Add selectable canary-first fleet updates

internal/server/http/fleet_update.go

@@ -32,6 +32,13 @@ import (
 type fleetUpdateStartReq struct {
 	TargetVersion string   `json:"target_version,omitempty"`
 	HostIDs       []string `json:"host_ids,omitempty"`
+	CanaryFirst   bool     `json:"canary_first,omitempty"`
+}
+
+type fleetUpdateCandidate struct {
+	Host     store.Host
+	Eligible bool
+	Reason   string
 }
 
 // fleetUpdateHostView is one row in the JSON response for GET
@@ -56,6 +63,7 @@ type fleetUpdateView struct {
 	CurrentHostID   string                `json:"current_host_id,omitempty"`
 	HaltedReason    string                `json:"halted_reason,omitempty"`
 	CompletedAt     *string               `json:"completed_at,omitempty"`
+	CanaryFirst     bool                  `json:"canary_first"`
 	Hosts           []fleetUpdateHostView `json:"hosts"`
 }
 
@@ -65,6 +73,7 @@ type fleetUpdateView struct {
 type fleetUpdatePage struct {
 	// Idle-state fields.
 	OutOfDateHosts []store.Host // online hosts whose version != target
+	Candidates     []fleetUpdateCandidate
 	TargetVersion  string
 
 	// Active-state fields. Nil when no fleet update has ever run.
@@ -100,6 +109,11 @@ func (s *Server) handleAPIFleetUpdateStart(w stdhttp.ResponseWriter, r *stdhttp.
 	if target == "" {
 		target = version.Version
 	}
+	if target != version.Version {
+		writeJSONError(w, stdhttp.StatusUnprocessableEntity, "unsupported_target_version",
+			"fleet updates can only target the running server version")
+		return
+	}
 	hostIDs := body.HostIDs
 	if len(hostIDs) == 0 {
 		derived, err := s.deriveOutOfDateOnlineHostIDs(r.Context(), target)
@@ -108,6 +122,19 @@ func (s *Server) handleAPIFleetUpdateStart(w stdhttp.ResponseWriter, r *stdhttp.
 			return
 		}
 		hostIDs = derived
+	} else {
+		validated, reasons, err := s.validateFleetUpdateHostIDs(r.Context(), target, hostIDs)
+		if err != nil {
+			writeJSONError(w, stdhttp.StatusInternalServerError, "internal", err.Error())
+			return
+		}
+		if len(reasons) > 0 {
+			writeJSON(w, stdhttp.StatusUnprocessableEntity, map[string]any{
+				"code": "ineligible_hosts", "message": "one or more selected hosts are ineligible", "reasons": reasons,
+			})
+			return
+		}
+		hostIDs = validated
 	}
 	if len(hostIDs) == 0 {
 		writeJSONError(w, stdhttp.StatusConflict, "no_hosts_eligible",
@@ -115,7 +142,20 @@ func (s *Server) handleAPIFleetUpdateStart(w stdhttp.ResponseWriter, r *stdhttp.
 		return
 	}
 
