This commit is contained in:
2026-11-16 16:08:15 +03:00
parent 0497b0c217
commit 442595b37c
29 changed files with 2930 additions and 63 deletions
+3 -6
View File
@@ -23,9 +23,6 @@ MINIO_ROOT_PASSWORD=replace-with-admin-password
S3_ACCESS_KEY_ID=messenger-api
S3_SECRET_ACCESS_KEY=replace-with-api-password
# Reserved for lab6 observability setup.
OTEL_SERVICE_NAME=messenger-api
# Worker must set a distinct service.name.
OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318
OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
OTEL_RESOURCE_ATTRIBUTES=deployment.environment.name=lab
# Compose sets service.name separately for API, dispatcher and worker.
OTEL_SDK_DISABLED=false
GRAFANA_ADMIN_PASSWORD=replace-with-grafana-password
+28 -14
View File
@@ -1,18 +1,32 @@
## Лабораторная 6
Автор / группа: TODO
Целевая оценка (3/4/5): TODO
Трек 5, если заявлен: TODO
Автор / группа: Ромашко Владимир, БВТ2301.
Что реализовано и что перенесено: TODO
Команды запуска из чистого clone: TODO
Результаты проверок и ссылка на REPORT.md: TODO
Ограничения / известные ошибки: TODO
Источники и использование ИИ: TODO
Целевая оценка (3/4/5): **4**.
- [ ] Весь обязательный минимум текущей и предыдущих работ сохранён с объявленными переходами.
- [ ] Контракт и публичные тесты не ослаблены.
- [ ] make check и make test выполнены; приложен реальный вывод.
- [ ] Сценарии отказов/безопасности своего уровня воспроизведены.
- [ ] Все секреты и приватные ключи находятся вне Git; отчёт не содержит tokens.
- [ ] README задания сохранён, REPORT.md содержит инструкцию именно моего решения.
Что реализовано и что перенесено: приложение lab5 с API-репликами, БД, S3, брокером и worker сохранено. Добавлены OTLP metrics/logs/traces, Collector, Prometheus/Loki/Tempo, Grafana provisioning, сквозной trace context с retry, SLO и два алерта.
Команды запуска из чистого clone:
```sh
cp .env.example .env
# Задать локальные пароли в .env.
docker compose up -d --build
curl http://127.0.0.1:8080/health/ready # дождаться 200
python -X utf8 scripts/check_contract.py
python -X utf8 tests/smoke.py --base-url http://127.0.0.1:8080
python -X utf8 student-tests/observability.py --scenario all
```
Результаты проверок: check, публичный smoke и проверка сигналов — PASS; сценарии worker/БД/S3/Tempo пройдены, алерты firing → resolved. Нагрузка: 2389 запросов за 30s, без ошибок. [REPORT.md](../REPORT.md), [вывод](../evidence/verification.txt), [сценарии](../evidence/observability.json), [нагрузка](../evidence/load.json).
Ограничения / известные ошибки: один узел, без HA/TLS и внешней доставки алертов; при длительном отказе backend допустима потеря телеметрии. Блокирующих ошибок в проверенных сценариях не выявлено.
Источники и использование ИИ: контракт лабораторной, документация OpenTelemetry и Loki; Codex — реализация, тестирование и оформление. Ссылки в REPORT.md.
- [x] Весь обязательный минимум текущей и предыдущих работ сохранён с объявленными переходами.
- [x] Контракт и публичные тесты не ослаблены.
- [x] make check и make test выполнены эквивалентными Python-командами; приложен реальный вывод.
- [x] Сценарии отказов/безопасности своего уровня воспроизведены.
- [x] Все секреты и приватные ключи находятся вне Git; отчёт не содержит tokens.
- [x] README задания сохранён, REPORT.md содержит инструкцию именно моего решения.
+49 -14
View File
@@ -1,37 +1,72 @@
# Отчёт — лабораторная 6
- Автор, группа: TODO
- Целевая оценка; выбранный трек 5, если нужен: TODO
- Commit/PR: TODO после публикации
- Использованные источники, библиотеки и помощь ИИ: TODO
- Автор, группа: Ромашко Владимир, БВТ2301.
- Целевая оценка: **4**.
- Источники: [контракт](contracts/observability.md), [OpenTelemetry](https://opentelemetry.io/docs/collector/configuration/), [Loki OTLP](https://grafana.com/docs/loki/latest/send-data/otel/). Библиотеки: FastAPI, SQLAlchemy, Redis, boto3, Pillow, OpenTelemetry SDK, psutil. Codex использовался для реализации, проверок и оформления.
## Запуск из чистого clone
TODO: версии runtime/образов, prerequisites, команды создания локальных secrets/.env, build, миграций, запуска, ожидания готовности, тестов и остановки. Указать необходимые действия оператора. Не вставлять секреты. Отдельно указать безопасную остановку и команду намеренного удаления учебных данных.
Нужны Docker с Compose и Python 3.9+ на хосте; свободные порты 8080, 3000, 9090, 3100, 3200. В контейнерах Python 3.12.8; PostgreSQL 17.4, Redis 7.4.2, Nginx 1.28.0; MinIO закреплён digest. Телеметрия: SDK 1.38.0 / semantic conventions 0.59b0, Collector 0.136.0, Prometheus 3.6.0, Loki 3.5.7, Tempo 2.9.0, Grafana 12.2.0.
```sh
cp .env.example .env
# Задать свои POSTGRES_PASSWORD, MINIO_ROOT_PASSWORD,
# S3_SECRET_ACCESS_KEY и GRAFANA_ADMIN_PASSWORD.
docker compose config --quiet
docker compose up -d --build
curl http://127.0.0.1:8080/health/ready # дождаться 200 {"status":"ok"}
python -X utf8 scripts/check_contract.py
python -X utf8 tests/smoke.py --base-url http://127.0.0.1:8080
python -X utf8 student-tests/observability.py --scenario all
python -X utf8 tests/load.py --base-url http://127.0.0.1:8080 --duration 30 --concurrency 4 --output evidence/load.json
docker compose down # безопасная остановка, volumes остаются
```
Миграции (включая 008 с trace context) и создание S3 bucket выполняются автоматически. Grafana: http://localhost:3000, `admin`, пароль из `.env`; dashboard **Lab 6 / Messenger / Lab 6** импортируется автоматически. Python-команды check/smoke эквивалентны `make check` / `make test`; на проверявшемся Windows-хосте make отсутствует. Для **намеренного удаления всех учебных данных**: `docker compose down -v` (при проверках не выполнялось).
## Архитектура и решения
TODO: схема процессов/хранилищ, источник истины, транзакционные границы, ограничения. Что перенесено из lab5, что изменено в этой работе. Соответствие переменных из .env.example реальной конфигурации.
Из lab5 сохранены Nginx → две API-реплики, PostgreSQL, Redis для сессий/кеша, S3, durable Redis Streams, outbox/dispatcher и worker. PostgreSQL — источник истины для заданий; attachment и outbox фиксируются одной транзакцией после записи staging, ack выполняется после устойчивого результата.
Добавлены API/worker/dispatcher → OTLP → Collector → Prometheus (метрики), Loki (логи), Tempo (трассы) → Grafana. История хранится в volumes, retention 72 часа. W3C context сохраняется в outbox и переносится через broker в worker и image process; retry остаётся в той же трассе. Экспортируются шаблоны маршрутов и технические поля, без SQL-параметров, токенов и содержимого сообщений.
Метрики: `http_requests_total` и `jobs_processed_total` — счётчики; `http_duration_seconds`, `job_duration_seconds`, `attachment_time_to_ready_seconds` — histogram в секундах; `queue_depth`, `db_pool_used/waiting`, `worker_busy/capacity` — количества; `queue_oldest_age_seconds` — секунды; `container_memory_bytes` / `container_cpu_seconds_total` — память/CPU. Точные имена, labels и запросы: [описание сигналов](infra/otel/README.md#сигналы), [dashboard JSON](infra/grafana/dashboards/messenger.json).
Переменные `.env.example` подставляются в Compose: пароли — соответствующим сервисам, `MAX_*`, `JOB_*`, `WORKER_CONCURRENCY` — приложению, `APP_PORT` — Nginx. `OTEL_SDK_DISABLED=false` включает телеметрию; service.name и внутренние адреса заданы отдельно по ролям.
SLI = 1 − 5xx / все запросы без health за **30 минут**, SLO ≥ **99%**. `ApiHighErrorRate`: >5% ошибок за 2m, минимум 20 запросов, `for: 20s`. `QueueLagHigh`: возраст queued >30s при свежем sample, `for: 15s`. Правила и [инструкции диагностики](infra/otel/README.md#queue-lag) сохранены в репозитории.
## Доказательства критериев
| Критерий README | Команда/сценарий | Наблюдаемый результат | Файл доказательства |
| --- | --- | --- | --- |
| Публичный smoke | make test | TODO | TODO |
| Остальные пункты уровня 3 | TODO | TODO | TODO |
| Все пункты уровня 4, если заявлен | TODO | TODO | TODO |
| Выбранный трек 5, если заявлен | TODO | TODO | TODO |
| Структура и публичный smoke | Команды check/smoke выше | PASS при включённой телеметрии | [verification.txt](evidence/verification.txt) |
| Минимум lab5 и общий стенд | `docker compose up -d --build`; smoke | API, upload 202, ready, оригинал и thumbnail работают; хранилища имеют volumes | [compose.yaml](compose.yaml), verification.txt |
| OTLP: метрики, логи, трассы API/worker | `observability.py --scenario smoke` | Все обязательные метрики найдены; связанные логи трёх сервисов | [observability.json: smoke](evidence/observability.json) |
| Импортируемый dashboard и безопасные атрибуты | Provisioning; сценарий smoke | 18 панелей, 3 datasource; проверены labels и отсутствие приватных входных данных | [provisioning](infra/grafana/provisioning), observability.json |
| Корреляция HTTP/job → log/trace | Сценарий smoke | Найдены request_id, trace_id и parent-связи до S3/DB | observability.json: smoke |
| W3C propagation, включая retry | `observability.py --scenario retry` | Попытки 1 и 2 в одной трассе, tracestate сохранён | observability.json: retry |
| p50/p95/p99, очередь, DB pool, worker, CPU/RAM, time-to-ready | Сценарий smoke; dashboard | Метрики существуют; percentile вычисляется по сумме histogram buckets | dashboard JSON, observability.json |
| SLI/SLO и два проверенных алерта | `--scenario queue`, `--scenario database` | Оба алерта firing → resolved после восстановления | [alerts.yaml](infra/otel/alerts.yaml), observability.json |
| Ограниченная буферизация и отказ telemetry backend | `--scenario backend` | API продолжает работу; очередь ≤32 batches, RSS ≤199.03 MiB | [collector.yaml](infra/otel/collector.yaml), observability.json |
TODO: добавить отдельную строку для каждого заявленного критерия. PASS без команды и наблюдения не является доказательством. Проверка инфраструктурного hook требует подтверждения, какие реальные процессы/хранилища перезапущены.
