Prompt Observability Slo

# Prompt — Observability and SLO design

> [🇷🇺 Русский](../ru/prompt-observability-slo.md) · 🇬🇧 **English**

[← Index](./README.md) · [Standards](./standards.md) · [Template](./template.md)

tags: #prompt #observability #slo #sre #service #single-shot
**Status:** `🧪 Experimental`
**Owner:** Pandamy619
**Last edit:** 2026-05-20
**Version:** v1
**Source idea:** [Google SRE Workbook — Implementing SLOs](https://sre.google/workbook/implementing-slos/), [Liz Fong-Jones — Multi-window multi-burn-rate alerting](https://sre.google/workbook/alerting-on-slos/)
**Delta from original:** requires SLI linkage to user journey (not internal metric), multi-window multi-burn-rate alerting format, and an explicit section on logging/tracing correlation fields.

---

## When to use

- When a service is moving to production and you need to decide what and how to observe
- When the team is drowning in alerts and wants to move to SLO-based alerting
- When you need to separate "I see the metric" from "I understand whether users are suffering"

## What to substitute

- `{{service_name}}` — string, e.g. `payments-api`
- `{{critical_user_journeys}}` — list, e.g. `create payment; check status; refund`
- `{{current_metrics}}` — string, e.g. `Prometheus on /metrics; have http_requests_total and http_request_duration_seconds`
- `{{logging_stack}}` — enum: `loki | elastic | datadog_logs | cloudwatch | other`
- `{{tracing_stack}}` — enum: `opentelemetry | jaeger | tempo | datadog_apm | none`
- `{{error_budget_policy}}` — string, e.g. `on budget exhaustion — freeze features until recovery`
- `{{team_maturity}}` — enum: `none | basic_metrics | mature_metrics_no_slo | slo_in_progress`

## Definition of done

Universal minimum — see [Definition of done](./standards.md#2-definition-of-done).

Specific criteria:

- For each critical user journey, at least 2 SLIs are defined (latency + availability)
- SLO is tied to user-visible behavior, not to internal metrics (CPU, memory aren't SLOs)
- Alerting is based on burn rate (multi-window multi-burn-rate), not instantaneous values
- Specific log fields are listed for correlation (`trace_id`, `request_id`, `user_id`)
- Dashboards are split into 3 levels: overview / per-journey / drill-down

## Prompt