Files
mtusi2026lab3/infra/haproxy/haproxy.cfg
T
Alexey BodanyukandClaude Opus 5.5 3f12698676 feat: HAProxy и N API-реплик, X-Instance-Id, общая идемпотентность, drain при SIGTERM
- compose: один масштабируемый сервис api (deploy.replicas=2, без ports и
  container_name), HAProxy 3.2 LTS по digest — единственная точка входа
  (127.0.0.1), разовый migrate до старта реплик, CPU/RAM-лимиты у всех сервисов,
  каждому сервису только его secrets.
- HAProxy: server-template + Docker DNS (re-resolve 1 с), активная проверка
  GET /health/ready каждую 1 с (fall 2 / rise 2), без backup-серверов,
  retry только conn-failure, ошибки самого LB — в формате контракта с X-Request-Id.
- API: X-Instance-Id = hostname контейнера на каждом ответе (и на 500/503);
  Idempotency-Key обязателен; ключ, сообщение и результат пишутся в ОДНОЙ
  транзакции (миграция 0002), конкурентный повтор ждёт уникальный индекс;
  другой body — 409, посторонний — всегда 404; attachment_ids=[] равен
  отсутствию поля; дубли ключей JSON и тело у POST /auth/sessions — 400.
- SIGTERM: readiness 503 + DRAIN_SECONDS=3 обслуживания, затем graceful stop.
- scripts: stop-one.sh (+ .cmd для Windows), restore-api.sh, restart.sh
  перезапускает и LB, lbctl.py — состояние бэкендов HAProxy -> контейнеры.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-16 13:32:13 +03:00

90 lines
4.6 KiB
INI

# HAProxy 3.2 LTS — the single public entry point of the lab 3 stand (compose service `lb`).
#
# Backend discovery: Docker DNS name `api` returns one A record per RUNNING replica of the
# compose service `api` (deploy.replicas / --scale). server-template keeps 8 slots and
# re-resolves every second, so scale up/down, a restarted container with a new IP and a
# full stack restart are picked up without restarting the LB. A stopped container drops
# out of DNS and its slot goes to MAINT.
#
# Readiness: an ACTIVE check `GET /health/ready` (PostgreSQL + Valkey reachable, not draining)
# every 1 s; 2 failures -> DOWN, 2 successes -> UP. Passive errors do not decide anything on
# their own, and there are NO backup/fallback servers without checks: a replica that answers
# /health/live but not /health/ready never gets business traffic.
# Worst case detection of a failed replica:
# inter 1 s + fall 2 x (check timeout 1 s) ~= 3 s (frozen process);
# refused connection / 503 ready: inter 1 s + fastinter 0.5 s ~= 1.5 s.
# Requests that could not even be SENT (connection refused/reset before the request)
# are retried on another replica (retry-on conn-failure + redispatch) - safe for POST,
# because the API has not seen them. Nothing that may have reached the API is retried by
# the LB: an uncertain POST is retried by the CLIENT with the same Idempotency-Key.
# No sticky sessions: sessions live in Valkey, every replica serves every user.
global
log stdout format raw local0 info
maxconn 4096
# Runtime API for `show servers state` / `show stat` (scripts and tests, no network port).
stats socket /var/lib/haproxy/admin.sock mode 600 level admin
hard-stop-after 15s
resolvers docker
nameserver dns1 127.0.0.11:53
accepted_payload_size 8192
resolve_retries 3
timeout resolve 1s
timeout retry 1s
hold valid 1s
hold other 1s
hold refused 1s
hold nx 1s
hold timeout 1s
hold obsolete 1s
defaults
mode http
log global
option httplog
option dontlognull
option http-keep-alive
timeout connect 1s
timeout client 30s
timeout server 10s # > DB_TIMEOUT_SECONDS (3 s): the API answers 503 itself first
timeout http-request 10s
timeout http-keep-alive 10s
timeout queue 3s
timeout check 1s
retries 2
retry-on conn-failure
option redispatch 1
# Errors generated by the LB itself, in the contract error format with the same request id.
# They carry no X-Instance-Id: no API replica produced them. 502/504 = the outcome of the
# request is unknown (a replica died mid-request / did not answer in time): retry with the
# same Idempotency-Key. 503 = no ready replica at all.
http-error status 502 content-type application/json lf-string "{\"error\":{\"code\":\"unavailable\",\"message\":\"API-реплика оборвала ответ; повторите с тем же Idempotency-Key\",\"request_id\":\"%[var(txn.request_id)]\"}}" hdr X-Request-Id "%[var(txn.request_id)]"
http-error status 503 content-type application/json lf-string "{\"error\":{\"code\":\"unavailable\",\"message\":\"Нет готовых API-реплик\",\"request_id\":\"%[var(txn.request_id)]\"}}" hdr X-Request-Id "%[var(txn.request_id)]"
http-error status 504 content-type application/json lf-string "{\"error\":{\"code\":\"unavailable\",\"message\":\"API-реплика не ответила вовремя; повторите с тем же Idempotency-Key\",\"request_id\":\"%[var(txn.request_id)]\"}}" hdr X-Request-Id "%[var(txn.request_id)]"
frontend public
bind :8080
# One request id end to end: keep a valid client X-Request-Id, otherwise create one; the
# API echoes it, and errors generated by the LB itself carry the same value.
http-request set-header X-Request-Id %[uuid()] unless { req.hdr(x-request-id) -m reg ^[A-Za-z0-9._:-]{1,128}$ }
http-request set-var(txn.request_id) req.hdr(x-request-id)
log-format "%ci:%cp [%tr] %ft %b/%s %TR/%Tw/%Tc/%Tr/%Ta %ST %B %tsc %ac/%fc/%bc/%sc/%rc %sq/%bq rid=%[var(txn.request_id)] \"%HM %HPO\""
default_backend api
backend api
balance roundrobin
option httpchk
http-check send meth GET uri /health/ready ver HTTP/1.1 hdr Host api
http-check expect status 200
default-server inter 1s fastinter 500ms downinter 1s fall 2 rise 2 check resolvers docker init-addr none resolve-prefer ipv4 maxconn 200
server-template api 8 api:8080
# Local-only status page (127.0.0.1:8404 on the host): which replica is UP/DOWN right now.
frontend stats
bind :8404
stats enable
stats uri /
stats refresh 2s
stats show-legends