Здесь `observability.py` — `python -X utf8 student-tests/observability.py`. Сценарии останавливают реальные контейнеры текущего Compose-проекта и восстанавливают их в `finally`; volumes не удаляют.
## Эксперименты и отказы
TODO: гипотеза, hardware/лимиты, версии, объём данных, профиль нагрузки, concurrency/arrival rate, длительность/прогрев, повторы, успешные RPS, ошибки, p50/p95/p99, CPU/RAM. Сохранить raw summary без tokens и объяснить ограничения измерения.
Проверки 16.11.2026. Гипотеза: телеметрия показывает причину отказа, а её backend не блокирует бизнес-операции. Docker Desktop: 8 CPU, 7.76 GiB RAM; лимиты API — по 1 CPU/512 MiB, worker — 2 CPU/2 GiB, Collector — 0.5 CPU/256 MiB.
Нагрузка: один прогон closed-loop POST сообщений, 4 клиента, 30 секунд, без отдельного прогрева и автоматических retry. **2389 успешных запросов, 0 ошибок, 79.219 RPS; p50/p95/p99 = 40.784/99.520/129.314 ms**. Исходная сводка: [load.json](evidence/load.json). Это локальная проверка, не исследование предельной производительности; CPU во время прогона отдельно не фиксировался, RSS Collector измерен в сценарии отказа ниже.
| Воздействие | Что ожидаем | Что наблюдали | Время восстановления / потеря данных |
| --- | --- | --- | --- |
| TODO | TODO | TODO | TODO |
| Остановить worker | Рост возраста очереди и QueueLagHigh | Depth=1, возраст 55s; producer есть, consumer появляется после запуска; затем ready, depth=0 | firing → resolved: 20.3s; принятое вложение сохранено |
| Остановить PostgreSQL | 503 и ApiHighErrorRate | Error ratio 42.6%; в trace `db.acquire`/`OperationalError` и связанный лог — отказ БД, а не медленный handler | После запуска запросы успешны; firing → resolved: 109.1s с учётом окна 2m |
| Временно остановить S3 | Retry без потери задания | Первая попытка failed, вторая ready, общий trace | Задание успешно завершено |
| Остановить Tempo на 60s | Приложение работает, буферы ограничены | 165 записей и обработка вложения успешны; очередь ≤32, RSS 198.52–199.02 MiB; после запуска новая трасса найдена | Потеря 808 принятых Collector spans, бизнес-операции успешны |
Порядок диагностики: dashboard → лог по request_id/trace_id → ошибочный или отсутствующий span зависимости → восстановление сервиса → проверка ready/успешных запросов и resolved. Для Tempo потеря рассчитана как Δaccepted − Δsent после опустошения очереди; повторы отправки не считаются отдельными потерянными spans. Подробные timestamps и ID — в [observability.json](evidence/observability.json).
## Ограничения и следующие шаги
TODO: какие отказы выдерживаются и какие нет; что не реализовано; где учебное упрощение; какое улучшение подтверждено, а какое пока гипотеза.
Один локальный узел, без HA и TLS; алерты только локальные. SDK ограничивает очереди 512 записями, Collector использует memory limiter/batch и retry до 15s: длительный отказ может терять телеметрию. Потери до приёма Collector указанной дельтой не измеряются. История удаляется по retention; метрики не заменяют журнал транзакций. При недоступной БД API возвращает 503, после восстановления работает снова. Треки оценки 5 не реализовывались.
+86 -1
View File
@@ -1,9 +1,10 @@
x-api: &api
image: mtusi2026lab6-api:local
pull_policy: never
dns_opt:
- timeout:1
- attempts:1
environment:
environment: &app-env
APP_HOST: 0.0.0.0
APP_PORT: 8080
POSTGRES_HOST: postgres
@@ -29,6 +30,9 @@ x-api: &api
JOB_TIMEOUT_SECONDS: ${JOB_TIMEOUT_SECONDS:-60}
JOB_MEMORY_MB: ${JOB_MEMORY_MB:-768}
STAGING_RETENTION_SECONDS: ${STAGING_RETENTION_SECONDS:-86400}
OTEL_SERVICE_NAME: messenger-api
OTEL_EXPORTER_OTLP_ENDPOINT: http://otel-collector:4318
OTEL_SDK_DISABLED: ${OTEL_SDK_DISABLED:-false}
networks: [backend]
cpus: 1.0
mem_limit: 512m
@@ -51,6 +55,9 @@ services:
dispatcher:
<<: *api
environment:
<<: *app-env
OTEL_SERVICE_NAME: messenger-dispatcher
command: [python, -m, src.jobs, dispatch]
healthcheck:
disable: true
@@ -64,6 +71,9 @@ services:
worker:
<<: *api
environment:
<<: *app-env
OTEL_SERVICE_NAME: messenger-worker
command: [python, -m, src.worker]
init: true
healthcheck:
@@ -101,6 +111,9 @@ services:
migrate:
<<: *api
environment:
<<: *app-env
OTEL_SDK_DISABLED: 'true'
command: [python, -m, src.database]
healthcheck:
disable: true
@@ -150,6 +163,9 @@ services:
cleanup:
<<: *api
environment:
<<: *app-env
OTEL_SDK_DISABLED: 'true'
command: [python, -m, src.cleanup]
healthcheck:
disable: true
@@ -217,6 +233,71 @@ services:
cpus: 0.5
mem_limit: 256m
otel-collector:
image: otel/opentelemetry-collector-contrib:0.136.0
command: [--config=/etc/otel/collector.yaml]
volumes:
- ./infra/otel/collector.yaml:/etc/otel/collector.yaml:ro
networks: [backend]
mem_limit: 256m
cpus: 0.5
restart: unless-stopped
prometheus:
image: prom/prometheus:v3.6.0
command: [--config.file=/etc/prometheus/prometheus.yaml, --storage.tsdb.retention.time=3d]
volumes:
- ./infra/otel/prometheus.yaml:/etc/prometheus/prometheus.yaml:ro
- ./infra/otel/alerts.yaml:/etc/prometheus/alerts.yaml:ro
- prometheus_data:/prometheus
ports: ['127.0.0.1:9090:9090']
networks: [edge, backend]
mem_limit: 512m
cpus: 0.5
restart: unless-stopped
loki:
image: grafana/loki:3.5.7
command: [-config.file=/etc/loki/local.yaml]
volumes:
- ./infra/otel/loki.yaml:/etc/loki/local.yaml:ro
- loki_data:/loki
ports: ['127.0.0.1:3100:3100']
networks: [edge, backend]
mem_limit: 384m
cpus: 0.5
restart: unless-stopped
tempo:
image: grafana/tempo:2.9.0
command: [-config.file=/etc/tempo.yaml]
volumes:
- ./infra/otel/tempo.yaml:/etc/tempo.yaml:ro
- tempo_data:/var/tempo
ports: ['127.0.0.1:3200:3200']
networks: [edge, backend]
mem_limit: 384m
cpus: 0.5
restart: unless-stopped
grafana:
image: grafana/grafana:12.2.0
environment:
GF_SECURITY_ADMIN_USER: admin
GF_SECURITY_ADMIN_PASSWORD: ${GRAFANA_ADMIN_PASSWORD:?Set GRAFANA_ADMIN_PASSWORD in .env}
GF_USERS_ALLOW_SIGN_UP: 'false'
GF_ANALYTICS_REPORTING_ENABLED: 'false'
GF_ANALYTICS_CHECK_FOR_UPDATES: 'false'
volumes:
- ./infra/grafana/provisioning:/etc/grafana/provisioning:ro
- ./infra/grafana/dashboards:/var/lib/grafana/dashboards:ro
- grafana_data:/var/lib/grafana
ports: ['127.0.0.1:3000:3000']
networks: [edge, backend]
mem_limit: 384m
cpus: 0.5
restart: unless-stopped
networks:
edge:
backend:
@@ -226,3 +307,7 @@ volumes:
postgres_data:
redis_data:
s3_data:
prometheus_data:
loki_data:
tempo_data:
grafana_data:
+17
View File
@@ -0,0 +1,17 @@
{
"lab": 6,
"workload": "closed-loop POST messages; no automatic retries",
"requested_duration_seconds": 30.0,
"elapsed_seconds": 30.157,
"concurrency": 4,
"successful_requests": 2389,
"failed_requests": 0,
"error_ratio": 0.0,
"successful_rps": 79.219,
"successful_latency_ms": {
"p50": 40.784,
"p95": 99.52,
"p99": 129.314
},
"note": "Latency percentiles cover successes only. Timeouts/errors are counted separately; closed-loop has coordinated omission and client overhead."
}
File diff suppressed because it is too large Load Diff
+5
View File
@@ -0,0 +1,5 @@
PASS: offline skeleton structure, local refs and Python syntax; application NOT tested.
PASS: lab6 public API scenario; infrastructure and grading evidence remain separate.
Running smoke
PASS: smoke
PASS: Grafana provisioning (18 panels, 3 datasources), all dashboard queries, no active alerts; local secrets absent from Git candidates; assignment/contracts/public tests unchanged.
