Compare commits

...
Author SHA1 Message Date
hermes 3a68e81c56 fix(hermes_infra/gitea): add dedicated act_runner cache-server
The built-in act_runner cache-server binds to an interface unreachable
from per-job container networks (default container.network: ""), so
ACTIONS_CACHE_URL resolves to 192.168.112.21:37051 and every job burns
~9 minutes on:

    ::warning::Failed to save:   reserveCache failed: connect ETIMEDOUT
    ::warning::Failed to restore: getCacheEntry  failed: connect ETIMEDOUT

Measured on PR #15 in admin/brencher (gitea/actions/runs/16):
each of 3 jobs ≈12 min wall-clock, of which 9 min is dead air waiting
for cache timeouts.

Run a dedicated act_runner cache-server on the same overlay network and
point the runner at it via external_server. secret shared via
ACT_RUNNER_EXTERNAL_SECRET env on both services.

Related: admin/brencher#15 (whose python-package.yml enables cache:'pip'
but currently gets no benefit because the cache backend is unreachable).

Refs: gitea/actions/runs/16 (job #248 logs)
2026-09-03 18:40:10 +00:00
hermes 6848d114a7 chore(hermes_squid): remove brencher-state-proxy
Drop the brencher-state-proxy sidecar from hermes_infra -- it is
no longer needed. The sidecar existed to relay /state from
brencher.rudolf.keenetic.link into the admin overlay; that relay is
unnecessary since brencher ingress handles the path itself.