-	fuID, err := s.deps.FleetWorker.Start(r.Context(), user.ID, target, hostIDs)
+	var fuID string
+	var err error
+	if body.CanaryFirst && len(hostIDs) > 1 {
+		worker, supported := s.deps.FleetWorker.(interface {
+			StartCanary(context.Context, string, string, []string) (string, error)
+		})
+		if !supported {
+			writeJSONError(w, stdhttp.StatusServiceUnavailable, "canary_unavailable", "")
+			return
+		}
+		fuID, err = worker.StartCanary(r.Context(), user.ID, target, hostIDs)
+	} else {
+		fuID, err = s.deps.FleetWorker.Start(r.Context(), user.ID, target, hostIDs)
+	}
 	if err != nil {
 		if errors.Is(err, store.ErrFleetUpdateRunning) {
 			writeJSONError(w, stdhttp.StatusConflict, "fleet_update_in_progress", err.Error())
@@ -129,6 +169,7 @@ func (s *Server) handleAPIFleetUpdateStart(w stdhttp.ResponseWriter, r *stdhttp.
 		"fleet_update_id": fuID,
 		"target_version":  target,
 		"host_count":      len(hostIDs),
+		"canary_first":    body.CanaryFirst,
 	})
 	_ = s.deps.Store.AppendAudit(r.Context(), store.AuditEntry{
 		ID: ulid.Make().String(), UserID: &user.ID, Actor: "user",
@@ -209,6 +250,7 @@ func (s *Server) handleAPIFleetUpdateGet(w stdhttp.ResponseWriter, r *stdhttp.Re
 		Status:          fu.Status,
 		CurrentHostID:   fu.CurrentHostID,
 		HaltedReason:    fu.HaltedReason,
+		CanaryFirst:     fu.PauseAfterFirst,
 		Hosts:           make([]fleetUpdateHostView, 0, len(hosts)),
 	}
 	if fu.CompletedAt != nil {
@@ -286,6 +328,27 @@ func (s *Server) buildFleetUpdatePage(r *stdhttp.Request) (fleetUpdatePage, erro
 	}
 	for _, h := range hosts {
 		page.HostNames[h.ID] = h.Name
+		candidate := fleetUpdateCandidate{Host: h}
+		switch {
+		case h.AgentVersion == "":
+			candidate.Reason = "version unknown"
+		case h.AgentVersion == page.TargetVersion:
+			candidate.Reason = "already current"
+		case s.deps.Hub == nil || !s.deps.Hub.Connected(h.ID):
+			candidate.Reason = "offline"
+		default:
+			updating, uerr := s.deps.Store.RunningUpdateJobForHost(r.Context(), h.ID)
+			if uerr != nil {
+				return page, uerr
+			}
+			if updating != "" {
+				candidate.Reason = "update already running"
+			} else {
+				candidate.Eligible = true
+				page.OutOfDateHosts = append(page.OutOfDateHosts, h)
+			}
+		}
+		page.Candidates = append(page.Candidates, candidate)
 	}
 
 	active, err := s.deps.Store.ActiveFleetUpdate(r.Context())
@@ -328,20 +391,54 @@ func (s *Server) buildFleetUpdatePage(r *stdhttp.Request) (fleetUpdatePage, erro
 		}
 	}
 
-	// Idle list (or "still out of date" reference even when an active
-	// roll is running — cheap to compute, harmless to attach).
-	for _, h := range hosts {
-		if h.Status != "online" {
-			continue
-		}
-		if h.AgentVersion == "" || h.AgentVersion == page.TargetVersion {
-			continue
-		}
-		page.OutOfDateHosts = append(page.OutOfDateHosts, h)
-	}
 	return page, nil
 }
 
+// validateFleetUpdateHostIDs deduplicates an explicit selection while
+// preserving review order and rejects the entire request if membership has
+// changed or any selected host is not dispatchable.
+func (s *Server) validateFleetUpdateHostIDs(ctx context.Context, target string, requested []string) ([]string, map[string]string, error) {
+	hosts, err := s.deps.Store.ListHosts(ctx)
+	if err != nil {
+		return nil, nil, err
+	}
+	byID := make(map[string]store.Host, len(hosts))
+	for _, h := range hosts {
+		byID[h.ID] = h
+	}
+	seen := map[string]bool{}
+	validated := make([]string, 0, len(requested))
+	reasons := map[string]string{}
+	for _, id := range requested {
+		if seen[id] {
+			continue
+		}
+		seen[id] = true
+		h, ok := byID[id]
+		switch {
+		case !ok:
+			reasons[id] = "host not found"
+		case h.AgentVersion == "":
+			reasons[id] = "agent version unknown"
+		case h.AgentVersion == target:
+			reasons[id] = "already at target version"
+		case s.deps.Hub == nil || !s.deps.Hub.Connected(id):
+			reasons[id] = "host offline"
+		default:
+			jobID, jerr := s.deps.Store.RunningUpdateJobForHost(ctx, id)
+			if jerr != nil {
+				return nil, nil, jerr
+			}
+			if jobID != "" {
+				reasons[id] = "update already in progress"
+			} else {
+				validated = append(validated, id)
+			}
+		}
+	}
+	return validated, reasons, nil
+}
+
 // deriveOutOfDateOnlineHostIDs returns the list of host IDs that
 // (a) are online (Hub.Connected) and (b) have an agent_version that's
 // non-empty AND != target. Used by the start endpoint when the caller