+716
View File
@@ -0,0 +1,716 @@
{
"uid": "messenger-lab6",
"title": "Messenger / Lab 6",
"schemaVersion": 41,
"version": 1,
"refresh": "5s",
"time": {
"from": "now-30m",
"to": "now"
},
"tags": [
"lab6"
],
"panels": [
{
"id": 1,
"title": "API requests / second",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 0,
"w": 12,
"h": 8
},
"description": "Business requests only; health probes are excluded.",
"targets": [
{
"refId": "A",
"expr": "sum by (method,route) (rate(http_requests_total{service_name=\"messenger-api\"}[2m]))",
"legendFormat": "{{method}} {{route}}"
}
],
"fieldConfig": {
"defaults": {
"unit": "reqps"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 2,
"title": "API server error ratio",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 0,
"w": 12,
"h": 8
},
"description": "5xx / all non-health requests, rolling 2m. 4xx are not server errors.",
"targets": [
{
"refId": "A",
"expr": "messenger:http_error_ratio:2m",
"legendFormat": "5xx / all"
}
],
"fieldConfig": {
"defaults": {
"unit": "percentunit"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 3,
"title": "API duration p50 / p95 / p99",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 8,
"w": 12,
"h": 8
},
"description": "Quantiles from merged histogram buckets across API replicas. Never averages of per-instance percentiles.",
"targets": [
{
"refId": "A",
"expr": "histogram_quantile(0.5, sum by (le) (rate(http_duration_seconds_bucket{service_name=\"messenger-api\"}[2m])))",
"legendFormat": "p50"
},
{
"refId": "B",
"expr": "histogram_quantile(0.95, sum by (le) (rate(http_duration_seconds_bucket{service_name=\"messenger-api\"}[2m])))",
"legendFormat": "p95"
},
{
"refId": "C",
"expr": "histogram_quantile(0.99, sum by (le) (rate(http_duration_seconds_bucket{service_name=\"messenger-api\"}[2m])))",
"legendFormat": "p99"
}
],
"fieldConfig": {
"defaults": {
"unit": "s"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 4,
"title": "Job attempts / second",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 8,
"w": 12,
"h": 8
},
"description": "One count per claimed attempt; failed includes retryable failures, skipped is a fenced attempt.",
"targets": [
{
"refId": "A",
"expr": "sum by (outcome) (rate(jobs_processed_total[2m]))",
"legendFormat": "{{outcome}}"
}
],
"fieldConfig": {
"defaults": {
"unit": "ops"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 5,
"title": "Worker processing duration p50 / p95 / p99",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 16,
"w": 12,
"h": 8
},
"description": "Includes isolated image process and durable result check; excludes queue wait.",
"targets": [
{
"refId": "A",
"expr": "histogram_quantile(0.5, sum by (le) (rate(job_duration_seconds_bucket[2m])))",
"legendFormat": "p50"
},
{
"refId": "B",
"expr": "histogram_quantile(0.95, sum by (le) (rate(job_duration_seconds_bucket[2m])))",
"legendFormat": "p95"
},
{
"refId": "C",
"expr": "histogram_quantile(0.99, sum by (le) (rate(job_duration_seconds_bucket[2m])))",
"legendFormat": "p99"
}
],
"fieldConfig": {
"defaults": {
"unit": "s"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 6,
"title": "Time from accepted upload to ready",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 16,
"w": 12,
"h": 8
},
"description": "Database timestamps: durable attachment creation to ready update. Includes queue wait and retries; observed on successful attempts.",
"targets": [
{
"refId": "A",
"expr": "histogram_quantile(0.5, sum by (le) (rate(attachment_time_to_ready_seconds_bucket[5m])))",
"legendFormat": "p50"
},
{
"refId": "B",
"expr": "histogram_quantile(0.95, sum by (le) (rate(attachment_time_to_ready_seconds_bucket[5m])))",
"legendFormat": "p95"
},
{
"refId": "C",
"expr": "histogram_quantile(0.99, sum by (le) (rate(attachment_time_to_ready_seconds_bucket[5m])))",
"legendFormat": "p99"
}
],
"fieldConfig": {
"defaults": {
"unit": "s"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 7,
"title": "Durable queue depth",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 24,
"w": 12,
"h": 8
},
"description": "PostgreSQL queued attachments, including unpublished outbox/retries; excludes processing jobs.",
"targets": [
{
"refId": "A",
"expr": "max(queue_depth)",
"legendFormat": "queued"
}
],
"fieldConfig": {
"defaults": {
"unit": "short"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 8,
"title": "Oldest queued attachment age",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 24,
"w": 12,
"h": 8
},
"description": "Age since acceptance. A stale sample is not zero; check queue sample age below.",
"targets": [
{
"refId": "A",
"expr": "max(queue_oldest_age_seconds)",
"legendFormat": "age"
}
],
"fieldConfig": {
"defaults": {
"unit": "s"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 9,
"title": "DB pool used / acquiring",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 32,
"w": 12,
"h": 8
},
"description": "Per long-lived process. Acquiring includes connection creation and pool wait; max pool size is 10.",
"targets": [
{
"refId": "A",
"expr": "db_pool_used",
"legendFormat": "{{service_name}} {{service_instance_id}} used"
},
{
"refId": "B",
"expr": "db_pool_waiting",
"legendFormat": "{{service_name}} {{service_instance_id}} acquiring"
}
],
"fieldConfig": {
"defaults": {
"unit": "short"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 10,
"title": "Worker saturation",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 32,
"w": 12,
"h": 8
},
"description": "Busy consumers / configured concurrency. Missing capacity after 30s indicates stopped worker.",
"targets": [
{
"refId": "A",
"expr": "sum(worker_busy)",
"legendFormat": "busy"
},
{
"refId": "B",
"expr": "sum(worker_capacity)",
"legendFormat": "capacity"
}
],
"fieldConfig": {
"defaults": {
"unit": "short"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 11,
"title": "Container memory (includes image children)",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 40,
"w": 12,
"h": 8
},
"description": "cgroup v2 memory.current. Includes child processes and charged cache.",
"targets": [
{
"refId": "A",
"expr": "container_memory_bytes",
"legendFormat": "{{service_name}} {{service_instance_id}}"
}
],
"fieldConfig": {
"defaults": {
"unit": "bytes"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 12,
"title": "Container CPU",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 40,
"w": 12,
"h": 8
},
"description": "cgroup v2 CPU seconds / second; 1.0 is one fully used CPU core.",
"targets": [
{
"refId": "A",
"expr": "rate(container_cpu_seconds_total[2m])",
"legendFormat": "{{service_name}} {{service_instance_id}}"
}
],
"fieldConfig": {
"defaults": {
"unit": "short"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 13,
"title": "API availability SLI / SLO",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 48,
"w": 12,
"h": 8
},
"description": "Target: >=99% non-5xx business requests in a rolling 30m window. No traffic is not evidence that SLO is met.",
"targets": [
{
"refId": "A",
"expr": "1 - ((sum(increase(http_requests_total{service_name=\"messenger-api\",status_class=\"5xx\"}[30m])) or vector(0)) / clamp_min(sum(increase(http_requests_total{service_name=\"messenger-api\"}[30m])), 1))",
"legendFormat": "SLI over 30m"
},
{
"refId": "B",
"expr": "vector(0.99)",
"legendFormat": "SLO"
}
],
"fieldConfig": {
"defaults": {
"unit": "percentunit"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 14,
"title": "Local alerts",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 48,
"w": 12,
"h": 8
},
"description": "1 means pending/firing. Series disappears after resolution. Rules and runbooks are stored in infra/otel.",
"targets": [
{
"refId": "A",
"expr": "ALERTS{alertname=~\"ApiHighErrorRate|QueueLagHigh\"}",
"legendFormat": "{{alertname}} {{alertstate}}"
}
],
"fieldConfig": {
"defaults": {
"unit": "short"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 15,
"title": "Collector memory",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 56,
"w": 12,
"h": 8
},
"description": "",
"targets": [
{
"refId": "A",
"expr": "otelcol_process_memory_rss_bytes",
"legendFormat": "RSS"
}
],
"fieldConfig": {
"defaults": {
"unit": "bytes"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 16,
"title": "Collector exporter queue / capacity",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 56,
"w": 12,
"h": 8
},
"description": "Bounded queues: 32 batches per backend, up to 256 records per batch.",
"targets": [
{
"refId": "A",
"expr": "otelcol_exporter_queue_size",
"legendFormat": "{{exporter}} queued"
},
{
"refId": "B",
"expr": "otelcol_exporter_queue_capacity",
"legendFormat": "{{exporter}} capacity"
}
],
"fieldConfig": {
"defaults": {
"unit": "short"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 17,
"title": "Collector failed sends / rejected enqueue",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 0,
"y": 64,
"w": 12,
"h": 8
},
"description": "Send failures may include retries, not unique lost spans. Exact experiment loss is measured by accepted/sent deltas after draining.",
"targets": [
{
"refId": "A",
"expr": "rate(otelcol_exporter_send_failed_spans_total[2m])",
"legendFormat": "{{exporter}} span send failures"
},
{
"refId": "B",
"expr": "rate(otelcol_exporter_send_failed_log_records_total[2m])",
"legendFormat": "{{exporter}} log send failures"
},
{
"refId": "C",
"expr": "rate(otelcol_exporter_enqueue_failed_spans_total[2m])",
"legendFormat": "{{exporter}} span queue rejections"
}
],
"fieldConfig": {
"defaults": {
"unit": "ops"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
},
{
"id": 18,
"title": "Queue sample age",
"type": "timeseries",
"datasource": {
"type": "prometheus",
"uid": "prometheus"
},
"gridPos": {
"x": 12,
"y": 64,
"w": 12,
"h": 8
},
"description": "Above 15s: queue measurements are stale (e.g. DB outage).",
"targets": [
{
"refId": "A",
"expr": "time() - max(queue_sample_timestamp_seconds)",
"legendFormat": "seconds since DB sample"
}
],
"fieldConfig": {
"defaults": {
"unit": "s"
},
"overrides": []
},
"options": {
"legend": {
"displayMode": "list",
"placement": "bottom"
}
}
}
],
"links": [
{
"title": "Explore logs",
"url": "/explore?left=%7B%22datasource%22%3A%22loki%22%7D"
},
{
"title": "Explore traces",
"url": "/explore?left=%7B%22datasource%22%3A%22tempo%22%7D"
}
]
}
@@ -0,0 +1,7 @@
apiVersion: 1
providers:
- name: messenger
folder: Lab 6
type: file
options:
path: /var/lib/grafana/dashboards
@@ -0,0 +1,41 @@
apiVersion: 1
datasources:
- name: Prometheus
uid: prometheus
type: prometheus
access: proxy
url: http://prometheus:9090
isDefault: true
jsonData:
timeInterval: 5s
exemplarTraceIdDestinations:
- name: trace_id
datasourceUid: tempo
- name: Loki
uid: loki
type: loki
access: proxy
url: http://loki:3100
jsonData:
derivedFields:
- name: trace_id
matcherType: label
matcherRegex: trace_id
datasourceUid: tempo
url: '$${__value.raw}'
urlDisplayLabel: View trace
- name: Tempo
uid: tempo
type: tempo
access: proxy
url: http://tempo:3200
jsonData:
tracesToLogsV2:
datasourceUid: loki
spanStartTimeShift: '-5m'
spanEndTimeShift: '5m'
tags: [{key: service.name, value: service_name}]
filterByTraceID: true
filterBySpanID: false
nodeGraph:
enabled: true
+137 -4
View File
@@ -1,7 +1,140 @@
# Конфигурация наблюдаемости — реализовать
# Лаба 6: запуск и диагностика
Добавьте `collector.yaml` с OTLP receivers, processors и exporters выбранной версии Collector. Укажите явные bind endpoints, внутренние сети, memory limits/batch и поведение при backpressure по уровню. Сопоставьте все три pipeline с реально развёрнутыми backend.
