forked from mtusi2026/mtusi2026lab3
Найдено первым прогоном из чистого clone: make test-failover (BASE_URL по умолчанию http://localhost:8080) успел сделать 8 записей за 15 с, и первая успешная запись после хука пришла на 8,4 с. Windows сначала стучится на ::1, где порт не был опубликован, и каждое новое соединение ждёт ~2 с. LB теперь опубликован на 127.0.0.1 и [::1] (оба loopback): 0,013 с на запрос. Тест политики портов допускает только loopback. collect_evidence: публичный тест всегда выбирает реплику, обслужившую его первый POST, а последовательность запросов фиксирована — перед каждым следующим прогоном rotation сдвигается лишними GET, чтобы по очереди останавливались обе реплики. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
185 lines
7.0 KiB
YAML
185 lines
7.0 KiB
YAML
# Lab 3 stand: client -> HAProxy (lb) -> API x N (one scalable service `api`) -> PostgreSQL + Valkey.
|
|
# From a clean clone (details and Windows notes: REPORT.md, docs/WINDOWS.md):
|
|
# python scripts/init_secrets.py # once: random local credentials into ./secrets (git-ignored)
|
|
# cp .env.example .env # optional overrides (API_REPLICAS, limits, ...)
|
|
# docker compose --env-file .env up -d --build --wait
|
|
# Order: postgres/valkey healthy -> migrate (ONE run, exits 0) -> api replicas healthy -> lb.
|
|
# Images are pinned by version AND digest. Passwords are passed only as secret files, and
|
|
# every service gets only the secrets it needs. Only the LB publishes ports (loopback only).
|
|
# Two networks: `edge` (lb <-> api) and `data` (api/migrate <-> postgres/valkey, internal):
|
|
# the LB cannot reach the stores, and the stores are not reachable from the edge.
|
|
name: mtusi-lab3
|
|
|
|
x-api-env: &api-env
|
|
APP_ENV: ${APP_ENV:-lab}
|
|
LOG_LEVEL: ${LOG_LEVEL:-info}
|
|
SESSION_TTL_SECONDS: ${SESSION_TTL_SECONDS:-3600}
|
|
SESSION_STORE: ${SESSION_STORE:-valkey}
|
|
DATABASE_URL_FILE: /run/secrets/database_url
|
|
CACHE_URL_FILE: /run/secrets/cache_url
|
|
DB_POOL_MIN: ${DB_POOL_MIN:-1}
|
|
DB_POOL_MAX: ${DB_POOL_MAX:-10}
|
|
DB_TIMEOUT_SECONDS: ${DB_TIMEOUT_SECONDS:-3}
|
|
CACHE_TIMEOUT_SECONDS: ${CACHE_TIMEOUT_SECONDS:-0.5}
|
|
PASSWORD_HASH_CONCURRENCY: ${PASSWORD_HASH_CONCURRENCY:-2}
|
|
IDEMPOTENCY_TTL_SECONDS: ${IDEMPOTENCY_TTL_SECONDS:-86400}
|
|
# Graceful stop: 3 s drain (readiness 503, still serving) + up to 8 s for in-flight requests.
|
|
DRAIN_SECONDS: ${DRAIN_SECONDS:-3}
|
|
SHUTDOWN_TIMEOUT_SECONDS: ${SHUTDOWN_TIMEOUT_SECONDS:-8}
|
|
DOCS_ENABLED: ${DOCS_ENABLED:-true}
|
|
# Track 1 improvement: `function` = one SQL call per send (default), `multi` = old path (A/B).
|
|
MESSAGE_WRITE_PATH: ${MESSAGE_WRITE_PATH:-function}
|
|
|
|
services:
|
|
postgres:
|
|
image: postgres:18.6-bookworm@sha256:afc7e2d441324c0388fa80c3d24f733b4194a4eb7f47dd8ee2b08eb1a24a647c
|
|
environment:
|
|
POSTGRES_USER: messenger
|
|
POSTGRES_DB: messenger
|
|
POSTGRES_PASSWORD_FILE: /run/secrets/postgres_password
|
|
# pg_stat_statements: per-statement counters for the track 5 experiments (bench/track1).
|
|
command: ["postgres", "-c", "shared_preload_libraries=pg_stat_statements", "-c", "max_connections=${PG_MAX_CONNECTIONS:-100}"]
|
|
secrets: [postgres_password]
|
|
volumes:
|
|
- pgdata:/var/lib/postgresql # PostgreSQL 18 image keeps PGDATA under this path
|
|
healthcheck:
|
|
test: ["CMD-SHELL", "pg_isready -U messenger -d messenger -q"]
|
|
interval: 2s
|
|
timeout: 3s
|
|
retries: 30
|
|
start_period: 10s
|
|
stop_grace_period: 30s # clean shutdown checkpoint instead of crash recovery
|
|
restart: unless-stopped
|
|
networks: [data]
|
|
deploy:
|
|
resources:
|
|
limits: {cpus: "${PG_CPUS:-2.0}", memory: "${PG_MEMORY:-1g}"}
|
|
|
|
valkey:
|
|
image: valkey/valkey:8.1.3-bookworm@sha256:fea8b3e67b15729d4bb70589eb03367bab9ad1ee89c876f54327fc7c6e618571
