NASP/argus
50
0
Fork 0
Code Issues 13 Pull Requests Actions Packages Projects 1 Releases Wiki Activity

[#23] 增加环境变量指定dev/user/instance，优先从环境变量，如果没提供再从hostname解析

Browse Source
This commit is contained in:
yuyr 2025-10-16 15:53:47 +08:00
parent 83b53dac7f
commit 22f40aaafc
15 changed files with 375 additions and 50 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/agent/README.md
View File

@ -34,6 +34,15 @@ Agent 不再依赖配置文件；所有参数均由环境变量与主机名推
| `MASTER_ENDPOINT` | 是 | N/A | Master 基础地址，可写 `http://host:3000` 或 `host:3000`（自动补全 `http://`）。 |
| `REPORT_INTERVAL_SECONDS` | 否 | `60` | 状态上报间隔（秒）。必须为正整数。 |
| `AGENT_HOSTNAME` | 否 | `$(hostname)` | 覆盖容器内主机名，便于测试或特殊命名需求。 |
| `AGENT_ENV` | 否 | 来源于主机名 | 运行环境标识（如 `dev`、`prod`）。与 `AGENT_USER`、`AGENT_INSTANCE` 必须同时设置。 |
| `AGENT_USER` | 否 | 来源于主机名 | 归属用户或团队标识。与 `AGENT_ENV`、`AGENT_INSTANCE` 必须同时设置。 |
| `AGENT_INSTANCE` | 否 | 来源于主机名 | 实例编号或别名。与 `AGENT_ENV`、`AGENT_USER` 必须同时设置。 |
主机名与元数据的解析优先级：
1. 若设置 `AGENT_ENV` / `AGENT_USER` / `AGENT_INSTANCE` 且全部存在，则直接使用这些值。
2. 否则按历史约定从主机名解析 `env-user-instance` 前缀。
3. 如果两者都无法得到完整结果，Agent 启动会失败并提示需要提供上述环境变量。
派生路径：

7
src/agent/app/collector.py
View File

@ -18,13 +18,12 @@ _HOSTNAME_PATTERN = re.compile(r"^([^-]+)-([^-]+)-([^-]+)-.*$")
def collect_metadata(config: AgentConfig) -> Dict[str, Any]:
"""汇总节点注册需要的静态信息。"""
hostname = config.hostname
env, user, instance = _parse_hostname(hostname)
meta = {
"hostname": hostname,
"ip": _detect_ip_address(),
"env": env,
"user": user,
"instance": instance,
"env": config.environment,
"user": config.user,
"instance": config.instance,
"cpu_number": _detect_cpu_count(),
"memory_in_bytes": _detect_memory_bytes(),
"gpu_number": _detect_gpu_count(),

47
src/agent/app/config.py
View File

@ -14,6 +14,9 @@ DEFAULT_REPORT_INTERVAL_SECONDS: Final[int] = 60
@dataclass(frozen=True)
class AgentConfig:
hostname: str
environment: str
user: str
instance: str
node_file: str
version: str
master_endpoint: str
@ -47,10 +50,51 @@ def _resolve_hostname() -> str:
return os.environ.get("AGENT_HOSTNAME") or socket.gethostname()
def _resolve_metadata_fields(hostname: str) -> tuple[str, str, str]:
env = os.environ.get("AGENT_ENV")
user = os.environ.get("AGENT_USER")
instance = os.environ.get("AGENT_INSTANCE")
if env and user and instance:
return env, user, instance
if any([env, user, instance]):
LOGGER = None
try:
from .log import get_logger
LOGGER = get_logger("argus.agent.config")
except Exception: # pragma: no cover - defensive
LOGGER = None
if LOGGER is not None:
LOGGER.warning(
"Incomplete metadata environment variables; falling back to hostname parsing",
extra={
"has_env": bool(env),
"has_user": bool(user),
"has_instance": bool(instance),
},
)
env = user = instance = None
from .collector import _parse_hostname # Local import to avoid circular dependency
env, user, instance = _parse_hostname(hostname)
if not all([env, user, instance]):
raise ValueError(
"Failed to determine metadata fields; set AGENT_ENV/USER/INSTANCE or use supported hostname pattern"
)
return env, user, instance
def load_config() -> AgentConfig:
"""从环境变量推导配置，移除了外部配置文件依赖。"""
hostname = _resolve_hostname()
environment, user, instance = _resolve_metadata_fields(hostname)
node_file = f"/private/argus/agent/{hostname}/node.json"
health_dir = f"/private/argus/agent/{hostname}/health/"
@ -66,6 +110,9 @@ def load_config() -> AgentConfig:
return AgentConfig(
hostname=hostname,
environment=environment,
user=user,
instance=instance,
node_file=node_file,
version=VERSION,
master_endpoint=master_endpoint,