```sh
cp .env.example .env
# Задать локальные пароли в .env (в Git не добавлять).
docker compose up -d --build
python -X utf8 tests/smoke.py --base-url http://127.0.0.1:8080
python -X utf8 tests/load.py --base-url http://127.0.0.1:8080 --duration 30 --concurrency 4 --output evidence/load.json
python -X utf8 student-tests/observability.py --scenario all
```
Добавьте конфигурации storage, Grafana provisioning/datasources, provisioning/dashboards и импортируемые dashboard JSON. На 4 — alert rules и runbooks. Пустой dashboard или debug exporter, который только печатает данные в stdout, не заменяют хранилища/доказательства из `contracts/observability.md`.
Последняя команда создаёт синтетические данные и временно останавливает **только
сервисы текущего Compose-проекта**. В `finally` запускает остановленные сервисы.
Не запускать параллельно с другими проверками этого стенда. Volumes не удаляются.
Санитизированные результаты сохраняются в `evidence/observability.json`.
Не публикуйте OTLP/metrics/административные ports в общую сеть без необходимости. В lab7 реализуйте TLS/mTLS по trust matrix; экспорт Authorization и message bodies запрещён.
Grafana: http://localhost:3000, пользователь `admin`, пароль `GRAFANA_ADMIN_PASSWORD`
из `.env`. Dashboard **Lab 6 / Messenger / Lab 6** и три datasource загружаются
автоматически. Prometheus: http://localhost:9090; Loki API: порт 3100; Tempo API:
3200. Все опубликованные порты слушают только loopback, OTLP наружу не опубликован.
Collector принимает OTLP/HTTP от API, dispatcher и worker. Prometheus хранит
метрики, Loki — логи, Tempo — трассы; Grafana обращается к этим хранилищам.
История сохраняется в именованных volumes, retention — 72 часа.
Версии закреплены в Compose и requirements.txt: OTel Python 1.38.0, semantic
conventions 0.59b0 (зависимость SDK); наши имена метрик перечислены ниже.
Это явная инструментация: auto-instrumentation, сбор SQL и HTTP headers отключены.
## Сигналы
| Prometheus name | Единицы и смысл |
| --- | --- |
| `http_requests_total` | Число бизнес-запросов; method, route template, status_class |
| `http_duration_seconds` | Histogram полного HTTP-запроса, включая отдачу stream |
| `jobs_processed_total` | Попытки обработки; job_type, outcome=ready/failed/skipped |
| `job_duration_seconds` | Histogram времени попытки, без ожидания в очереди |
| `attachment_time_to_ready_seconds` | Histogram от записи attachment до durable ready, включая retry |
| `queue_depth` | Количество queued в PostgreSQL, включая outbox и отложенные retry |
| `queue_oldest_age_seconds` | Возраст старейшего queued от момента приёма, не время обработки |
| `queue_sample_timestamp_seconds` | Unix-время последнего успешного опроса очереди |
| `db_pool_used`, `db_pool_waiting` | Занятые соединения / потоки, получающие соединение (ожидание + reconnect) |
| `worker_busy`, `worker_capacity` | Число занятых consumer / лимит concurrency |
| `process_memory_bytes`, `process_cpu_seconds_total` | RSS и CPU процесса |
| `container_memory_bytes`, `container_cpu_seconds_total` | cgroup v2, включая дочерние процессы обработки |
Histogram buckets, секунды: .005, .01, .025, .05, .1, .25, .5, 1, 2, 3, 5, 10,
30, 60, 120, 300, +Inf. Dashboard содержит PromQL, единицы и описания каждого
графика. p50/p95/p99 считаются `histogram_quantile` по сумме buckets всех реплик.
Метрики очереди отдаёт только dispatcher; это не длина Redis Stream (в нём
ещё могут быть уже взятые сообщения). При недоступной БД sample становится
устаревшим; возраст sample выведен отдельно, QueueLagHigh не использует stale data.
Resource: service.name, service.version=1.6.0, service.instance.id=hostname,
deployment.environment.name=lab. API: messenger-api; worker и его дочерний
процесс: messenger-worker; publisher: messenger-dispatcher. Дочерний процесс
экспортирует только traces/logs, чтобы не создавать отдельные краткоживущие
ряды метрик. Parent измеряет длительность попытки и результат после COMMIT.
Метрики наблюдательные: авария после COMMIT до записи счётчика может потерять
измерение, но не результат обработки. Предметных ID в labels метрик нет.
## Связь сигналов
SERVER span API → traceparent/tracestate в transactional outbox → PRODUCER span
dispatcher → W3C context в JSON сообщения Redis → CONSUMER span worker → context
через окружение изолированного image process → S3/DB CLIENT spans. Retry публикует
новый producer/consumer в той же трассе, с `job.attempt`. Baggage не переносится.
Context сохранён в миграции 008, поэтому рестарт API не разрывает причинную связь.
Логи содержат timestamp/severity/service.name/event и trace_id/span_id активного
span. `http_request` содержит ограниченный `request_id`, совпадающий с
X-Request-Id ответа. Loki сохраняет технические поля как structured metadata;
индексирует только service.name и окружение. Для корреляции достаточно request_id,
trace_id и номера попытки: имена файлов и предметные ID не экспортируются.
В Grafana Explore / Loki:
```logql
{service_name="messenger-api"} | request_id="ЗНАЧЕНИЕ_X_REQUEST_ID"
{service_name=~"messenger-.*"} | trace_id="TRACE_ID"
```
Поле trace_id открывает Tempo; из Tempo доступен переход к логам. В трассе upload
видны `POST /api/v1/chats/{chat_id}/attachments`, `attachment-jobs publish`,
`attachment-jobs process`, `image.process`, `s3.GetObject`, `s3.PutObject`,
`db.SELECT`/`db.UPDATE`. Исключения содержат только тип, без message/stack/SQL.
## SLI/SLO и alerts
SLI доступности = 1 − число 5xx / число всех бизнес-запросов за скользящие **30 минут**.
SLO: **не менее 99%**. Health исключён из числителя и знаменателя; 4xx не считается
отказом сервера. Без запросов SLI не оцениваем. Это учебное окно, а не месячный SLO.
`ApiHighErrorRate`: 5xx > 5% за 2 минуты, минимум 20 запросов, `for: 20s`.
`QueueLagHigh`: queued > 0 и возраст > 30s, свежий sample, `for: 15s`.
Пороги короткие для воспроизводимой локальной демонстрации; алерты не являются
burn-rate исследованием. Rules вычисляет Prometheus; firing/pending видны в
http://localhost:9090/alerts и на dashboard. После восстановления alert пропадает
из instant `ALERTS`; история переходов остаётся в range query. Доставка наружу не настроена.
## Queue lag
1. Открыть queue depth/age и worker capacity/busy. Растущая очередь при отсутствии
capacity означает остановку worker; свежий sample подтверждает доступность БД.
2. В Loki найти `job_published`, затем открыть trace: producer есть, consumer ещё нет.
3. `docker compose start worker`; дождаться `job_processed` outcome=ready, depth=0.
4. Проверить, что QueueLagHigh больше не firing и в той же трассе появился consumer.
## API errors
1. При ApiHighErrorRate найти http_request со status_class=5xx и взять trace_id.
2. В Tempo проверить `db.acquire`/`db.SELECT`: error.type и длительность покажут
отказ подключения/БД. В Loki есть `db_error`/`operation_failed`. Медленный handler
без отказа зависимости не имел бы ошибочного DB span. Посмотреть db_pool_waiting.
3. Проверить PostgreSQL: `docker compose ps postgres`; восстановить
`docker compose start postgres`, дождаться readiness и успешного бизнес-запроса.
4. Через окно 2m error ratio снизится, alert перейдёт в resolved. Sample очереди
тоже должен стать свежим. Сценарий проверки сохраняет трассу ошибочного запроса.
## Недоступность telemetry backend
SDK: очереди по 512 spans/logs, batch 64, HTTP timeout 1s; metric export каждые 5s.
Отправка выполняется в фоновых потоках. Collector: memory limiter 192 MiB,
spike 48 MiB, container limit 256 MiB; batch ≤256 записей; отдельная очередь
32 batches / 2 consumers на Loki и Tempo; retry до 15s, timeout 2s, overflow
не блокирует приём. Конечные ограничения защищают память, но допускают потерю
телеметрии при длительном отказе. API и pipeline не зависят от готовности backend.
Сценарий `backend` останавливает Tempo, продолжает создавать сообщения и
обрабатывать изображения, снимает RSS/очередь Collector, затем запускает Tempo
и проверяет новую трассу. `send_failed_spans` может считать повторные попытки:
его нельзя выдавать за число уникально потерянных spans. Для отдельного окна
тест вычисляет accepted − sent после опустошения очереди (без рестартов Collector)
и сохраняет эти дельты вместе с наблюдаемой памятью. Это потери уже принятых
Collector spans; отброшенные до приёма SDK/receiver в этот счёт не входят.
Prometheus всё это время
работает, поэтому сохраняется история отказа и восстановления.
Первичные источники: [Collector](https://opentelemetry.io/docs/collector/configuration/),
[native Loki OTLP](https://grafana.com/docs/loki/latest/send-data/otel/),
[bounded exporter queues](https://github.com/open-telemetry/opentelemetry-collector/blob/v0.136.0/exporter/exporterhelper/README.md).