Changes vs the brencher-state-proxy commit that was merged in #4:
- delete hermes_squid/brencher-state-proxy.conf
- remove the brencher-state-proxy: service block from
  hermes_squid/stack-compose.yml
- remove the brencher_state_proxy_conf: top-level config object
- drop brencher-state-proxy from the NO_PROXY / no_proxy lists
- update the surrounding comments
2026-08-29 22:47:36 +00:00
4 changed files with 35 additions and 165 deletions
-97
View File
@@ -1,97 +0,0 @@
# nginx config for the brencher-state-proxy sidecar.
#
# Path-locked reverse proxy for the brencher /state endpoint:
# - /state -> proxied to https://brencher.rudolf.keenetic.link/state
# - everything else -> 403 Forbidden
#
# This lets hermes-agent reach /state without going through squid's
# host-level allowlist (which can't see inside TLS) and without exposing
# the rest of brencher's surface area.
#
# Listen on 8080 (plain HTTP). No TLS -- this is a private internal
# API on the admin overlay, reachable only from inside the swarm. No
# host port is published.
#
# Upstream is the public URL on purpose -- the sidecar reaches the public
# brencher.rudolf.keenetic.link through the nginx-ingress overlay the same
# way any external client would, which means it doesn't need to be on the
# brencher container's private networks.
server {
listen 8080;
server_name _;
resolver 127.0.0.11 192.168.1.1 valid=30s ipv6=off;
resolver_timeout 5s;
# Tiny access log; default nginx format
access_log /var/log/nginx/brencher-state-access.log;
error_log /var/log/nginx/brencher-state-error.log warn;
# Healthcheck for swarm -- not /state, deliberately, so a 200 here
# doesn't accidentally imply /state is reachable from outside the
# allowlist.
location = /healthz {
access_log off;
return 200 "ok\n";
add_header Content-Type text/plain;
}
# The only allowed path. Exact match: query strings or trailing
# anything else fall through to the default 403.
location = /state {
set $brencher_upstream https://brencher.rudolf.keenetic.link;
proxy_pass $brencher_upstream/state;
proxy_http_version 1.1;
# Upstream is HTTPS; the sidecar trusts the LE-issued cert via
# the system CA bundle. SNI explicitly named so any future
# transparent-proxy middlebox doesn't break the handshake.
proxy_ssl_server_name on;
proxy_ssl_name brencher.rudolf.keenetic.link;
proxy_set_header Host brencher.rudolf.keenetic.link;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto https;
proxy_connect_timeout 5s;
proxy_read_timeout 30s;
proxy_send_timeout 10s;
# Don't cache -- state should be live.
proxy_buffering off;
add_header Cache-Control "no-store" always;
}
# Secondary brencher (runs on the docker host via host.docker.internal,
# reached as http://host.docker.internal:5002 -- same trick the brencher
# container itself uses in docker-compose.yml's extra_hosts). Only the
# /state2 path is exposed; everything else returns 403 like /state.
location = /state2 {
proxy_pass http://host.docker.internal:5002/state;
proxy_http_version 1.1;
# Cleartext HTTP upstream (port 5002), no TLS to wrangle.
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto https;
proxy_connect_timeout 5s;
proxy_read_timeout 30s;
proxy_send_timeout 10s;
proxy_buffering off;
add_header Cache-Control "no-store" always;
}
# Catch-all: deny everything else, including /state/ (with trailing
# slash), /state?foo=bar (with query), and any other path.
location / {
access_log off;
return 403 "forbidden: only /state is allowed through this proxy\n";
add_header Content-Type text/plain;
}
}
+11 -49
View File
@@ -64,10 +64,10 @@ services:
command: gateway run
user: "0:0"
networks:
# No admin overlay membership here: the sidecars (brencher-state-proxy,
# docker_socket_proxy) are on hermes_proxy too, so DNS resolves
# without expanding hermes-gateway's L3 footprint. The NO_PROXY
# list below ensures those requests skip squid.
# No admin overlay membership here: the docker_socket_proxy sidecar
# is on hermes_proxy too, so DNS resolves without expanding
# hermes-gateway's L3 footprint. The NO_PROXY list below ensures
# those requests skip squid.
- hermes_proxy
volumes:
- hermes_gateway_data:/opt/data
@@ -84,21 +84,21 @@ services:
HTTPS_PROXY: http://squid:3128
http_proxy: http://squid:3128
https_proxy: http://squid:3128
# The sidecars added in this PR (brencher-state-proxy, docker_socket_proxy)
# are reachable on the hermes_proxy overlay. They are NOT public services, so
# traffic to them MUST NOT be routed through squid (which can't reach
# them anyway and would 403 the request).
# The docker_socket_proxy sidecar is reachable on the hermes_proxy
# overlay. It is NOT a public service, so traffic to it MUST NOT be
# routed through squid (which can't reach it anyway and would 403
# the request).
# Comma-separated list. Format accepted by curl/requests/urllib:
# bare hostnames, IPs, CIDR ranges. Order doesn't matter; case-insensitive.
# 127.0.0.1 - loopback
# .local - mDNS / link-local (covers host.docker.internal)
# squid, hermes-gateway, hermes-agent - same-overlay services that
# hermes-gateway itself needs to talk to without going through squid
# brencher-state-proxy, docker_socket_proxy - the new sidecars
# docker_socket_proxy - the sidecar
# 10.0.0.0/8 - the swarm's typical overlay subnet (defensive: covers
# any IP-direct call that bypasses DNS)
NO_PROXY: localhost,127.0.0.1,.local,squid,hermes-gateway,hermes-agent,brencher-state-proxy,docker_socket_proxy,10.0.0.0/8