BIN
src/agent/dist/argus-agent vendored
View File

Binary file not shown.

0
src/agent/tests/__init__.py Normal file
View File

30
src/agent/tests/docker-compose.yml
View File

@ -60,6 +60,36 @@ services:
ipv4_address: 172.28.0.20
restart: always
agent_env:
image: ubuntu:22.04
container_name: argus-agent-env-e2e
hostname: host_abc
depends_on:
- master
- bind
environment:
- MASTER_ENDPOINT=http://master.argus.com:3000
- REPORT_INTERVAL_SECONDS=2
- AGENT_ENV=prod
- AGENT_USER=ml
- AGENT_INSTANCE=node-3
- AGENT_HOSTNAME=host_abc
- "ARGUS_BUILD_UID=${ARGUS_BUILD_UID:-2133}"
- "ARGUS_BUILD_GID=${ARGUS_BUILD_GID:-2015}"
volumes:
- ./private/argus/agent/host_abc:/private/argus/agent/host_abc
- ./private/argus/agent/host_abc/health:/private/argus/agent/host_abc/health
- ./private/argus/etc:/private/argus/etc
- ../dist/argus-agent:/usr/local/bin/argus-agent:ro
- ./scripts/agent_entrypoint.sh:/usr/local/bin/agent-entrypoint.sh:ro
- ../scripts/agent_deployment_verify.sh:/usr/local/bin/agent_deployment_verify.sh:ro
entrypoint:
- /usr/local/bin/agent-entrypoint.sh
networks:
default:
ipv4_address: 172.28.0.21
restart: always
networks:
default:
driver: bridge

8
src/agent/tests/scripts/00_e2e_test.sh
View File

@ -7,10 +7,10 @@ SCRIPTS=(
"02_up.sh"
"03_wait_and_assert_registration.sh"
"04_write_health_files.sh"
"08_verify_agent.sh"
"05_assert_status_on_master.sh"
"06_restart_agent_and_reregister.sh"
"07_down.sh"
"05_verify_agent.sh"
"06_assert_status_on_master.sh"
"07_restart_agent_and_reregister.sh"
"08_down.sh"
)
for script in "${SCRIPTS[@]}"; do

2
src/agent/tests/scripts/02_up.sh
View File

@ -41,7 +41,7 @@ compose() {
fi
}
docker container rm -f argus-agent-e2e argus-master-agent-e2e argus-bind-agent-e2e >/dev/null 2>&1 || true
docker container rm -f argus-agent-e2e argus-agent-env-e2e argus-master-agent-e2e argus-bind-agent-e2e >/dev/null 2>&1 || true
docker network rm tests_default >/dev/null 2>&1 || true