+28
View File
@@ -0,0 +1,28 @@
groups:
- name: messenger
rules:
- record: messenger:http_error_ratio:2m
expr: |
(sum(rate(http_requests_total{service_name="messenger-api",status_class="5xx"}[2m])) or vector(0))
/ clamp_min(sum(rate(http_requests_total{service_name="messenger-api"}[2m])), 0.001)
- alert: ApiHighErrorRate
expr: messenger:http_error_ratio:2m > 0.05 and sum(increase(http_requests_total{service_name="messenger-api"}[2m])) >= 20
for: 20s
labels:
severity: warning
annotations:
summary: 'API: more than 5% server errors over 2 minutes'
description: 'SLI is 5xx / all non-health requests. At least 20 requests in the window.'
runbook: 'infra/otel/README.md#api-errors'
- alert: QueueLagHigh
expr: |
max(queue_oldest_age_seconds) > 30
and max(queue_depth) > 0
and time() - max(queue_sample_timestamp_seconds) < 15
for: 15s
labels:
severity: warning
annotations:
summary: 'Oldest queued attachment waits more than 30 seconds'
description: 'Check worker capacity/busy and job errors; backlog is sampled by dispatcher.'
runbook: 'infra/otel/README.md#queue-lag'
+69
View File
@@ -0,0 +1,69 @@
receivers:
otlp:
protocols:
http:
endpoint: 0.0.0.0:4318
processors:
memory_limiter:
check_interval: 1s
limit_mib: 192
spike_limit_mib: 48
batch:
timeout: 1s
send_batch_size: 128
send_batch_max_size: 256
exporters:
prometheus:
endpoint: 0.0.0.0:8889
resource_to_telemetry_conversion:
enabled: true
enable_open_metrics: true
metric_expiration: 30s
translation_strategy: UnderscoreEscapingWithSuffixes
otlphttp/loki:
endpoint: http://loki:3100/otlp
timeout: 2s
sending_queue:
queue_size: 32
num_consumers: 2
block_on_overflow: false
retry_on_failure:
initial_interval: 1s
max_interval: 5s
max_elapsed_time: 15s
otlphttp/tempo:
endpoint: http://tempo:4318
timeout: 2s
sending_queue:
queue_size: 32
num_consumers: 2
block_on_overflow: false
retry_on_failure:
initial_interval: 1s
max_interval: 5s
max_elapsed_time: 15s
service:
telemetry:
metrics:
readers:
- pull:
exporter:
prometheus:
host: 0.0.0.0
port: 8888
pipelines:
metrics:
receivers: [otlp]
processors: [memory_limiter, batch]
exporters: [prometheus]
logs:
receivers: [otlp]
processors: [memory_limiter, batch]
exporters: [otlphttp/loki]
traces:
receivers: [otlp]
processors: [memory_limiter, batch]
exporters: [otlphttp/tempo]
+37
View File
@@ -0,0 +1,37 @@
auth_enabled: false
server:
http_listen_port: 3100
common:
path_prefix: /loki
replication_factor: 1
ring:
kvstore:
store: inmemory
storage:
filesystem:
chunks_directory: /loki/chunks
rules_directory: /loki/rules
schema_config:
configs:
- from: 2024-01-01
store: tsdb
object_store: filesystem
schema: v13
index:
prefix: index_
period: 24h
limits_config:
allow_structured_metadata: true
retention_period: 72h
otlp_config:
resource_attributes:
ignore_defaults: true
attributes_config:
- action: index_label
attributes: [service.name, deployment.environment.name]
compactor:
working_directory: /loki/compactor
retention_enabled: true
delete_request_store: filesystem
analytics:
reporting_enabled: false
+12
View File
@@ -0,0 +1,12 @@
global:
scrape_interval: 5s
evaluation_interval: 5s
rule_files:
- /etc/prometheus/alerts.yaml
scrape_configs:
- job_name: messenger
static_configs:
- targets: ['otel-collector:8889']
- job_name: collector
static_configs:
- targets: ['otel-collector:8888']
+23
View File
@@ -0,0 +1,23 @@
server:
http_listen_port: 3200
distributor:
receivers:
otlp:
protocols:
http:
endpoint: 0.0.0.0:4318
ingester:
trace_idle_period: 10s
max_block_duration: 5m
compactor:
compaction:
block_retention: 72h
storage:
trace:
backend: local
wal:
path: /var/tempo/wal
local:
path: /var/tempo/blocks
usage_report:
reporting_enabled: false
+1
View File
@@ -0,0 +1 @@
ALTER TABLE attachment_jobs ADD COLUMN trace_context jsonb NOT NULL DEFAULT '{}';
+4
View File
@@ -17,3 +17,7 @@ redis==5.2.1
argon2-cffi==23.1.0
boto3==1.40.70
Pillow==12.0.0
opentelemetry-api==1.38.0
opentelemetry-sdk==1.38.0
opentelemetry-exporter-otlp-proto-http==1.38.0
psutil==7.1.2
+1
View File
@@ -9,5 +9,6 @@ if __name__ == '__main__':
host=os.getenv('APP_HOST', '0.0.0.0'),
port=int(os.getenv('APP_PORT', '8080')),
workers=1,
access_log=False, # Structured logs use route templates, never query strings.
timeout_graceful_shutdown=5,
)
+2 -1
View File
@@ -10,6 +10,7 @@ from .admission import reserve_capacity
from .models import Attachment, AttachmentJob, ChatMember
from .object_storage import BUCKET, PREFIX, s3_client
from .store import ensure_uuid, serialize, store
from .telemetry import carrier
MAX_BYTES = min(int(os.getenv('MAX_UPLOAD_BYTES', '10485760')), 10485760)
MAX_PIXELS = min(int(os.getenv('MAX_IMAGE_PIXELS', '40000000')), 40000000)
@@ -71,7 +72,7 @@ def save_attachment(source, chat_id, actor_id, filename, content_type, size, dig
size_bytes=size, sha256=digest, staging_key=staging_key)
session.add(attachment)
session.flush()
session.add(AttachmentJob(attachment_id=attachment_id))
session.add(AttachmentJob(attachment_id=attachment_id, trace_context=carrier()))
result = metadata(attachment)
# An uncertain COMMIT is not followed by DELETE: cleanup checks the DB later.
return result
+57 -2
View File
@@ -1,8 +1,12 @@
import os
from pathlib import Path
from sqlalchemy import create_engine, func, insert, select
from sqlalchemy import create_engine, event, func, insert, select
from sqlalchemy.engine import URL, make_url
from sqlalchemy.pool import QueuePool
from threading import Lock
from . import telemetry as otel
from .models import SchemaMigration
@@ -23,12 +27,63 @@ def database_url():
)
class ObservedPool(QueuePool):
"""Count threads acquiring a connection, including pool wait and reconnect time."""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.waiting = 0
self.wait_lock = Lock()
def connect(self):
with self.wait_lock:
self.waiting += 1
try:
with otel.dependency_span('db.acquire', **{'db.system': 'postgresql'}):
return super().connect()
finally:
with self.wait_lock:
self.waiting -= 1
def create_db_engine():
return create_engine(
engine = create_engine(
database_url(), pool_size=10, max_overflow=0, pool_timeout=2, pool_pre_ping=True,
poolclass=ObservedPool, hide_parameters=True,
connect_args={'connect_timeout': 2,
'options': '-c statement_timeout=3000 -c lock_timeout=2000'},
)
# Only operation names; SQL text, bound values and connection URLs are never exported.
@event.listens_for(engine, 'before_cursor_execute')
def before(conn, cursor, statement, parameters, context, executemany):
if not otel.trace.get_current_span().get_span_context().is_valid:
return
operation = statement.lstrip().split(None, 1)[0].upper() if statement.strip() else 'OTHER'
if operation not in {'SELECT', 'INSERT', 'UPDATE', 'DELETE'}:
operation = 'OTHER'
context.otel_span = otel.tracer.start_span('db.' + operation,
kind=otel.SpanKind.CLIENT, attributes={'db.system': 'postgresql', 'db.operation.name': operation})
@event.listens_for(engine, 'after_cursor_execute')
def after(conn, cursor, statement, parameters, context, executemany):
current = getattr(context, 'otel_span', None)
if current:
current.end()
@event.listens_for(engine, 'handle_error')
def error(context):
current = getattr(context.execution_context, 'otel_span', None)
if current:
current.set_status(otel.Status(otel.StatusCode.ERROR))
current.set_attribute('error.type', type(context.original_exception).__name__)
current.end()
otel.event('db_error', severity=otel.logging.ERROR, error_type=type(context.original_exception).__name__)
if os.getenv('OTEL_CHILD_PROCESS') != 'true':
otel.meter.create_observable_gauge('db_pool_used', callbacks=[
lambda _: [otel.metrics.Observation(engine.pool.checkedout())]])
otel.meter.create_observable_gauge('db_pool_waiting', callbacks=[
lambda _: [otel.metrics.Observation(engine.pool.waiting)]])
return engine
def migrate(engine):
+7 -1
View File
@@ -12,6 +12,7 @@ from .database import create_db_engine
from .jobs import now
from .models import Attachment, AttachmentJob
from .object_storage import BUCKET, s3_client
from . import telemetry as otel
def fit_size(size, bound):
@@ -98,9 +99,13 @@ def main():
import resource
memory = int(os.getenv('JOB_MEMORY_MB', '768')) * 1024 * 1024
resource.setrlimit(resource.RLIMIT_AS, (memory, memory))
otel.setup(child=True)
engine = create_db_engine()
try:
process(sessionmaker(engine, expire_on_commit=False), sys.argv[1], sys.argv[2])
with otel.span('image.process', context=otel.propagator.extract({
'traceparent': os.getenv('TRACEPARENT', ''), 'tracestate': os.getenv('TRACESTATE', '')})):
process(sessionmaker(engine, expire_on_commit=False), sys.argv[1], sys.argv[2])
otel.event('image_completed')
except (UnidentifiedImageError, Image.DecompressionBombError, ValueError, SyntaxError):
return 20
except MemoryError:
@@ -109,6 +114,7 @@ def main():
return 22 # No credentials or decoder/database internals in logs.
