Files
knowledge-governance/internal/librarystack/stack.go
T
姚章浩 a42b101a7b
build / build (push) Successful in 1m25s
fix(library): 基线崩溃续作与维护原状态保留
父审发现三处会让目标卡死或状态对不上实际入口,逐条修:

1. 登记成功但没来得及收尾:崩溃重入会撞上"已有生效版本"直接失败,
   目标留在维护里、占用也不释放。现在按「本作业+同一快照」识别为自己的
   版本,只做收尾(恢复服务/开入口/释放两把占用),不重新捕获、不重写;
   别人的版本仍然拒绝。捕获完成但登记失败则明确失败并写"未对目标写入"。
2. 重入覆盖原运行状态:崩溃时服务已被停住,再读一次只会读到 stopped,
   写回去会让恢复阶段永远不启动写入者。AcquireInstance/LibraryMaintenance
   在 owner 相同的重入时保留已持久化原状态;新作业接管已释放窗口用自己的观察。
   真实 PostgreSQL 集成用例覆盖(先红后绿)。
3. 版本登记移进维护窗口:捕获/校验/登记要么一起成立、要么目标原样恢复;
   记录失败时窗口已开放,治理状态同步释放,不再"显示维护中、实际已开"。
   索引映射读不到时拒绝登记,不再用空摘要写证据。

顺带:
- 基线窗口统一收尾抽成 settleBaselineWindow,四种结局各有测试(成功/健康失败/
  held 保持维护/已恢复则释放),并用真实 DB 用例证明 baseline 与 sync/restore
  共用写互斥、幂等重试不产生第二个作业。
- 新增 cmd/librarydrill:维护隔离 → 整库预备份 → 一致性校验 → canary 漂移检测
  → 按预备份恢复 → 逐类复核 → 按原状态恢复服务。用与 worker 相同的 LIBRARY_*
  配置,绝不触碰封存源。
- 目标/预备份/维护栈装配收成 internal/librarystack 一份(worker 与演练共用)。

验证:go build/vet/test 全过;TEST_DATABASE_URL 指向临时 PostgreSQL 跑
TestMaintenanceReentryKeepsOriginalServiceState(修复前失败:got app=stopped)
与 TestLibraryWriteMutexCoversBaselineJobs;本地门禁 15/15。
2026-10-29 19:08:58 +08:00