59
src/agent/tests/scripts/03_wait_and_assert_registration.sh
View File

@ -6,11 +6,14 @@ TEST_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
TMP_ROOT="$TEST_ROOT/tmp"
API_BASE="http://localhost:32300/api/v1/master"
AGENT_HOSTNAME="dev-e2euser-e2einst-pod-0"
ENV_AGENT_HOSTNAME="host_abc"
NODE_FILE="$TEST_ROOT/private/argus/agent/$AGENT_HOSTNAME/node.json"
ENV_NODE_FILE="$TEST_ROOT/private/argus/agent/$ENV_AGENT_HOSTNAME/node.json"
mkdir -p "$TMP_ROOT"
node_id=""
primary_node_id=""
env_node_id=""
for _ in {1..30}; do
sleep 2
response=$(curl -sS "$API_BASE/nodes" || true)
@ -19,24 +22,49 @@ for _ in {1..30}; do
fi
list_file="$TMP_ROOT/nodes_list.json"
echo "$response" > "$list_file"
node_id=$(python3 - "$list_file" <<'PY'
readarray -t node_ids < <(python3 - "$list_file" "$AGENT_HOSTNAME" "$ENV_AGENT_HOSTNAME" <<'PY'
import json, sys
with open(sys.argv[1]) as handle:
nodes = json.load(handle)
print(nodes[0]["id"] if nodes else "")
target_primary = sys.argv[2]
target_env = sys.argv[3]
primary_id = ""
env_id = ""
for node in nodes:
if node.get("name") == target_primary:
primary_id = node.get("id", "")
if node.get("name") == target_env:
env_id = node.get("id", "")
print(primary_id)
print(env_id)
PY
)
if [[ -n "$node_id" ]]; then
primary_node_id="${node_ids[0]}"
env_node_id="${node_ids[1]}"
if [[ -n "$primary_node_id" && -n "$env_node_id" ]]; then
break
fi
done
if [[ -z "$node_id" ]]; then
echo "[ERROR] Agent did not register within timeout" >&2
if [[ -z "$primary_node_id" ]]; then
echo "[ERROR] Primary agent did not register within timeout" >&2
exit 1
fi
echo "$node_id" > "$TMP_ROOT/node_id"
if [[ -z "$env_node_id" ]]; then
echo "[ERROR] Env-variable agent did not register within timeout" >&2
exit 1
fi
echo "$primary_node_id" > "$TMP_ROOT/node_id"
echo "$env_node_id" > "$TMP_ROOT/node_id_host_abc"
if [[ ! -f "$NODE_FILE" ]]; then
echo "[ERROR] node.json not created at $NODE_FILE" >&2
@ -50,8 +78,20 @@ with open(sys.argv[1]) as handle:
assert "id" in node and node["id"], "node.json missing id"
PY
if [[ ! -f "$ENV_NODE_FILE" ]]; then
echo "[ERROR] node.json not created at $ENV_NODE_FILE" >&2
exit 1
fi
python3 - "$ENV_NODE_FILE" <<'PY'
import json, sys
with open(sys.argv[1]) as handle:
node = json.load(handle)
assert "id" in node and node["id"], "env agent node.json missing id"
PY
detail_file="$TMP_ROOT/initial_detail.json"
curl -sS "$API_BASE/nodes/$node_id" -o "$detail_file"
curl -sS "$API_BASE/nodes/$primary_node_id" -o "$detail_file"
python3 - "$detail_file" "$TMP_ROOT/initial_ip" <<'PY'
import json, sys, pathlib
with open(sys.argv[1]) as handle:
@ -62,4 +102,5 @@ if not ip:
pathlib.Path(sys.argv[2]).write_text(ip)
PY
echo "[INFO] Agent registered with node id $node_id"
echo "[INFO] Agent registered with node id $primary_node_id"
echo "[INFO] Env-variable agent registered with node id $env_node_id"