finally:
engine.dispose()
otel.shutdown()
return 0
+23 -4
View File
@@ -10,13 +10,14 @@ from uuid import UUID, uuid4
from redis import Redis
from redis.exceptions import ResponseError
from sqlalchemy import and_, or_, select
from sqlalchemy import and_, or_, select, func
from sqlalchemy.orm import sessionmaker
from .database import create_db_engine
from .admission import reserve_capacity
from .models import Attachment, AttachmentJob
from .object_storage import BUCKET, s3_client
from . import telemetry as otel
STREAM = 'attachment-jobs'
GROUP = 'image-workers'
@@ -62,7 +63,8 @@ def fail(session, job, attachment, code, permanent=False):
attachment.error_code = None
delay = min(60, 2 ** min(job.failures, 6)) * random.uniform(0.5, 1.5)
job.available_at = now() + timedelta(seconds=delay)
logging.info('job=%s attempt=%s result=%s code=%s', job.id, job.attempt, attachment.status, code)
otel.event('job_retry' if attachment.status == 'queued' else 'job_failed',
attempt=job.attempt, outcome=attachment.status, error_code=code)
def dispatch_once(sessions, client):
@@ -80,10 +82,24 @@ def dispatch_once(sessions, client):
fail(session, job, attachment, 'lease_expired')
continue
if job.published_at is None:
client.xadd(STREAM, {'job': json.dumps(envelope(job))})
with otel.span('attachment-jobs publish', kind=otel.SpanKind.PRODUCER,
context=otel.propagator.extract(job.trace_context), attributes={
'messaging.system': 'redis', 'messaging.destination.name': STREAM,
'messaging.operation.type': 'publish', 'job.attempt': job.attempt + 1}):
client.xadd(STREAM, {'job': json.dumps({**envelope(job), **otel.carrier()})})
otel.event('job_published', attempt=job.attempt + 1)
job.published_at = now()
def sample_queue(sessions):
# Database is the durable backlog: includes outbox rows not yet in Redis and retries.
with sessions() as session:
depth, oldest = session.execute(select(func.count(), func.min(Attachment.created_at))
.select_from(Attachment).where(Attachment.status == 'queued')).one()
otel.queue_state.update(depth=depth, age=max(0, (now() - oldest).total_seconds()) if oldest else 0,
updated=time.time())
def claim(sessions, message):
if (message.get('schema_version') != 1 or message.get('type') != 'attachment.process.v1'
or message.get('pipeline_version') != 1):
@@ -148,6 +164,7 @@ def main():
retry.add_argument('--reason', required=True)
args = parser.parse_args()
logging.basicConfig(level=logging.INFO)
otel.setup()
engine = create_db_engine()
sessions = sessionmaker(engine, expire_on_commit=False)
try:
@@ -169,11 +186,13 @@ def main():
try:
ensure_group(client)
dispatch_once(sessions, client)
sample_queue(sessions)
except Exception as exc:
logging.warning('Dispatch deferred: %s', type(exc).__name__)
otel.event('dispatch_deferred', error_type=type(exc).__name__)
time.sleep(1)
finally:
engine.dispose()
otel.shutdown()
if __name__ == '__main__':
+7
View File
@@ -18,10 +18,13 @@ from .routers.users import router as users_router
from .routers.chats import router as chats_router
from .routers.auth import router as auth_router
from .routers.attachments import router as attachments_router
from . import telemetry
@asynccontextmanager
async def lifespan(app):
telemetry.setup()
telemetry.seed_http(app.routes)
engine = create_db_engine()
try:
store.sessions = sessionmaker(engine, expire_on_commit=False)
@@ -29,6 +32,7 @@ async def lifespan(app):
finally:
engine.dispose()
sessions.client.close()
telemetry.shutdown()
app = FastAPI(title="MTUSI Messenger", lifespan=lifespan)
@@ -46,3 +50,6 @@ app.include_router(users_router)
app.include_router(chats_router)
app.include_router(auth_router)
app.include_router(attachments_router)
# Outside FastAPI's error middleware: includes generated 500 responses and full streams.
app = telemetry.HttpTelemetry(app)
+1
View File
@@ -97,6 +97,7 @@ class AttachmentJob(Base):
lease_token: Mapped[str | None] = mapped_column(Uuid(as_uuid=False))
created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.clock_timestamp())
replay_audit: Mapped[list] = mapped_column(JSONB, server_default='[]')
trace_context: Mapped[dict] = mapped_column(JSONB, server_default='{}')
class IdempotencyRecord(Base):
+11 -1
View File
@@ -3,6 +3,7 @@ from functools import lru_cache
import boto3
from botocore.config import Config
from .telemetry import dependency_span
BUCKET = os.getenv('S3_BUCKET', 'messenger-attachments')
PREFIX = 'originals/'
@@ -10,7 +11,7 @@ PREFIX = 'originals/'
@lru_cache(maxsize=1)
def s3_client():
return boto3.client(
client = boto3.client(
's3', endpoint_url=os.getenv('S3_ENDPOINT', 'http://s3:9000'),
region_name=os.getenv('S3_REGION', 'us-east-1'),
aws_access_key_id=os.environ['S3_ACCESS_KEY_ID'],
@@ -21,3 +22,12 @@ def s3_client():
request_checksum_calculation='when_required',
response_checksum_validation='when_required'),
)
# Instrument SDK operations without object keys, signed URLs, headers or bytes.
original = client._make_api_call
def call(operation, parameters):
with dependency_span('s3.' + operation, **{'rpc.system': 'aws-api', 'rpc.service': 'S3',
'rpc.method': operation}):
return original(operation, parameters)
client._make_api_call = call
return client
+206
View File
@@ -0,0 +1,206 @@
"""Small, explicit OTel instrumentation. Only allowlisted technical fields leave the app."""
import json
import logging
import os
import re
import socket
import time
from contextlib import contextmanager
from datetime import datetime, timezone
from pathlib import Path
import psutil
from opentelemetry import metrics, trace
from opentelemetry.trace import SpanKind, Status, StatusCode
from opentelemetry.trace.propagation.tracecontext import TraceContextTextMapPropagator
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.sdk.metrics import MeterProvider, Histogram
from opentelemetry.sdk.metrics.export import PeriodicExportingMetricReader
from opentelemetry.sdk.metrics.view import View, ExplicitBucketHistogramAggregation
from opentelemetry.sdk._logs import LoggerProvider, LoggingHandler
from opentelemetry.sdk._logs.export import BatchLogRecordProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.exporter.otlp.proto.http.metric_exporter import OTLPMetricExporter
from opentelemetry.exporter.otlp.proto.http._log_exporter import OTLPLogExporter
tracer = trace.get_tracer('messenger', '1.6.0')
meter = metrics.get_meter('messenger', '1.6.0')
propagator = TraceContextTextMapPropagator() # No baggage: it may contain user data.
logger = logging.getLogger('messenger.events')
logger.propagate = False
logger.setLevel(logging.INFO)
providers = []
LOG_FIELDS = {'method', 'route', 'status_class', 'status', 'duration_seconds', 'request_id',
'operation', 'error_type', 'attempt', 'outcome', 'error_code', 'job_type'}
http_requests = meter.create_counter('http_requests')
http_duration = meter.create_histogram('http_duration', unit='s')
jobs_processed = meter.create_counter('jobs_processed')
job_duration = meter.create_histogram('job_duration', unit='s')
time_to_ready = meter.create_histogram('attachment_time_to_ready', unit='s')
worker_busy = meter.create_up_down_counter('worker_busy')
queue_state = {'depth': 0, 'age': 0, 'updated': 0}
class JsonFormatter(logging.Formatter):
def format(self, record):
context = trace.get_current_span().get_span_context()
return json.dumps({
'timestamp': datetime.fromtimestamp(record.created, timezone.utc).isoformat(),
'severity': record.levelname, 'service.name': os.getenv('OTEL_SERVICE_NAME', 'messenger-api'),
'event': record.getMessage(),
'trace_id': format(context.trace_id, '032x') if context.is_valid else '',
'span_id': format(context.span_id, '016x') if context.is_valid else '',
**{key: record.__dict__[key] for key in LOG_FIELDS if key in record.__dict__},
})
def event(name, *, severity=logging.INFO, **fields):
# Call sites pass fixed event names, safe enums/numbers and technical correlation IDs.
logger.log(severity, name, extra={'event': name, **{k: v for k, v in fields.items() if k in LOG_FIELDS}})
def setup(*, child=False):
if providers or os.getenv('OTEL_SDK_DISABLED', 'false').lower() == 'true':
return
resource = Resource({
'service.name': os.getenv('OTEL_SERVICE_NAME', 'messenger-api'),
'service.version': '1.6.0', 'service.instance.id': socket.gethostname(),
'deployment.environment.name': 'lab',
})
endpoint = os.getenv('OTEL_EXPORTER_OTLP_ENDPOINT', 'http://otel-collector:4318').rstrip('/')
traces = TracerProvider(resource=resource)
traces.add_span_processor(BatchSpanProcessor(
OTLPSpanExporter(endpoint=endpoint + '/v1/traces', timeout=1),
max_queue_size=512, max_export_batch_size=64, schedule_delay_millis=1000))
trace.set_tracer_provider(traces)
logs = LoggerProvider(resource=resource)
logs.add_log_record_processor(BatchLogRecordProcessor(
OTLPLogExporter(endpoint=endpoint + '/v1/logs', timeout=1),
max_queue_size=512, max_export_batch_size=64, schedule_delay_millis=1000))
console = logging.StreamHandler()
console.setFormatter(JsonFormatter())
logger.addHandler(console)
logger.addHandler(LoggingHandler(logger_provider=logs))
providers.extend([traces, logs])
if child:
return # Short-lived image processes emit spans/logs; parent owns worker metrics.