202 lines
8.2 KiB
Go

// Package librarystack 是"目标整库读写栈"的唯一装配点:同步、恢复、基线,
// 以及主机侧的真实恢复演练都从这里拿同一个目标/预备份存储/维护协调器。
//
// 为什么单独成包:装配一旦被复制两份,"目标是谁、用哪份契约"就会随调用路径漂移。
// 这个包只做装配与配置校验,不包含任何业务编排。
package librarystack
import (
"context"
"errors"
"fmt"
"strings"
"knowledge-governance/internal/appconfig"
"knowledge-governance/internal/librarymaintenance"
"knowledge-governance/internal/librarysnapshot"
v071 "knowledge-governance/internal/releaseadapter/v071"
)
// OpenMaintenance 按部署配置装配实例级维护目标;未接线时返回明确原因。
func OpenMaintenance(config appconfig.Config) (*librarymaintenance.Manager, error) {
database, err := librarymaintenance.DatabaseIdentity(config.LibraryProductionDatabaseURL)
if err != nil {
return nil, err
}
return librarymaintenance.New(librarymaintenance.Config{
Instance: config.LibraryTargetInstance,
Project: config.LibraryTargetProject,
Writers: config.LibraryTargetWriters,
NonWriters: config.LibraryTargetNonWriters,
DockerSocket: config.LibraryTargetDockerSocket,
IngressURL: config.LibraryTargetIngressURL,
IngressService: config.LibraryTargetIngressService,
IngressDrainFile: config.LibraryTargetIngressDrainFile,
Database: database,
Index: config.LibraryProductionIndex,
Bucket: config.LibraryProductionS3Bucket,
DatabaseURL: config.LibraryProductionDatabaseURL,
RedisAddr: config.LibraryTargetRedisAddr,
RedisPassword: config.LibraryTargetRedisPassword,
RedisDB: config.LibraryTargetRedisDB,
})
}
// scopeConfig 是"某个 scope 的生产整库目标"的部署注入配置。
type scopeConfig struct {
target v071.ProductionScope
s3 v071.S3ObjectStoreConfig
}
func scopeConfigOf(config appconfig.Config, scope string) (scopeConfig, error) {
missing := make([]string, 0, 5)
if config.LibraryProductionDatabaseURL == "" {
missing = append(missing, "LIBRARY_PRODUCTION_DATABASE_URL")
}
if config.LibraryProductionElasticURL == "" {
missing = append(missing, "LIBRARY_PRODUCTION_ELASTIC_URL")
}
if config.LibraryProductionTenantID == 0 {
missing = append(missing, "LIBRARY_PRODUCTION_TENANT_ID")
}
kb := config.LibraryProductionInternalKB
if scope == "external" {
kb = config.LibraryProductionExternalKB
}
if kb == "" {
missing = append(missing, "LIBRARY_PRODUCTION_KB_"+strings.ToUpper(scope))
}
if config.LibraryProductionS3Endpoint == "" || config.LibraryProductionS3Bucket == "" ||
config.LibraryProductionS3AccessKey == "" || config.LibraryProductionS3SecretKey == "" {
missing = append(missing, "LIBRARY_PRODUCTION_S3_ENDPOINT/BUCKET/ACCESS_KEY/SECRET_KEY")
}
if len(missing) > 0 {
return scopeConfig{}, fmt.Errorf("生产整库目标未配置:缺少 %s", strings.Join(missing, "、"))
}
return scopeConfig{
target: v071.ProductionScope{TenantID: config.LibraryProductionTenantID, KnowledgeBaseID: kb},
s3: v071.S3ObjectStoreConfig{
Provider: config.LibraryProductionS3Provider, Endpoint: config.LibraryProductionS3Endpoint,
Region: config.LibraryProductionS3Region, Bucket: config.LibraryProductionS3Bucket,
AccessKeyID: config.LibraryProductionS3AccessKey, SecretAccessKey: config.LibraryProductionS3SecretKey,
Secure: strings.HasPrefix(config.LibraryProductionS3Endpoint, "https://"),
ForcePathStyle: true, MaxObjectBytes: 256 * 1024 * 1024,
},
}, nil
}
func newTarget(ctx context.Context, config appconfig.Config, scope scopeConfig) (*v071.ProductionLibraryTarget, error) {
objects, err := v071.NewS3ObjectStore(scope.s3)
if err != nil {
return nil, errors.New("生产对象存储配置无效")
}
return v071.NewProductionLibraryTarget(ctx, v071.ProductionLibraryConfig{
Postgres: config.LibraryProductionDatabaseURL, SearchURL: config.LibraryProductionElasticURL,
SearchIndex: config.LibraryProductionIndex, Objects: objects, ObjectsReader: objects,
ArchiveRoot: config.LibraryArchiveRoot,
ObjectsBucket: config.LibraryProductionS3Bucket,
MaxDocuments: 200000,
}, scope.target)
}
func newBackups(config appconfig.Config, scope string, target *v071.ProductionLibraryTarget,
contract librarysnapshot.Compatibility) (*v071.ProductionBackupStore, error) {
return v071.NewProductionBackupStore(v071.ProductionBackupConfig{
Pool: target.Pool(), Search: target.Search(), Objects: target.Objects(),
ArchiveRoot: config.LibraryArchiveRoot, Scope: scope, Identity: target.Identity(),
Compatibility: contract, MaxDocuments: 200000, MaxObjectByte: 256 * 1024 * 1024,
})
}
// OpenProductionStack 用部署注入的配置装配目标与独立预备份存储。
// 契约以"要同步的这份快照"为准:生产必须向它看齐,而不是反过来。
func OpenProductionStack(ctx context.Context, config appconfig.Config, scope, snapshotID, knownMapping string) (
*v071.ProductionLibraryTarget, *v071.ProductionBackupStore, error) {
scopeCfg, err := scopeConfigOf(config, scope)
if err != nil {
return nil, nil, err
}
target, err := newTarget(ctx, config, scopeCfg)
if err != nil {
return nil, nil, err
}
archive, err := librarysnapshot.Open(config.LibraryArchiveRoot, 256*1024*1024)
if err != nil {
target.Close()
return nil, nil, fmt.Errorf("归档不可用,无法读取快照契约: %w", err)
}
manifest, err := archive.Load(snapshotID)
archive.Close()
if err != nil {
target.Close()
return nil, nil, fmt.Errorf("读取目标快照失败: %w", err)
}
contract := manifest.Compatibility
if contract.EmbeddingDimension == 0 {
contract.EmbeddingModel = config.LibraryEmbeddingModel
contract.EmbeddingDimension = config.LibraryEmbeddingDimension
}
// 生产索引存在时必须与快照契约一致;不存在或为空基线时跳过比对
// (空索引的映射没有可比性,非空之后仍然严格比对)。
if mapping, mapErr := target.Search().MappingSHA256(ctx); mapErr == nil {
// 映射基线:优先用"上一次生产生效版本登记的实际映射",
// 它才是这份生产索引的既有事实;快照契约只管"来自哪份数据"。
baseline := knownMapping
if baseline == "" {
baseline = mapping
}
if baseline != "" && mapping != baseline {
current, currentErr := target.Current(ctx)
emptyBaseline := currentErr == nil && current.Counts["index"] == 0 && current.Counts["chunks"] == 0
if !emptyBaseline {
target.Close()
return nil, nil, errors.New("生产索引映射与上次生效版本不一致(索引被改动过),先对齐索引映射")
}
}
}
backupContract := contract
if knownMapping != "" {
backupContract.IndexMappingSHA256 = knownMapping
}
backups, err := newBackups(config, scope, target, backupContract)
if err != nil {
target.Close()
return nil, nil, err
}
return target, backups, nil
}
// OpenBaselineStack 装配"给目标自己建基线"所需的目标与捕获器。
//
// 与同步的差别只有契约来源:基线还没有快照,它要捕获的正是目标当前状态,
// 所以契约取目标此刻的真实索引映射,绝不去借一份别人的快照当契约。
func OpenBaselineStack(ctx context.Context, config appconfig.Config, scope string) (
*v071.ProductionLibraryTarget, *v071.ProductionBackupStore, error) {
scopeCfg, err := scopeConfigOf(config, scope)
if err != nil {
return nil, nil, err
}
target, err := newTarget(ctx, config, scopeCfg)
if err != nil {
return nil, nil, err
}
mapping, mapErr := target.Search().MappingSHA256(ctx)
if mapErr != nil {
target.Close()
return nil, nil, fmt.Errorf("读取目标索引映射失败,无法确定基线契约: %w", mapErr)
}
// 只写能证明的字段:embedding 配置来自部署声明,索引映射来自目标此刻的真实映射;
// 目标实例的 WeKnora/schema 版本这里没有可靠来源,就不填。
contract := librarysnapshot.Compatibility{
EmbeddingModel: config.LibraryEmbeddingModel,
EmbeddingDimension: config.LibraryEmbeddingDimension,
IndexMappingSHA256: mapping,
}
backups, err := newBackups(config, scope, target, contract)
if err != nil {
target.Close()
return nil, nil, err
}
return target, backups, nil
}