eval

internal/alerts/laya.go

@@ -0,0 +1,180 @@
+package alerts
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// Laya asks the laya model (POST {URL}/predict) what it would do about an
+// alert. It runs in shadow mode: the verdict is recorded on the alert and
+// shown in notifications, but nothing is ever acted on, muted or dropped
+// because of it. A failed call leaves the alert exactly as it was.
+type Laya struct {
+	URL    string // base URL, e.g. https://laya.k6n.net
+	Client *http.Client
+}
+
+// LayaActions are the remediations laya may choose between. They are
+// suggestions only; none is wired to an automated action yet.
+var LayaActions = map[string]string{
+	"ignore":             "transient blip or noise, nobody needs to look at it",
+	"restart_deployment": "the service is stuck, leaking or crashed; restarting its pods is likely to clear it",
+	"check_capacity":     "requests time out or resources are exhausted; check load, connection pools, CPU/memory limits and scaling",
+	"fix_dependency":     "a database, upstream service or network path this service depends on is failing",
+	"investigate_code":   "a bug or bad release in this service; needs a code fix or a rollback",
+}
+
+var layaUrgency = []string{"cosmetic", "degraded performance", "service outage"}
+
+// LayaVerdict is laya's answer, reduced to what a person needs to judge it.
+type LayaVerdict struct {
+	Model            string             `json:"model,omitempty"`
+	Action           string             `json:"action"`
+	ActionConfidence float64            `json:"actionConfidence"` // probability of the chosen action
+	ActionProbs      map[string]float64 `json:"actionProbs,omitempty"`
+	Human            float64            `json:"human"` // probability a person must intervene
+	Urgency          string             `json:"urgency"`
+	UrgencyScore     float64            `json:"urgencyScore"` // expected value on 0 (cosmetic) .. 2 (outage)
+	UrgencyConf      float64            `json:"urgencyConfidence"`
+}
+
+func layaRequest(state string) map[string]any {
+	return map[string]any{
+		"state": state,
+		"questions": map[string]any{
+			"remediation_action": map[string]any{
+				"type":         "choice",
+				"instructions": "Which operational response fits this production error alert best?",
+				"criteria":     LayaActions,
+			},
+			"requires_human_attention": map[string]any{
+				"type":         "noul",
+				"instructions": "Does this alert need a human engineer to intervene, or will it resolve or matter little on its own?",
+				"criteria": map[string]string{
+					"true":  "errors are new or rising well above normal and users or requests are failing",
+					"false": "a transient blip, client noise or bot traffic",
+				},
+			},
+			"urgency": map[string]any{
+				"type":         "score",
+				"instructions": "Rate the operational severity of this alert.",
+				"criteria":     layaUrgency,
+			},
+		},
+	}
+}
+
+type layaAnswer struct {
+	Choice           string             `json:"choice"`
+	Probabilities    map[string]float64 `json:"probabilities"`
+	AnswerConfidence float64            `json:"answer_confidence"`
+	Noul             float64            `json:"noul"`
+	Score            float64            `json:"score"`
+	Legend           map[string]string  `json:"legend"`
+}
+
+type layaResponse struct {
+	Model   string                `json:"model"`
+	Answers map[string]layaAnswer `json:"answers"`
+}
+
+// Classify asks laya about the alert. The state is the alert's plain-text
+// message, so the model sees exactly the facts a person would, nothing more.
+func (l *Laya) Classify(ctx context.Context, a Alert) (*LayaVerdict, error) {
+	a.Laya = nil
+	return l.classifyState(ctx, a.Text(""))
+}
+
+func (l *Laya) classifyState(ctx context.Context, state string) (*LayaVerdict, error) {
+	raw, err := json.Marshal(layaRequest(state))
+	if err != nil {
+		return nil, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, strings.TrimRight(l.URL, "/")+"/predict", bytes.NewReader(raw))
+	if err != nil {
+		return nil, errors.New("laya: invalid URL")
+	}
+	req.Header.Set("Content-Type", "application/json")
+	client := l.Client
+	if client == nil {
+		client = &http.Client{Timeout: 10 * time.Second}
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("laya: %w", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode/100 != 2 {
+		detail, _ := io.ReadAll(io.LimitReader(resp.Body, 300))
+		return nil, fmt.Errorf("laya: HTTP %d: %s", resp.StatusCode, strings.TrimSpace(string(detail)))
+	}
+	var lr layaResponse
+	if err := json.NewDecoder(resp.Body).Decode(&lr); err != nil {
+		return nil, fmt.Errorf("laya: invalid response: %w", err)
+	}
+	return lr.verdict()
+}
+
+func (lr layaResponse) verdict() (*LayaVerdict, error) {
+	act, ok1 := lr.Answers["remediation_action"]
+	hum, ok2 := lr.Answers["requires_human_attention"]
+	urg, ok3 := lr.Answers["urgency"]
+	if !ok1 || !ok2 || !ok3 || act.Choice == "" {
+		return nil, errors.New("laya: response is missing answers")
+	}
+	v := &LayaVerdict{Model: lr.Model, Action: act.Choice, ActionConfidence: act.Probabilities[act.Choice], ActionProbs: act.Probabilities,
+		Human: hum.Noul, UrgencyScore: urg.Score}
+	// The urgency label is the most probable bucket, not the rounded score:
+	// a 50/50 cosmetic/outage split must not read as "degraded".
+	best := -1.0
+	for k, p := range urg.Probabilities {
+		if p > best || (p == best && k < v.Urgency) {
+			best, v.Urgency = p, k
+		}
+	}
+	v.UrgencyConf = best
+	if label, ok := urg.Legend[v.Urgency]; ok {
+		v.Urgency = label
+	} else if i, err := strconv.Atoi(v.Urgency); err == nil && i >= 0 && i < len(layaUrgency) {
+		v.Urgency = layaUrgency[i]
+	}
+	return v, nil
+}
+
+// Summary is a one-line rendering, e.g.
+// "check_capacity 62% · human 71% · degraded performance 75%".
+func (v LayaVerdict) Summary() string {
+	return fmt.Sprintf("%s %s · human %s · %s %s", v.Action, pct(v.ActionConfidence), pct(v.Human), v.Urgency, pct(v.UrgencyConf))
+}
+
+// RunnersUp lists the other actions by probability, for judging how close
+// the call was.
+func (v LayaVerdict) RunnersUp() string {
+	type kv struct {
+		k string
+		p float64
+	}
+	var rest []kv
+	for k, p := range v.ActionProbs {
+		if k != v.Action {
+			rest = append(rest, kv{k, p})
+		}
+	}
+	sort.Slice(rest, func(i, j int) bool { return rest[i].p > rest[j].p })
+	parts := make([]string, 0, len(rest))
+	for _, r := range rest {
+		parts = append(parts, r.k+" "+pct(r.p))
+	}
+	return strings.Join(parts, ", ")
+}
+
+func pct(p float64) string { return strconv.Itoa(int(p*100+0.5)) + "%" }