|
|
# The password goes into a config include readable only by the valkey user,
|
|
# so it never appears in the process arguments. docker-entrypoint.sh drops root.
|
|
command:
|
|
- sh
|
|
- -c
|
|
- >-
|
|
umask 077 &&
|
|
printf 'requirepass %s\n' "$$(cat /run/secrets/valkey_password)" > /tmp/requirepass.conf &&
|
|
chown valkey /tmp/requirepass.conf &&
|
|
exec docker-entrypoint.sh valkey-server /etc/valkey/valkey.conf
|
|
secrets: [valkey_password]
|
|
volumes:
|
|
- ./infra/valkey/valkey.conf:/etc/valkey/valkey.conf:ro
|
|
- valkeydata:/data
|
|
healthcheck:
|
|
test: ["CMD-SHELL", "REDISCLI_AUTH=\"$$(cat /run/secrets/valkey_password)\" valkey-cli ping | grep -q PONG"]
|
|
interval: 2s
|
|
timeout: 3s
|
|
retries: 30
|
|
start_period: 5s
|
|
restart: unless-stopped
|
|
networks: [data]
|
|
deploy:
|
|
resources:
|
|
limits: {cpus: "${VALKEY_CPUS:-0.5}", memory: 256m}
|
|
|
|
# One-shot schema migration: runs ONCE per `up`, before any API replica starts
|
|
# (api depends_on service_completed_successfully). Replicas never migrate themselves.
|
|
migrate:
|
|
build:
|
|
context: .
|
|
args:
|
|
BUILD_REVISION: ${BUILD_REVISION:-dev}
|
|
image: ${API_IMAGE:-mtusi-messenger:lab3}
|
|
command: ["python", "-m", "messenger.migrate"]
|
|
environment:
|
|
DATABASE_URL_FILE: /run/secrets/database_url
|
|
secrets: [database_url]
|
|
depends_on:
|
|
postgres:
|
|
condition: service_healthy
|
|
healthcheck:
|
|
disable: true
|
|
restart: "no"
|
|
networks: [data]
|
|
deploy:
|
|
resources:
|
|
limits: {cpus: "0.5", memory: 256m}
|
|
|
|
api:
|
|
build:
|
|
context: .
|
|
args:
|
|
BUILD_REVISION: ${BUILD_REVISION:-dev}
|
|
image: ${API_IMAGE:-mtusi-messenger:lab3}
|
|
environment: *api-env
|
|
secrets: [database_url, cache_url]
|
|
# No `ports` and no container_name: replicas are reachable only through `lb`,
|
|
# and `--scale api=N` / deploy.replicas works. X-Instance-Id = container hostname.
|
|
expose: ["8080"]
|
|
healthcheck:
|
|
test: ["CMD", "python", "-m", "messenger.healthcheck"]
|
|
interval: 2s
|
|
timeout: 2s
|
|
retries: 3
|
|
start_period: 5s
|
|
depends_on:
|
|
postgres:
|
|
condition: service_healthy
|
|
valkey:
|
|
condition: service_healthy
|
|
migrate:
|
|
condition: service_completed_successfully
|
|
stop_grace_period: 15s # > DRAIN_SECONDS + SHUTDOWN_TIMEOUT_SECONDS (11 s)
|
|
restart: unless-stopped
|
|
networks: [edge, data]
|
|
deploy:
|
|
replicas: ${API_REPLICAS:-2}
|
|
resources:
|
|
limits: {cpus: "${API_CPUS:-1.0}", memory: "${API_MEMORY:-256m}"}
|
|
|
|
lb:
|
|
image: haproxy:3.2.26-alpine@sha256:254e6370f01cd654e0744edf607c67f33777c1b6a9dc0d4b28bcba7171460122
|
|
volumes:
|
|
- ./infra/haproxy/haproxy.cfg:/usr/local/etc/haproxy/haproxy.cfg:ro
|
|
ports:
|
|
# The ONLY public entry point, on the loopback interfaces only. Both IPv4 and IPv6:
|
|
# on Windows `localhost` resolves to ::1 first, and a refused ::1 connection costs ~2 s
|
|
# per request before the fallback to 127.0.0.1 (measured: 2.0 s vs 0.007 s).
|
|
- "127.0.0.1:${APP_PUBLISH_PORT:-8080}:8080"
|
|
- "[::1]:${APP_PUBLISH_PORT:-8080}:8080"
|
|
- "127.0.0.1:${LB_STATS_PORT:-8404}:8404" # local status page (UP/DOWN of replicas)
|
|
healthcheck:
|
|
# Config is valid and the process answers; whether any API is ready is the LB's own job.
|
|
test: ["CMD-SHELL", "haproxy -c -q -f /usr/local/etc/haproxy/haproxy.cfg && wget -q -O /dev/null http://127.0.0.1:8404/"]
|
|
interval: 5s
|
|
timeout: 3s
|
|
retries: 3
|
|
depends_on:
|
|
api:
|
|
condition: service_healthy
|
|
restart: unless-stopped
|
|
networks: [edge]
|
|
deploy:
|
|
resources:
|
|
limits: {cpus: "${LB_CPUS:-1.0}", memory: 128m}
|
|
|
|
networks:
|
|
edge:
|
|
data:
|
|
internal: true # no route outside: stores talk only to api/migrate
|
|
|
|
volumes:
|
|
pgdata:
|
|
valkeydata:
|
|
|
|
secrets:
|
|
postgres_password:
|
|
file: ./secrets/postgres_password
|
|
valkey_password:
|
|
file: ./secrets/valkey_password
|
|
database_url:
|
|
file: ./secrets/database_url
|
|
cache_url:
|
|
file: ./secrets/cache_url
|