60
src/agent/tests/scripts/05_verify_agent.sh Executable file
View File

@ -0,0 +1,60 @@
#!/usr/bin/env bash
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
TEST_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
REPO_ROOT="$(cd "$TEST_ROOT/.." && pwd)"
VERIFY_SCRIPT="$REPO_ROOT/scripts/agent_deployment_verify.sh"
ENV_NODE_ID_FILE="$TEST_ROOT/tmp/node_id_host_abc"
PRIMARY_CONTAINER="argus-agent-e2e"
ENV_CONTAINER="argus-agent-env-e2e"
PRIMARY_HOSTNAME="dev-e2euser-e2einst-pod-0"
ENV_HOSTNAME="host_abc"
if ! docker ps --format '{{.Names}}' | grep -q "^${PRIMARY_CONTAINER}$"; then
echo "[WARN] agent container not running; skip verification"
exit 0
fi
if docker exec -i "$PRIMARY_CONTAINER" bash -lc 'command -v curl >/dev/null && command -v jq >/dev/null'; then
echo "[INFO] curl/jq already installed in agent container"
else
echo "[INFO] Installing curl/jq in agent container"
docker exec -i "$PRIMARY_CONTAINER" bash -lc 'apt-get update >/dev/null 2>&1 && apt-get install -y curl jq >/dev/null 2>&1' || true
fi
if [[ ! -f "$VERIFY_SCRIPT" ]]; then
echo "[ERROR] Verification script missing at $VERIFY_SCRIPT" >&2
exit 1
fi
run_verifier() {
local container="$1" hostname="$2"
if ! docker ps --format '{{.Names}}' | grep -q "^${container}$"; then
echo "[WARN] container $container not running; skip"
return
fi
if ! docker exec -i "$container" bash -lc 'command -v /usr/local/bin/agent_deployment_verify.sh >/dev/null'; then
echo "[ERROR] /usr/local/bin/agent_deployment_verify.sh missing in $container" >&2
exit 1
fi
echo "[INFO] Running verification for $hostname in $container"
docker exec -i "$container" env VERIFY_HOSTNAME="$hostname" /usr/local/bin/agent_deployment_verify.sh
}
run_verifier "$PRIMARY_CONTAINER" "$PRIMARY_HOSTNAME"
if docker ps --format '{{.Names}}' | grep -q "^${ENV_CONTAINER}$"; then
if docker exec -i "$ENV_CONTAINER" bash -lc 'command -v curl >/dev/null && command -v jq >/dev/null'; then
echo "[INFO] curl/jq already installed in env agent container"
else
echo "[INFO] Installing curl/jq in env agent container"
docker exec -i "$ENV_CONTAINER" bash -lc 'apt-get update >/dev/null 2>&1 && apt-get install -y curl jq >/dev/null 2>&1' || true
fi
run_verifier "$ENV_CONTAINER" "$ENV_HOSTNAME"
else
echo "[WARN] env-driven agent container not running; skip secondary verification"
fi

33
src/agent/tests/scripts/05_assert_status_on_master.sh → src/agent/tests/scripts/06_assert_status_on_master.sh
View File

@ -6,6 +6,8 @@ TEST_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
TMP_ROOT="$TEST_ROOT/tmp"
API_BASE="http://localhost:32300/api/v1/master"
NODE_ID="$(cat "$TMP_ROOT/node_id")"
ENV_NODE_ID="$(cat "$TMP_ROOT/node_id_host_abc")"
ENV_HOSTNAME="host_abc"
NODES_JSON="$TEST_ROOT/private/argus/metric/prometheus/nodes.json"
success=false
@ -41,13 +43,36 @@ if [[ ! -f "$NODES_JSON" ]]; then
exit 1
fi
python3 - "$NODES_JSON" <<'PY'
python3 - "$NODES_JSON" "$NODE_ID" "$ENV_NODE_ID" <<'PY'
import json, sys
with open(sys.argv[1]) as handle:
nodes = json.load(handle)
assert len(nodes) == 1, nodes
entry = nodes[0]
assert entry["node_id"], entry
expected_primary = sys.argv[2]
expected_env = sys.argv[3]
ids = {entry.get("node_id") for entry in nodes}
assert expected_primary in ids, nodes
assert expected_env in ids, nodes
assert len(nodes) >= 2, nodes
PY
echo "[INFO] Master reflects agent health and nodes.json entries"
env_detail_file="$TMP_ROOT/env_agent_detail.json"
curl -sS "$API_BASE/nodes/$ENV_NODE_ID" -o "$env_detail_file"
python3 - "$env_detail_file" "$ENV_HOSTNAME" <<'PY'
import json, sys
with open(sys.argv[1]) as handle:
node = json.load(handle)
expected_name = sys.argv[2]
assert node.get("name") == expected_name, node
meta = node.get("meta_data", {})
assert meta.get("env") == "prod", meta
assert meta.get("user") == "ml", meta
assert meta.get("instance") == "node-3", meta
PY
echo "[INFO] Env-variable agent reports expected metadata"

7
src/agent/tests/scripts/06_restart_agent_and_reregister.sh → src/agent/tests/scripts/07_restart_agent_and_reregister.sh
View File

@ -10,6 +10,7 @@ AGENT_HOSTNAME="dev-e2euser-e2einst-pod-0"
NETWORK_NAME="tests_default"
NEW_AGENT_IP="172.28.0.200"
ENTRYPOINT_SCRIPT="$SCRIPT_DIR/agent_entrypoint.sh"
VERIFY_SCRIPT="$TEST_ROOT/../scripts/agent_deployment_verify.sh"
ENV_FILE="$TEST_ROOT/.env"
# 中文提示：重启场景也需要同样的入口脚本，确保 DNS 注册逻辑一致
@ -18,6 +19,11 @@ if [[ ! -f "$ENTRYPOINT_SCRIPT" ]]; then
exit 1
fi
if [[ ! -f "$VERIFY_SCRIPT" ]]; then
echo "[ERROR] agent verification script missing at $VERIFY_SCRIPT" >&2
exit 1
fi
if [[ ! -f "$TMP_ROOT/agent_binary_path" ]]; then
echo "[ERROR] Agent binary path missing; rerun bootstrap" >&2
exit 1
@ -94,6 +100,7 @@ if ! docker run -d \
-v "$TEST_ROOT/private/argus/etc:/private/argus/etc" \
-v "$AGENT_BINARY:/usr/local/bin/argus-agent:ro" \
-v "$ENTRYPOINT_SCRIPT:/usr/local/bin/agent-entrypoint.sh:ro" \
-v "$VERIFY_SCRIPT:/usr/local/bin/agent_deployment_verify.sh:ro" \
-e MASTER_ENDPOINT=http://master.argus.com:3000 \
-e REPORT_INTERVAL_SECONDS=2 \
-e ARGUS_BUILD_UID="$AGENT_UID" \

2
src/agent/tests/scripts/07_down.sh → src/agent/tests/scripts/08_down.sh
View File

@ -13,7 +13,7 @@ compose() {
fi
}
docker container rm -f argus-agent-e2e >/dev/null 2>&1 || true
docker container rm -f argus-agent-e2e argus-agent-env-e2e >/dev/null 2>&1 || true
pushd "$TEST_ROOT" >/dev/null
compose down --remove-orphans

26
src/agent/tests/scripts/08_verify_agent.sh
View File

@ -1,26 +0,0 @@
#!/usr/bin/env bash
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
TEST_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
VERIFY_SCRIPT="$(cd "$TEST_ROOT/.." && pwd)/scripts/agent_deployment_verify.sh"
if ! docker ps --format '{{.Names}}' | grep -q '^argus-agent-e2e$'; then
echo "[WARN] agent container not running; skip verification"
exit 0
fi
if docker exec -i argus-agent-e2e bash -lc 'command -v curl >/dev/null && command -v jq >/dev/null'; then
echo "[INFO] curl/jq already installed in agent container"
else
echo "[INFO] Installing curl/jq in agent container"
docker exec -i argus-agent-e2e bash -lc 'apt-get update >/dev/null 2>&1 && apt-get install -y curl jq >/dev/null 2>&1' || true
fi
if docker exec -i argus-agent-e2e bash -lc 'command -v /usr/local/bin/agent_deployment_verify.sh >/dev/null'; then
docker exec -i argus-agent-e2e /usr/local/bin/agent_deployment_verify.sh
elif [[ -x "$VERIFY_SCRIPT" ]]; then
docker exec -i argus-agent-e2e "$VERIFY_SCRIPT"
else
echo "[WARN] agent_deployment_verify.sh not found"
fi