reader = PeriodicExportingMetricReader(
OTLPMetricExporter(endpoint=endpoint + '/v1/metrics', timeout=1),
export_interval_millis=5000, export_timeout_millis=2000)
buckets = [0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2, 3, 5, 10, 30, 60, 120, 300]
metric_provider = MeterProvider(resource=resource, metric_readers=[reader], views=[
View(instrument_type=Histogram,
aggregation=ExplicitBucketHistogramAggregation(buckets))])
metrics.set_meter_provider(metric_provider)
providers.append(metric_provider)
process = psutil.Process()
meter.create_observable_gauge('process_memory', unit='By', callbacks=[
lambda _: [metrics.Observation(process.memory_info().rss)]])
meter.create_observable_counter('process_cpu', unit='s', callbacks=[
lambda _: [metrics.Observation(sum(process.cpu_times()[:2]))]])
cgroup = Path('/sys/fs/cgroup')
if (cgroup / 'memory.current').exists():
meter.create_observable_gauge('container_memory', unit='By', callbacks=[
lambda _: [metrics.Observation(int((cgroup / 'memory.current').read_text()))]])
def cpu_usage(_):
values = dict(line.split() for line in (cgroup / 'cpu.stat').read_text().splitlines())
return [metrics.Observation(int(values['usage_usec']) / 1_000_000)]
meter.create_observable_counter('container_cpu', unit='s', callbacks=[cpu_usage])
if os.getenv('OTEL_SERVICE_NAME') == 'messenger-worker':
capacity = min(4, max(1, int(os.getenv('WORKER_CONCURRENCY', '2'))))
meter.create_observable_gauge('worker_capacity', callbacks=[
lambda _: [metrics.Observation(capacity)]])
worker_busy.add(0)
if os.getenv('OTEL_SERVICE_NAME') == 'messenger-dispatcher':
for name, key, unit in [('queue_depth', 'depth', ''), ('queue_oldest_age', 'age', 's'),
('queue_sample_timestamp', 'updated', 's')]:
meter.create_observable_gauge(name, unit=unit, callbacks=[
lambda _, key=key: [metrics.Observation(queue_state[key], {'queue': 'attachment-jobs'})]])
def shutdown():
for provider in reversed(providers):
provider.shutdown()
providers.clear()
def seed_http(routes):
# Export zero before the first failure: Prometheus rate needs a prior sample.
for route in routes:
if route.path.startswith('/health/'):
continue
for method in route.methods or ():
for status in ('2xx', '3xx', '4xx', '5xx'):
http_requests.add(0, {'method': method, 'route': route.path, 'status_class': status})
def carrier():
result = {}
propagator.inject(result)
return result
@contextmanager
def span(name, *, kind=SpanKind.INTERNAL, context=None, attributes=None):
# Never export exception messages/stack traces (SQL errors can embed parameters).
with tracer.start_as_current_span(name, kind=kind, context=context, attributes=attributes,
record_exception=False, set_status_on_exception=False) as current:
try:
yield current
except Exception as exc:
current.set_status(Status(StatusCode.ERROR))
current.set_attribute('error.type', type(exc).__name__)
event('operation_failed', severity=logging.ERROR, operation=name, error_type=type(exc).__name__)
raise
@contextmanager
def dependency_span(name, **attributes):
# Background health probes and queue polling must not create disconnected trace noise.
if not trace.get_current_span().get_span_context().is_valid:
yield
else:
with span(name, kind=SpanKind.CLIENT, attributes=attributes):
yield
class HttpTelemetry:
def __init__(self, app):
self.app = app
async def __call__(self, scope, receive, send):
if scope['type'] != 'http' or scope['path'].startswith('/health/'):
return await self.app(scope, receive, send)
headers = {k.decode('latin1'): v.decode('latin1') for k, v in scope['headers']
if k in (b'traceparent', b'tracestate')}
method = scope['method'] if scope['method'] in {'GET', 'POST', 'PUT', 'PATCH', 'DELETE', 'HEAD', 'OPTIONS'} else 'OTHER'
status = 500
started = time.monotonic()
with span('HTTP ' + method, kind=SpanKind.SERVER, context=propagator.extract(headers)) as current:
async def response(message):
nonlocal status
if message['type'] == 'http.response.start':
status = message['status']
await send(message)
try:
await self.app(scope, receive, response)
finally:
route = getattr(scope.get('route'), 'path', 'unmatched')
labels = {'method': method, 'route': route, 'status_class': f'{status // 100}xx'}
elapsed = time.monotonic() - started
http_requests.add(1, labels)
http_duration.record(elapsed, labels)
current.update_name(method + ' ' + route)
current.set_attributes({'http.request.method': method, 'http.route': route,
'http.response.status_code': status})
if status >= 500:
current.set_status(Status(StatusCode.ERROR))
request_id = scope.get('state', {}).get('request_id', '')
if not re.fullmatch(r'[A-Za-z0-9_-]{1,64}', request_id):
request_id = ''
current.set_attribute('request_id', request_id)
event('http_request', severity=logging.ERROR if status >= 500 else logging.INFO,
**labels, status=status, duration_seconds=elapsed,
request_id=request_id)
+61 -15
View File
@@ -12,30 +12,75 @@ from sqlalchemy.orm import sessionmaker
from .database import create_db_engine
from .jobs import GROUP, STREAM, LEASE_SECONDS, TIMEOUT, broker, claim, ensure_group, record_failure
from .models import Attachment, AttachmentJob
from . import telemetry as otel
def handle(sessions, client, message_id, fields):
try:
claimed = claim(sessions, json.loads(fields['job']))
message = json.loads(fields['job'])
if not isinstance(message, dict):
raise ValueError('invalid envelope')
except (ValueError, KeyError, TypeError):
logging.error('Invalid broker envelope: id=%s', message_id)
otel.event('invalid_job')
acknowledge(client, message_id)
return
with otel.span('attachment-jobs process', kind=otel.SpanKind.CONSUMER,
context=otel.propagator.extract(message), attributes={
'messaging.system': 'redis', 'messaging.destination.name': STREAM,
'messaging.operation.type': 'process', 'job.attempt': message.get('attempt', 0)}):
process_message(sessions, client, message_id, message)
def process_message(sessions, client, message_id, message):
try:
claimed = claim(sessions, message)
except (ValueError, KeyError, TypeError):
otel.event('invalid_job')
claimed = None
if claimed:
job_id, token = claimed
started = time.monotonic()
outcome = 'failed'
otel.worker_busy.add(1)
try:
result = subprocess.run([sys.executable, '-m', 'src.image_processor', job_id, token],
timeout=TIMEOUT, check=False)
code = result.returncode
except subprocess.TimeoutExpired:
code = 23
if code in (91, 92):
# Simulate a lost worker before ack: leave the Redis pending entry and lease.
return
if code:
record_failure(sessions, job_id, token,
{20: 'invalid_image', 21: 'resource_limit', 23: 'processing_timeout'}
.get(code, 'processing_failed'), permanent=code in (20, 21))
environment = {**os.environ, 'OTEL_CHILD_PROCESS': 'true'}
environment.update({key.upper(): value for key, value in otel.carrier().items()})
try:
result = subprocess.run([sys.executable, '-m', 'src.image_processor', job_id, token],
timeout=TIMEOUT, check=False, env=environment)
code = result.returncode
except subprocess.TimeoutExpired:
code = 23
if code in (91, 92):
# Simulate a lost worker before ack: leave the pending entry and lease.
return
if code:
record_failure(sessions, job_id, token,
{20: 'invalid_image', 21: 'resource_limit', 23: 'processing_timeout'}
.get(code, 'processing_failed'), permanent=code in (20, 21))
otel.trace.get_current_span().set_status(otel.Status(otel.StatusCode.ERROR))
else:
with sessions() as session:
job = session.get(AttachmentJob, job_id)
attachment = session.get(Attachment, job.attachment_id)
outcome = 'ready' if attachment.status == 'ready' else 'skipped'
if outcome == 'ready':
otel.time_to_ready.record((attachment.updated_at - attachment.created_at).total_seconds(),
{'job_type': 'attachment.process.v1'})
finally:
otel.worker_busy.add(-1)
labels = {'job_type': 'attachment.process.v1', 'outcome': outcome}
elapsed = time.monotonic() - started
otel.jobs_processed.add(1, labels)
otel.job_duration.record(elapsed, labels)
otel.event('job_processed', severity=logging.ERROR if outcome == 'failed' else logging.INFO,
**labels, duration_seconds=elapsed)
# Only after a durable ready/failed/retry (or an existing durable lease).
acknowledge(client, message_id)
def acknowledge(client, message_id):
with client.pipeline(transaction=True) as pipe:
pipe.xack(STREAM, GROUP, message_id)
pipe.xdel(STREAM, message_id)
@@ -58,12 +103,13 @@ def consume(sessions):
for message_id, fields in messages:
handle(sessions, client, message_id, fields)
except Exception as exc:
logging.warning('Worker deferred: %s', type(exc).__name__)
otel.event('worker_deferred', error_type=type(exc).__name__)
time.sleep(1)
def main():
logging.basicConfig(level=logging.INFO)
otel.setup()
engine = create_db_engine()
sessions = sessionmaker(engine, expire_on_commit=False)
concurrency = min(4, max(1, int(os.getenv('WORKER_CONCURRENCY', '2'))))
+290
View File
@@ -0,0 +1,290 @@
"""Lab 6 acceptance: real signals, two incidents, retry propagation and bounded outage.
Uses only the standard library. Stops services only in this Compose project and
always starts them again; never deletes volumes. Run separately from other tests.