no_proxy: localhost,127.0.0.1,.local,squid,hermes-gateway,hermes-agent,brencher-state-proxy,docker_socket_proxy,10.0.0.0/8
NO_PROXY: localhost,127.0.0.1,.local,squid,hermes-gateway,hermes-agent,docker_socket_proxy,10.0.0.0/8
no_proxy: localhost,127.0.0.1,.local,squid,hermes-gateway,hermes-agent,docker_socket_proxy,10.0.0.0/8
depends_on:
- squid
deploy:
@@ -193,42 +193,6 @@ services:
delay: 5s
max_attempts: 3
brencher-state-proxy:
# Path-locked reverse proxy for brencher.rudolf.keenetic.link/state.
# Lets hermes-agent reach /state without going through squid (which
# can't see inside TLS) and without exposing the rest of brencher.
# See hermes_squid/brencher-state-proxy.conf for the path rules.
image: nginx:alpine
hostname: brencher-state-proxy
extra_hosts:
# Lets the sidecar reach the secondary brencher at
# http://host.docker.internal:5002/state -- same trick the brencher
# container itself uses in its own docker-compose.yml.
- "host.docker.internal:host-gateway"
networks:
# hermes_proxy: same as docker_socket_proxy. hermes-gateway is
# already on this overlay, so DNS works. Egress is allowed despite
# the `internal: true` flag (which only blocks external ingress,
# not outbound traffic). The sidecar's upstream
# proxy_pass https://brencher.rudolf.keenetic.link/state is a
# public-internet HTTPS call, not a service-name lookup, so
# nginx_ingress membership is not required.
- hermes_proxy
configs:
- source: brencher_state_proxy_conf
target: /etc/nginx/conf.d/brencher-state-proxy.conf
deploy:
mode: replicated
replicas: 1
placement:
constraints:
- node.hostname == odroid # must run where the secret is mounted (and where the public ingress routes back to)
restart_policy:
condition: on-failure
delay: 5s
max_attempts: 3
networks:
hermes_proxy:
driver: overlay
@@ -247,8 +211,6 @@ configs:
file: ./squid.conf
nginx_hermes_conf:
file: ./nginx-hermes.conf
brencher_state_proxy_conf:
file: ./brencher-state-proxy.conf
volumes:
squid_etc:
+11 -17
View File
@@ -73,24 +73,18 @@ cache:
# Enable cache server to use actions/cache.
enabled: true
# The directory to store the cache data.
# If it's empty, the cache data will be stored in $HOME/.cache/actcache.
dir: ""
# The host of the cache server.
# It's not for the address to listen, but the address to connect from job containers.
# So 0.0.0.0 is a bad choice, leave it empty to detect automatically.
host: ""
# The port of the cache server.
# 0 means to use a random available port.
port: 0
# The external cache server URL. Valid only when enable is true.
# If it's specified, act_runner will use this URL as the ACTIONS_CACHE_URL rather than start a server by itself.
# The URL should generally end with "/".
# Requires external_secret below to be set to the same value on both this runner and the cache-server.
external_server: ""
# Shared secret between this runner and the external `act_runner cache-server`. Required when external_server
# (or `act_runner cache-server`) is in use: the runner pre-registers each job's ACTIONS_RUNTIME_TOKEN with the
# cache-server, and the cache-server enforces bearer auth + per-repo cache isolation.
external_secret: ""
# Point at the dedicated gitea_cache_server container on the `gitea` overlay.
# The runner's built-in cache server binds to an interface unreachable from
# per-job container networks, causing ~9 min/job wasted on
# "reserveCache failed: connect ETIMEDOUT" (see gitea/actions/runs/16 job #248).
host: "gitea_cache_server"
port: 37051
# Trailing slash required.
external_server: "http://gitea_cache_server:37051/"
# Shared with the cache-server via ACT_RUNNER_EXTERNAL_SECRET env on both
# services. Rotate via Docker secret; do not commit a real value.
external_secret: "CHANGE_ME_SHARED_SECRET"
container:
# Specifies the network to which the container will connect.
+13 -2
View File
@@ -1,4 +1,3 @@
services:
server:
image: docker.gitea.com/gitea:1.27.0
@@ -26,6 +25,7 @@ services:
GITEA_RUNNER_NAME: "runner1"
GITEA_RUNNER_LABELS: "runner_label1"
# GITEA_RUNNER_EPHEMERAL: "1"
ACT_RUNNER_EXTERNAL_SECRET: "CHANGE_ME_SHARED_SECRET"
networks:
- gitea
# - nginx_ingress
@@ -35,6 +35,16 @@ services:
volumes:
- /var/run/docker.sock:/var/run/docker.sock
cache_server:
image: docker.io/gitea/act_runner:nightly
command: ["cache-server"]
environment:
ACT_RUNNER_EXTERNAL_SECRET: "CHANGE_ME_SHARED_SECRET"
networks:
- gitea
volumes:
- cache_data:/root/.cache/actcache
networks:
nginx_ingress:
external: true
@@ -52,4 +62,5 @@ volumes:
type: nfs
o: addr=192.168.1.96
device: :/root/volumes/gitea
# device: :/media/DATA/runtime/gitea
# device: :/media/DATA/runtime/gitea
cache_data: