Files
LokiStack/backend/services/loki_config.go
T
opsx a32f93f533 feat(deployments,analysis): 保留弹框重设计 + 分析页命名空间筛选 + 过期 chunk 真正删除
=== 一、保留策略弹框重设计(RetentionDialog)===
旧版问题:
1. 主输入是 YAML textarea,用户必须会写 `limits_config.retention_period: 30d`
2. 三个按钮:取消 / 仅保存 / 保存并应用 —— 「仅保存」只改部署记录不重载 Loki,
   今天已经造成一次真实故障:DB 记录 7 天、运行时 ConfigMap 是 24h,配置漂移
3. 「查看完整配置」是独立按钮 + 独立状态,多一次点击
4. 硬编码提示 `retention_delete_delay: 2h`,而实际已改成 10m,属于误导

新版:
- 主输入改为**保留天数数字框**(1–3650),YAML 由前端拼,用户不用写
- 快捷预设 1/7/14/30/90,实时显示「= 720 小时」
- >7 天给出存储成本提示
- 完整运行 config 改成 `<details>` 折叠,展开即自动加载(去掉独立按钮)
- **按钮只剩 2 个:取消 / 保存并应用**
- 删掉硬编码的 compactor 提示块

=== 二、日志分析页加命名空间筛选 ===
根因:analysis 页写死 namespace: '',等价 `{namespace=~".+"}` —— 一次分析扫全集群
(实测约 370 GB/天),5 条 count_over_time 跑 30s 都出不来,页面永远转圈。

- 新增命名空间下拉(从 Loki label values 拉取,localStorage 记忆)
- 默认取业务命名空间 astralx(其次列表首个),不再默认全集群
- 「全部命名空间(较慢)」降级为显式选项
- 命名空间未就绪时先显示 loading,避免先发一次全集群查询
- 5 条查询改**分批(每批 2 条)**执行 + 429 自动重试一次

=== 三、后端并发/超时调整 ===
- 重查询并发上限 2 → 4,且拿不到令牌时**排队最多 10s**再拒绝
  (原来立即 429,导致分析页 5 条并发里有 3 条直接失败)
- maxLogQueryLimit 1000 → 5000,与前端下拉完全对齐(原来前端选 5000 会被静默截断)

=== 四、过期 chunk 真正删除(关键修复)===
实测发现:索引表已从 COS 删除,但 chunk 对象仍残留 21,354 个 / 24.8 GB,
最老的 45 小时前 —— 因为 Loki 默认 `retention_delete_delay` = **2h**,
标记后要等 2 小时才真正删 chunk。

- 配置显式设置 `compactor.retention_delete_delay: 10m`
- 已用 `loki -verify-config` 校验(config is valid)
- 删除前后对比:
    对象数 46,536 → 23,934;占用 53.61 GB → 27.25 GB
    超期对象 21,354 个 / 24.8 GB → **0 个**
    桶内最老对象 45.2h → 24.4h(= 保留期 1 天)
- deployment_config_test 同步调整:retention_delete_delay 从"禁止出现"移入"必须存在"

镜像:lghost/lokistack:latest + v0.1 = sha256:b4d98e016bca
2026-10-28 17:05:57 +00:00

414 lines
11 KiB
Go

// LokiDeployer — loki.yaml 配置生成:模板 + 对象存储/规则存储/ring 拓扑。
package services
import (
"bytes"
"fmt"
"text/template"
"lokistack/pkg/crypto"
)
const lokiConfigTemplate = `auth_enabled: {{ .AuthEnabled }}
common:
path_prefix: /var/lib/loki
{{- if .FilesystemStorage }}
storage:
filesystem:
chunks_directory: /var/lib/loki/chunks
rules_directory: /var/lib/loki/rules
{{- end }}
replication_factor: {{ .ReplicationFactor }}
ring:
instance_addr: {{ .RingInstanceAddr }}
kvstore:
store: {{ .RingKVStore }}
{{- if .StorageConfigBlock }}
storage_config:
{{ .StorageConfigBlock }}
{{- end }}
{{- if .MemberlistBlock }}
{{ .MemberlistBlock }}
{{- end }}
ingester:
wal:
dir: /var/lib/loki/wal
flush_on_shutdown: true
replay_memory_ceiling: 1GB
lifecycler:
ring:
kvstore:
store: {{ .RingKVStore }}
replication_factor: {{ .ReplicationFactor }}
{{- if .MultiReplica }}
autoforget_unhealthy: true
{{- end }}
compactor:
working_directory: /var/lib/loki/compactor
retention_enabled: true
delete_request_store: {{ .CompactorStoreType }}
# 过期 chunk 的「标记 → 真正删除」间隔。默认 2h 会让对象存储里
# 多留 2 小时的数据(实测桶内超过保留期的对象一度占到 24.6 GB)。
# 预发环境取 10m:既保留短暂的安全窗口,又不会长期占用存储成本。
retention_delete_delay: 10m
limits_config:
retention_period: {{ .RetentionPeriod }}h
# 单实例 Loki 的并发护栏:默认 32 高得离谱,一条重查询就能把单进程打满
max_query_parallelism: 8
# ── 查询结果缓存(官方推荐;此前完全没配,重复查询要重算一遍)───────
# align_queries_with_step 让不同时间范围的同类查询对齐到同一 step,
# 显著提高缓存命中率;命中后同样的"日志量统计"直接从缓存返回。
query_range:
align_queries_with_step: true
cache_results: true
results_cache:
cache:
embedded_cache:
enabled: true
max_size_mb: 100
schema_config:
configs:
- from: "2024-01-01"
store: tsdb
object_store: {{ .ObjectStore }}
schema: v13
index:
prefix: tsdb_index_
period: 24h
ruler:
storage:
type: {{ .RulerStoreType }}
{{- if .RulerStorageBlock }}
{{ .RulerStorageBlock }}
{{- end }}
{{- if eq .RulerStoreType "local" }}
local:
directory: /var/lib/loki/rules
{{- end }}
rule_path: /var/lib/loki/rules-temp
ring:
kvstore:
store: {{ .RingKVStore }}
analytics:
reporting_enabled: false
`
type lokiConfigData struct {
AuthEnabled bool
ReplicationFactor int
RetentionPeriod int
StorageConfigBlock string
ObjectStore string
RulerStoreType string
CompactorStoreType string
RingInstanceAddr string // 单副本 127.0.0.1;多副本 0.0.0.0(须可被对端寻址)
RingKVStore string // 单副本 inmemory;多副本 memberlist
MemberlistBlock string // 多副本时非空,注入 memberlist 顶级配置
RulerStorageBlock string // 对象存储时非空,注入 ruler.storage.s3/gcs/azure 嵌套配置
MultiReplica bool
FilesystemStorage bool
}
func (d *LokiDeployer) generateLokiConfig() string {
retentionHours := d.deployment.RetentionDays * 24
if retentionHours <= 0 {
retentionHours = DefaultRetentionDays * 24
}
storageType := d.effectiveStorageType()
storageBlock := d.buildStorageConfigBlock()
objectStore := "filesystem"
rulerStore := "local"
compactorStore := "filesystem"
switch storageType {
case "s3", "oss", "minio", "cos", "obs":
objectStore = "s3"
rulerStore = "s3"
// 删除请求存本地盘,避免各云 S3 兼容实现差异导致 compactor 启动失败
compactorStore = "filesystem"
case "gcs":
objectStore = "gcs"
rulerStore = "gcs"
compactorStore = "gcs"
case "azure":
objectStore = "azure"
rulerStore = "azure"
compactorStore = "azure"
}
// Ring 拓扑:单副本用 inmemory + 127.0.0.1(无需跨实例通信);
// 多副本必须切 memberlist + 0.0.0.0,否则 ring 无法收敛、写入失败。
ringAddr := "127.0.0.1"
ringStore := "inmemory"
memberlistBlock := ""
if d.deployment.Replicas >= 2 {
ringAddr = "0.0.0.0"
ringStore = "memberlist"
ns := d.deployment.Namespace
if ns == "" {
ns = "log"
}
memberlistBlock = fmt.Sprintf(`memberlist:
join_members:
- loki-headless.%s.svc.cluster.local
bind_port: 7946
dead_node_reclaim_time: 30s
gossip_to_dead_nodes_time: 2m
left_ingesters_timeout: 5m`, ns)
}
data := lokiConfigData{
AuthEnabled: d.deployment.AuthEnabled,
ReplicationFactor: d.replicationFactor(),
RetentionPeriod: retentionHours,
StorageConfigBlock: storageBlock,
ObjectStore: objectStore,
RulerStoreType: rulerStore,
CompactorStoreType: compactorStore,
RingInstanceAddr: ringAddr,
RingKVStore: ringStore,
MemberlistBlock: memberlistBlock,
RulerStorageBlock: d.buildRulerStorageBlock(),
MultiReplica: d.deployment.Replicas >= 2,
FilesystemStorage: objectStore == "filesystem",
}
tmpl, err := template.New("loki-config").Parse(lokiConfigTemplate)
if err != nil {
return d.generateLokiConfigFallback(retentionHours)
}
var buf bytes.Buffer
if err := tmpl.Execute(&buf, data); err != nil {
return d.generateLokiConfigFallback(retentionHours)
}
return buf.String()
}
func (d *LokiDeployer) buildStorageConfigBlock() string {
if d.deployment.StorageConfig == nil {
return ""
}
sc := d.deployment.StorageConfig
accessKey, secretKey := d.storageCredentialValues()
switch sc.StorageType {
case "s3":
endpoint := sc.Endpoint
if endpoint == "" {
endpoint = fmt.Sprintf("s3.%s.amazonaws.com", sc.Region)
}
insecure := "false"
if sc.Insecure {
insecure = "true"
}
return fmt.Sprintf(` aws:
bucketnames: %s
endpoint: %s
region: %s
access_key_id: %s
secret_access_key: %s
insecure: %s
s3forcepathstyle: true`, sc.Bucket, endpoint, sc.Region, accessKey, secretKey, insecure)
case "oss", "cos", "obs":
insecure := "false"
if sc.Insecure {
insecure = "true"
}
// 云厂商 S3 兼容端点有两种写法:
// 区域端点 cos.ap-hongkong.myqcloud.com ← 正确
// 虚拟主机式 mybucket-1250000000.cos.<region>... ← 桶名已在 host 里
// Loki 的 S3 客户端会自己把 bucketnames 拼到 endpoint 前面(s3forcepathstyle=false
// 即虚拟主机风格),所以 endpoint 必须是区域端点。若用户在平台里填了带桶名的写法,
// 这里剥掉桶名前缀,否则会拼出 <bucket>.<bucket>.cos... 导致 TLS 证书校验失败、
// 日志写入失败(ruler/ingester 持续报 x509 证书不匹配)。
endpoint := stripBucketPrefix(sc.Endpoint, sc.Bucket)
return fmt.Sprintf(` aws:
bucketnames: %s
endpoint: %s
region: %s
access_key_id: %s
secret_access_key: %s
insecure: %s
s3forcepathstyle: false`, sc.Bucket, endpoint, sc.Region, accessKey, secretKey, insecure)
case "minio":
insecure := "false"
if sc.Insecure {
insecure = "true"
}
return fmt.Sprintf(` aws:
bucketnames: %s
endpoint: %s
access_key_id: %s
secret_access_key: %s
insecure: %s
s3forcepathstyle: true`, sc.Bucket, stripBucketPrefix(sc.Endpoint, sc.Bucket), accessKey, secretKey, insecure)
case "gcs":
return fmt.Sprintf(` gcs:
bucket_name: %s`, sc.Bucket)
case "azure":
return fmt.Sprintf(` azure:
container_name: %s
account_name: %s
account_key: %s`, sc.Bucket, accessKey, secretKey)
default:
return ""
}
}
// buildRulerStorageBlock 生成 ruler.storage 下与存储类型同名的嵌套配置
// (Loki 3.x 的 ruler 存储客户端不回退读取 storage_config,缺少此块会以空 bucket 启动失败)
func (d *LokiDeployer) buildRulerStorageBlock() string {
if d.deployment.StorageConfig == nil {
return ""
}
sc := d.deployment.StorageConfig
accessKey, secretKey := d.storageCredentialValues()
switch sc.StorageType {
case "s3":
endpoint := sc.Endpoint
if endpoint == "" {
endpoint = fmt.Sprintf("s3.%s.amazonaws.com", sc.Region)
}
insecure := "false"
if sc.Insecure {
insecure = "true"
}
return fmt.Sprintf(` s3:
bucketnames: %s
endpoint: %s
region: %s
access_key_id: %s
secret_access_key: %s
insecure: %s
s3forcepathstyle: true`, sc.Bucket, endpoint, sc.Region, accessKey, secretKey, insecure)
case "oss", "cos", "obs":
insecure := "false"
if sc.Insecure {
insecure = "true"
}
return fmt.Sprintf(` s3:
bucketnames: %s
endpoint: %s
region: %s
access_key_id: %s
secret_access_key: %s
insecure: %s
s3forcepathstyle: false`, sc.Bucket, sc.Endpoint, sc.Region, accessKey, secretKey, insecure)
case "minio":
insecure := "false"
if sc.Insecure {
insecure = "true"
}
return fmt.Sprintf(` s3:
bucketnames: %s
endpoint: %s
access_key_id: %s
secret_access_key: %s
insecure: %s
s3forcepathstyle: true`, sc.Bucket, sc.Endpoint, accessKey, secretKey, insecure)
case "gcs":
return fmt.Sprintf(` gcs:
bucket_name: %s`, sc.Bucket)
case "azure":
return fmt.Sprintf(` azure:
container_name: %s
account_name: %s
account_key: %s`, sc.Bucket, accessKey, secretKey)
default:
return ""
}
}
func (d *LokiDeployer) storageCredentialValues() (string, string) {
if d.deployment.StorageConfig == nil {
return "", ""
}
if d.deployment.Mode == "k8s-yaml" || d.deployment.Mode == "host-docker" {
return "${LOKI_STORAGE_ACCESS_KEY}", "${LOKI_STORAGE_SECRET_KEY}"
}
// host-docker 模式没有 Secret 注入,必须把库里 ENC: 密文解密后再写进配置
return crypto.DecryptField(d.deployment.StorageConfig.AccessKey),
crypto.DecryptField(d.deployment.StorageConfig.SecretKey)
}
// generateLokiConfigFallback 当模板解析失败时的降级配置
func (d *LokiDeployer) generateLokiConfigFallback(retentionHours int) string {
return fmt.Sprintf(`auth_enabled: %v
server:
http_listen_port: 3100
grpc_listen_port: 9095
http_server_read_timeout: 3m
http_server_write_timeout: 3m
log_level: info
common:
path_prefix: /var/lib/loki
storage:
filesystem:
chunks_directory: /var/lib/loki/chunks
rules_directory: /var/lib/loki/rules
replication_factor: %d
ring:
instance_addr: 127.0.0.1
kvstore:
store: inmemory
ingester:
wal:
enabled: true
dir: /var/lib/loki/wal
lifecycler:
ring:
kvstore:
store: inmemory
replication_factor: %d
final_sleep: 0s
chunk_idle_period: 30m
max_chunk_age: 2h
querier:
max_concurrent: 10
compactor:
working_directory: /var/lib/loki/compactor
retention_enabled: true
retention_delete_delay: 2h
limits_config:
retention_period: %dh
ingestion_rate_mb: 64
ingestion_burst_size_mb: 128
schema_config:
configs:
- from: "2024-01-01"
store: tsdb
object_store: filesystem
schema: v13
index:
prefix: tsdb_index_
period: 24h
analytics:
reporting_enabled: false
`, d.deployment.AuthEnabled, d.replicationFactor(), d.replicationFactor(), retentionHours)
}
// ==================== 辅助函数 ====================
func (d *LokiDeployer) replicationFactor() int {
if d.deployment.Replicas >= 3 {
return 3
}
if d.deployment.Replicas >= 2 {
return 2
}
return 1
}