"""
import argparse
import json
import subprocess
import sys
import time
import urllib.error
import urllib.parse
import urllib.request
import uuid
from datetime import datetime, timezone
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / 'tests'))
from client import Client, check
PROM = 'http://127.0.0.1:9090'
LOKI = 'http://127.0.0.1:3100'
TEMPO = 'http://127.0.0.1:3200'
def stamp():
return datetime.now(timezone.utc).isoformat()
def get(url, **params):
if params:
url += '?' + urllib.parse.urlencode(params)
with urllib.request.urlopen(urllib.request.Request(url, headers={'Accept': 'application/json'}), timeout=10) as r:
return json.load(r)
def query(expr):
data = get(PROM + '/api/v1/query', query=expr)
check(data['status'] == 'success', 'PromQL failed')
return data['data']['result']
def number(expr):
values = query(expr)
return float(values[0]['value'][1]) if values else 0.0
def wait(label, fn, timeout=90):
end = time.monotonic() + timeout
last = None
while time.monotonic() < end:
try:
value = fn()
if value:
return value
except (urllib.error.URLError, TimeoutError, AssertionError) as exc:
last = type(exc).__name__
time.sleep(2)
raise AssertionError(f'Timeout: {label} ({last})')
def compose(*args):
subprocess.run(['docker', 'compose', *args], cwd=ROOT, check=True,
stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, timeout=90)
def trace_data(trace_id):
data = get(TEMPO + '/api/traces/' + trace_id)
spans = []
for batch in data.get('batches', data.get('resourceSpans', [])):
resource = {a['key']: a['value'].get('stringValue') for a in batch.get('resource', {}).get('attributes', [])}
for scope in batch.get('scopeSpans', batch.get('instrumentationLibrarySpans', [])):
for span in scope.get('spans', []):
spans.append({**span, 'service': resource.get('service.name')})
return data, spans
def logs(trace_id):
return get(LOKI + '/loki/api/v1/query_range',
query='{service_name=~"messenger-.*"} | trace_id="' + trace_id + '"',
start=str(time.time_ns() - 3_600_000_000_000), limit=1000)['data']['result']
def attrs(span):
return {a['key']: next(iter(a['value'].values())) for a in span.get('attributes', [])}
def pipeline(trace_id, *, retry=False):
data, spans = trace_data(trace_id)
names = {s['name'] for s in spans}
required = {'POST /api/v1/chats/{chat_id}/attachments', 'attachment-jobs publish',
'attachment-jobs process', 'image.process', 's3.GetObject', 's3.PutObject', 'db.UPDATE'}
if not required <= names:
return None
by_id = {s['spanId']: s for s in spans}
for name, parent_name in [('attachment-jobs publish', 'POST /api/v1/chats/{chat_id}/attachments'),
('attachment-jobs process', 'attachment-jobs publish'),
('image.process', 'attachment-jobs process')]:
for s in (s for s in spans if s['name'] == name):
check(by_id.get(s.get('parentSpanId'), {}).get('name') == parent_name,
'Broken trace parent for ' + name)
check(s.get('traceState') == 'lab=check', 'W3C tracestate lost at ' + name)
attempts = sorted({int(attrs(s)['job.attempt']) for s in spans
if s['name'] == 'attachment-jobs process'})
if retry and len(attempts) < 2:
return None
records = logs(trace_id)
services = {r['stream'].get('service_name') for r in records}
if not {'messenger-api', 'messenger-worker', 'messenger-dispatcher'} <= services:
return None
return {'trace_id': trace_id, 'span_count': len(spans), 'span_names': sorted(names),
'log_services': sorted(services), 'attempts': attempts, 'parent_chain_verified': True,
'tracestate_preserved': True,
'_raw': json.dumps([data, records])}
class Scenario:
def __init__(self, base_url):
self.client = Client(base_url)
self.alice, _, _, self.chat = self.client.bootstrap()
self.private_markers = [self.alice['token'], self.alice['registration']['password']]
def upload(self):
trace_id, request_id = uuid.uuid4().hex, str(uuid.uuid4())
filename = 'private_filename_' + uuid.uuid4().hex + '.png'
self.private_markers.append(filename)
result = self.client.call('POST', '/api/v1/chats/' + self.chat['id'] + '/attachments?filename=' + filename,
202, actor=self.alice, raw=(ROOT / 'tests/fixtures/sample.png').read_bytes(),
headers={'Content-Type': 'image/png', 'traceparent': f'00-{trace_id}-{uuid.uuid4().hex[:16]}-01',
'tracestate': 'lab=check', 'X-Request-Id': request_id}, schema='Attachment')
check(result['headers']['X-Request-Id'] == request_id, 'Request ID changed')
return result['json'], trace_id, request_id
def check_pipeline(self, trace_id, request_id, retry=False):
result = wait('pipeline spans/logs', lambda: pipeline(trace_id, retry=retry))
raw = result.pop('_raw')
check(all(value not in raw for value in self.private_markers), 'Private input leaked into telemetry')
check(request_id in raw, 'Request ID is missing from trace/log correlation')
result['request_id'] = request_id
result['private_inputs_absent'] = True
return result
def business(self, expected=200, trace_id=None):
headers = {'traceparent': f'00-{trace_id}-{uuid.uuid4().hex[:16]}-01'} if trace_id else {}
return self.client.call('GET', '/api/v1/chats/' + self.chat['id'], expected,
actor=self.alice, headers=headers, schema='Chat' if expected == 200 else None)
def smoke(self):
att, trace_id, request_id = self.upload()
self.client.wait_ready(self.alice, att)
result = self.check_pipeline(trace_id, request_id)
names = ['http_requests_total', 'http_duration_seconds_bucket', 'jobs_processed_total',
'job_duration_seconds_bucket', 'attachment_time_to_ready_seconds_bucket',
'queue_depth', 'queue_oldest_age_seconds', 'db_pool_used', 'db_pool_waiting',
'worker_busy', 'worker_capacity', 'container_cpu_seconds_total', 'container_memory_bytes']
for name in names:
wait(name, lambda name=name: query(name))
series = get(PROM + '/api/v1/series', **{'match[]': '{job="messenger"}', 'start': str(time.time() - 300)})['data']
forbidden = {'user_id', 'chat_id', 'message_id', 'attachment_id', 'request_id', 'filename', 'url'}
check(all(not forbidden.intersection(row) for row in series), 'High-cardinality metric label')
check(all(self.chat['id'] not in json.dumps(row) for row in series), 'Raw route ID in metrics')
result['metrics_present'] = names
return result
def queue(self):
result = {'started_at': stamp()}
compose('stop', 'worker')
try:
att, trace_id, request_id = self.upload()
self.client.message(self.alice, self.chat, 'queued attachment', attachments=[att['id']])
wait('QueueLagHigh firing', lambda: number('ALERTS{alertname="QueueLagHigh",alertstate="firing"}') == 1)
result.update(firing_at=stamp(), depth=number('max(queue_depth)'),
oldest_age_seconds=number('max(queue_oldest_age_seconds)'))
finally:
compose('start', 'worker')
ready = self.client.wait_ready(self.alice, att)
wait('queue alert resolved', lambda: not query('ALERTS{alertname="QueueLagHigh"}'))
result.update(resolved_at=stamp(), ready=ready['status'], depth_after=number('max(queue_depth)'),
pipeline=self.check_pipeline(trace_id, request_id))
return result
def database(self):
result = {'started_at': stamp()}
trace_id = uuid.uuid4().hex
compose('stop', 'postgres')
try:
errors = 0
def load_until_firing():
nonlocal errors
self.business(expected=503, trace_id=trace_id if errors == 0 else None)
errors += 1
return number('ALERTS{alertname="ApiHighErrorRate",alertstate="firing"}') == 1
wait('ApiHighErrorRate firing under sustained traffic', load_until_firing, timeout=120)
result['failed_requests'] = errors
result.update(firing_at=stamp(), error_ratio=number('messenger:http_error_ratio:2m'))
def failed_trace():
_, spans = trace_data(trace_id)
failed = [s for s in spans if s['name'].startswith('db.')
and s.get('status', {}).get('code') in (2, 'STATUS_CODE_ERROR')]
return [{'name': s['name'], 'error_type': attrs(s).get('error.type'),
'duration_seconds': (int(s['endTimeUnixNano']) - int(s['startTimeUnixNano'])) / 1e9}
for s in failed]
result['db_error_spans'] = wait('failed DB span', failed_trace)
result['trace_id'] = trace_id
result['correlated_log_streams'] = len(wait('DB error logs', lambda: logs(trace_id)))
finally:
compose('start', 'postgres')
wait('API recovered', self.business)
for _ in range(10):
self.business()
wait('API alert resolved', lambda: not query('ALERTS{alertname="ApiHighErrorRate"}'), timeout=160)
result['resolved_at'] = stamp()
return result
def retry(self):
failures = number('sum(jobs_processed_total{outcome="failed"})')
compose('stop', 'worker')
try:
att, trace_id, request_id = self.upload()
compose('stop', 's3')
# `compose start worker` can also start its dependency graph, including S3.
compose('up', '-d', '--no-deps', '--no-recreate', 'worker')
wait('failed processing attempt', lambda: number('sum(jobs_processed_total{outcome="failed"})') > failures, timeout=40)
finally:
compose('start', 's3', 'worker')
wait('S3/API ready', lambda: self.client.call('GET', '/health/ready', schema='Health'))
self.client.wait_ready(self.alice, att)
return self.check_pipeline(trace_id, request_id, retry=True)
def backend(self):
accepted = 'sum(otelcol_receiver_accepted_spans_total)'
sent = 'sum(otelcol_exporter_sent_spans_total{exporter="otlphttp/tempo"})'
queue = 'sum(otelcol_exporter_queue_size{exporter="otlphttp/tempo"})'
# A quiet interval on both ends makes accepted-minus-sent a loss count,
# not a count of failed export attempts or currently buffered spans.
time.sleep(8)
wait('trace queue drained', lambda: number(queue) == 0)
before = {'accepted': number(accepted), 'sent': number(sent)}
result = {'started_at': stamp(), 'before': before}
samples, writes = [], 0
compose('stop', 'tempo')
try:
end = time.monotonic() + 60
att, _, _ = self.upload()
self.client.wait_ready(self.alice, att)
while time.monotonic() < end:
body = 'private_message_' + uuid.uuid4().hex
self.private_markers.append(body)
self.client.message(self.alice, self.chat, body)
writes += 1
samples.append({'at': stamp(), 'collector_rss_bytes': number('otelcol_process_memory_rss_bytes'),
'queue_batches': number(queue)})
time.sleep(0.3)
check(writes > 10, 'Not enough writes during telemetry outage')
check(max(s['queue_batches'] for s in samples) <= 32, 'Collector queue is unbounded')
check(0 < max(s['collector_rss_bytes'] for s in samples) < 256 * 1024 * 1024,
'Collector exceeded its memory budget')
finally:
compose('start', 'tempo')
wait('trace queue drained', lambda: number(queue) == 0)
time.sleep(10) # Include final Prometheus scrape after drain.
after = {'accepted': number(accepted), 'sent': number(sent)}
lost = (after['accepted'] - before['accepted']) - (after['sent'] - before['sent'])
check(lost >= 0, 'Loss window was not quiescent; rerun backend separately')
result.update(after=after, lost_accepted_spans=lost, successful_writes=writes,
attachment_ready_during_outage=True, samples=samples, recovered_at=stamp())
att, trace_id, request_id = self.upload()
self.client.wait_ready(self.alice, att)
result['recovered_pipeline'] = self.check_pipeline(trace_id, request_id)
return result
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--scenario', choices=['smoke', 'queue', 'database', 'retry', 'backend', 'all'], default='smoke')
parser.add_argument('--base-url', default='http://127.0.0.1:8080')
args = parser.parse_args()
scenario = Scenario(args.base_url)
path = ROOT / 'evidence/observability.json'
evidence = json.loads(path.read_text(encoding='utf-8')) if path.exists() else {}
for name in (['smoke', 'queue', 'database', 'retry', 'backend'] if args.scenario == 'all' else [args.scenario]):
print('Running ' + name, flush=True)
evidence[name] = {'checked_at': stamp(), **getattr(scenario, name)()}
path.write_text(json.dumps(evidence, indent=2, ensure_ascii=False) + '\n', encoding='utf-8')
print('PASS: ' + name, flush=True)
if __name__ == '__main__':
main()