpp-maglev/internal/config"
+	"git.ipng.ch/ipng/vpp-maglev/internal/health"
+)
+
+// TestWarmupTrackerBasic pins the state-machine transitions the tracker
+// owns: inMinDelay, isReleased, tryRelease, finishAll, isAllDone. Time
+// is manipulated by backdating startAt — there is no hidden global
+// clock, so a test can assert behaviour at any simulated point in
+// the [0, maxDelay] window by rewinding startAt before each step.
+func TestWarmupTrackerBasic(t *testing.T) {
+	t.Run("min-delay gates everything", func(t *testing.T) {
+		w := newWarmupTracker()
+		w.configure(5*time.Second, 30*time.Second)
+
+		// At t=0 we are inside min-delay.
+		if !w.inMinDelay() {
+			t.Error("t=0: expected inMinDelay=true")
+		}
+		if w.isReleased("fe1") {
+			t.Error("t=0: expected isReleased(fe1)=false")
+		}
+		if w.tryRelease("fe1") {
+			t.Error("t=0: tryRelease should fail inside min-delay")
+		}
+		if w.isAllDone() {
+			t.Error("t=0: expected isAllDone=false")
+		}
+	})
+
+	t.Run("per-VIP release after min-delay", func(t *testing.T) {
+		w := newWarmupTracker()
+		w.configure(5*time.Second, 30*time.Second)
+		// Simulate t=10s by backdating startAt 10s into the past.
+		w.startAt = time.Now().Add(-10 * time.Second)
+
+		if w.inMinDelay() {
+			t.Error("t=10s: expected inMinDelay=false")
+		}
+		if w.isReleased("fe1") {
+			t.Error("t=10s: fe1 should not be released yet")
+		}
+		if !w.tryRelease("fe1") {
+			t.Error("t=10s: tryRelease(fe1) should succeed")
+		}
+		if !w.isReleased("fe1") {
+			t.Error("after tryRelease: isReleased(fe1) should be true")
+		}
+		// A second call returns true (already-released path).
+		if !w.tryRelease("fe1") {
+			t.Error("second tryRelease(fe1) should return true (already released)")
+		}
+		// Other VIPs are independent.
+		if w.isReleased("fe2") {
+			t.Error("releasing fe1 should not affect fe2")
+		}
+	})
+
+	t.Run("finishAll opens all gates", func(t *testing.T) {
+		w := newWarmupTracker()
+		w.configure(5*time.Second, 30*time.Second)
+
+		// Inside min-delay: fe1 not released.
+		if w.isReleased("fe1") {
+			t.Error("pre-finishAll: fe1 should not be released")
+		}
+		w.finishAll()
+		if !w.isAllDone() {
+			t.Error("post-finishAll: isAllDone should be true")
+		}
+		if !w.isReleased("fe1") {
+			t.Error("post-finishAll: every frontend should be released")
+		}
+		if w.inMinDelay() {
+			t.Error("post-finishAll: inMinDelay should be false")
+		}
+		// Second call is idempotent.
+		w.finishAll()
+	})
+
+	t.Run("doneChan closes on finishAll", func(t *testing.T) {
+		w := newWarmupTracker()
+		w.configure(5*time.Second, 30*time.Second)
+
+		select {
+		case <-w.doneChan():
+			t.Fatal("doneChan should not be readable before finishAll")
+		default:
+		}
+		w.finishAll()
+		select {
+		case <-w.doneChan():
+			// expected
+		case <-time.After(100 * time.Millisecond):
+			t.Fatal("doneChan should be readable after finishAll")
+		}
+	})
+
+	t.Run("configure is idempotent after first call", func(t *testing.T) {
+		w := newWarmupTracker()
+		w.configure(5*time.Second, 30*time.Second)
+		// Reload with shorter delays should be a no-op so a config
+		// reload mid-warmup doesn't move the goalposts.
+		w.configure(1*time.Second, 2*time.Second)
+		if w.minDelay != 5*time.Second {
+			t.Errorf("minDelay got %v, want %v", w.minDelay, 5*time.Second)
+		}
+		if w.maxDelay != 30*time.Second {
+			t.Errorf("maxDelay got %v, want %v", w.maxDelay, 30*time.Second)
+		}
+	})
+}
+
+// staticStateSource is a minimal StateSource for allBackendsKnown tests.
+// It holds a fixed config and a static backend-state map; BackendState
+// returns (state, true) for entries in the map and (StateUnknown, false)
+// for everything else — matching how checker.Checker would report a
+// backend that isn't under its watch.
+type staticStateSource struct {
+	cfg    *config.Config
+	states map[string]health.State
+}
+
+func (s *staticStateSource) Config() *config.Config { return s.cfg }
+func (s *staticStateSource) BackendState(name string) (health.State, bool) {
+	st, ok := s.states[name]
+	return st, ok
+}
+
+// TestAllBackendsKnown pins the per-VIP release precondition. A
+// frontend is eligible for release during the warmup phase iff every
+// backend it references has reached a non-Unknown state. StateDown,
+// StatePaused, and StateDisabled all count as "known" — the property
+// is "has the checker reported at least once", not "is healthy".
+func TestAllBackendsKnown(t *testing.T) {
+	ip := func(s string) net.IP { return net.ParseIP(s) }
+	fe := config.Frontend{
+		Address:  ip("192.0.2.1"),
+		Protocol: "tcp",
+		Port:     80,
+		Pools: []config.Pool{
+			{Name: "primary", Backends: map[string]config.PoolBackend{
+				"b1": {Weight: 100},
+				"b2": {Weight: 100},
+			}},
+			{Name: "fallback", Backends: map[string]config.PoolBackend{
+				"b3": {Weight: 100},
+			}},
+		},
+	}
+
+	cases := []struct {
+		name   string
+		states map[string]health.State
+		want   bool
+	}{
+		{
+			name: "all up → known",
+			states: map[string]health.State{
+				"b1": health.StateUp, "b2": health.StateUp, "b3": health.StateUp,
+			},
+			want: true,
+		},
+		{
+			name: "mixed up/down/disabled → known",
+			// All of up, down, paused, disabled are "non-Unknown" and
+			// therefore counts as known; allBackendsKnown is about
+			// "has the checker reported once", not about health.
+			states: map[string]health.State{
+				"b1": health.StateDown, "b2": health.StateDisabled, "b3": health.StatePaused,
+			},
+			want: true,
+		},
+		{
+			name: "one still unknown → not known",
+			states: map[string]health.State{
+				"b1": health.StateUp, "b2": health.StateUnknown, "b3": health.StateUp,
+			},
+			want: false,
+		},
+		{
+			name:   "backend not reported at all (checker doesn't know it) → not known",
+			states: map[string]health.State{"b1": health.StateUp, "b2": health.StateUp},
+			want:   false,
+		},
+		{
+			name: "fallback pool unknown → not known",
+			states: map[string]health.State{
+				"b1": health.StateUp, "b2": health.StateUp, "b3": health.StateUnknown,
+			},
+			want: false,
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			src := &staticStateSource{cfg: &config.Config{}, states: tc.states}
+			got := allBackendsKnown(fe, src)
+			if got != tc.want {
+				t.Errorf("got %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
+
+// TestWarmupTrackerZeroDelays pins the "warmup disabled" escape hatch:
+// with both delays set to 0, tryRelease succeeds immediately and
+// isReleased returns true for every frontend without needing any
+// state transitions. This is the configuration an operator picks
+// when they'd rather take the brief startup black-hole than wait
+// out the warmup — typically for tests and dev setups.
+func TestWarmupTrackerZeroDelays(t *testing.T) {
+	w := newWarmupTracker()
+	w.configure(0, 0)
+	if w.inMinDelay() {
+		t.Error("min=0: expected inMinDelay=false immediately")
+	}
+	if !w.tryRelease("fe1") {
+		t.Error("min=0: tryRelease should succeed immediately")
+	}
+	if !w.isReleased("fe1") {
+		t.Error("min=0: isReleased(fe1) should be true after tryRelease")
+	}
+}