133
src/agent/tests/test_config_metadata.py Normal file
View File

@ -0,0 +1,133 @@
from __future__ import annotations
import os
import unittest
from contextlib import contextmanager
from unittest.mock import patch
from app.config import AgentConfig, load_config
@contextmanager
def temp_env(**overrides: str | None):
originals: dict[str, str | None] = {}
try:
for key, value in overrides.items():
originals[key] = os.environ.get(key)
if value is None:
os.environ.pop(key, None)
else:
os.environ[key] = value
yield
finally:
for key, original in originals.items():
if original is None:
os.environ.pop(key, None)
else:
os.environ[key] = original
class LoadConfigMetadataTests(unittest.TestCase):
@patch("app.config.Path.mkdir")
def test_metadata_from_environment_variables(self, mock_mkdir):
with temp_env(
MASTER_ENDPOINT="http://master.local",
AGENT_HOSTNAME="dev-user-one-pod",
AGENT_ENV="prod",
AGENT_USER="ops",
AGENT_INSTANCE="node-1",
):
config = load_config()
self.assertEqual(config.environment, "prod")
self.assertEqual(config.user, "ops")
self.assertEqual(config.instance, "node-1")
mock_mkdir.assert_called()
@patch("app.config.Path.mkdir")
def test_metadata_falls_back_to_hostname(self, mock_mkdir):
with temp_env(
MASTER_ENDPOINT="http://master.local",
AGENT_HOSTNAME="qa-team-abc-pod-2",
AGENT_ENV=None,
AGENT_USER=None,
AGENT_INSTANCE=None,
):
config = load_config()
self.assertEqual(config.environment, "qa")
self.assertEqual(config.user, "team")
self.assertEqual(config.instance, "abc")
mock_mkdir.assert_called()
@patch("app.config.Path.mkdir")
def test_partial_environment_variables_fallback(self, mock_mkdir):
with temp_env(
MASTER_ENDPOINT="http://master.local",
AGENT_HOSTNAME="stage-ml-001-node",
AGENT_ENV="prod",
AGENT_USER=None,
AGENT_INSTANCE=None,
):
config = load_config()
self.assertEqual(config.environment, "stage")
self.assertEqual(config.user, "ml")
self.assertEqual(config.instance, "001")
mock_mkdir.assert_called()
@patch("app.config.Path.mkdir")
def test_invalid_hostname_raises_error(self, mock_mkdir):
with temp_env(
MASTER_ENDPOINT="http://master.local",
AGENT_HOSTNAME="invalidhostname",
AGENT_ENV=None,
AGENT_USER=None,
AGENT_INSTANCE=None,
):
with self.assertRaises(ValueError):
load_config()
mock_mkdir.assert_not_called()
class CollectMetadataTests(unittest.TestCase):
@patch("app.collector._detect_ip_address", return_value="127.0.0.1")
@patch("app.collector._detect_gpu_count", return_value=0)
@patch("app.collector._detect_memory_bytes", return_value=1024)
@patch("app.collector._detect_cpu_count", return_value=8)
def test_collect_metadata_uses_config_fields(
self,
mock_cpu,
mock_memory,
mock_gpu,
mock_ip,
):
config = AgentConfig(
hostname="dev-user-001-pod",
environment="prod",
user="ops",
instance="node-1",
node_file="/tmp/node.json",
version="1.0.0",
master_endpoint="http://master.local",
report_interval_seconds=60,
health_dir="/tmp/health",
)
from app.collector import collect_metadata
metadata = collect_metadata(config)
self.assertEqual(metadata["env"], "prod")
self.assertEqual(metadata["user"], "ops")
self.assertEqual(metadata["instance"], "node-1")
self.assertEqual(metadata["hostname"], "dev-user-001-pod")
self.assertEqual(metadata["ip"], "127.0.0.1")
self.assertEqual(metadata["cpu_number"], 8)
self.assertEqual(metadata["memory_in_bytes"], 1024)
self.assertEqual(metadata["gpu_number"], 0)
if __name__ == "__main__":
unittest.main()