237 lines
8.3 KiB
Go
237 lines
8.3 KiB
Go
package service
|
||
|
||
import (
|
||
"context"
|
||
"fmt"
|
||
"log/slog"
|
||
"strings"
|
||
"time"
|
||
|
||
"hyapp/pkg/appcode"
|
||
"hyapp/pkg/logx"
|
||
"hyapp/services/room-service/internal/room/outbox"
|
||
)
|
||
|
||
const (
|
||
// outboxRetryExponentialBackoff 是房间事件补偿的默认策略，避免外部故障时每秒刷同一条错误。
|
||
outboxRetryExponentialBackoff = "exponential_backoff"
|
||
// outboxWorkerName 固定写入日志，便于后续按 worker 聚合检索。
|
||
outboxWorkerName = "room_outbox"
|
||
)
|
||
|
||
// OutboxWorkerOptions 控制 room_outbox pending 事件的持续补偿循环。
|
||
type OutboxWorkerOptions struct {
|
||
// PollInterval 是两轮 pending 扫描之间的间隔；非正数会被归一到安全默认值。
|
||
PollInterval time.Duration
|
||
// BatchSize 是单轮最多处理条数；非正数会被归一，避免全表扫描。
|
||
BatchSize int
|
||
// PublishTimeout 限制单条外部投递耗时，shutdown 时当前事件最多阻塞该时长。
|
||
PublishTimeout time.Duration
|
||
// RetryStrategy 当前只支持 exponential_backoff。
|
||
RetryStrategy string
|
||
// MaxRetryCount 是转入 failed 前允许的最大失败次数。
|
||
MaxRetryCount int
|
||
// InitialBackoff 是首次失败后的等待时间。
|
||
InitialBackoff time.Duration
|
||
// MaxBackoff 是指数退避上限。
|
||
MaxBackoff time.Duration
|
||
}
|
||
|
||
func defaultOutboxWorkerOptions() OutboxWorkerOptions {
|
||
// 默认值优先避免 busy loop、无限重试和日志污染。
|
||
return OutboxWorkerOptions{
|
||
PollInterval: time.Second,
|
||
BatchSize: 100,
|
||
PublishTimeout: 3 * time.Second,
|
||
RetryStrategy: outboxRetryExponentialBackoff,
|
||
MaxRetryCount: 10,
|
||
InitialBackoff: 5 * time.Second,
|
||
MaxBackoff: 5 * time.Minute,
|
||
}
|
||
}
|
||
|
||
func normalizeOutboxWorkerOptions(options OutboxWorkerOptions) OutboxWorkerOptions {
|
||
// worker 层再次归一化，防止测试或手动构造绕过 config.Normalize。
|
||
defaults := defaultOutboxWorkerOptions()
|
||
if options.PollInterval <= 0 {
|
||
// 非正间隔不能直接传给 ticker，否则会 panic 或形成忙循环。
|
||
options.PollInterval = defaults.PollInterval
|
||
}
|
||
if options.BatchSize <= 0 {
|
||
// 非正批量按安全默认值处理，避免 repository 扫描不受控。
|
||
options.BatchSize = defaults.BatchSize
|
||
}
|
||
if options.PublishTimeout <= 0 {
|
||
// 单条外部投递必须有 deadline，shutdown 才能等待有界。
|
||
options.PublishTimeout = defaults.PublishTimeout
|
||
}
|
||
if options.MaxRetryCount <= 0 {
|
||
options.MaxRetryCount = defaults.MaxRetryCount
|
||
}
|
||
if options.InitialBackoff <= 0 {
|
||
options.InitialBackoff = defaults.InitialBackoff
|
||
}
|
||
if options.MaxBackoff <= 0 {
|
||
options.MaxBackoff = defaults.MaxBackoff
|
||
}
|
||
if options.MaxBackoff < options.InitialBackoff {
|
||
options.MaxBackoff = options.InitialBackoff
|
||
}
|
||
options.RetryStrategy = strings.TrimSpace(options.RetryStrategy)
|
||
if options.RetryStrategy == "" {
|
||
// 空策略按指数退避处理，保持和配置默认值一致。
|
||
options.RetryStrategy = defaults.RetryStrategy
|
||
}
|
||
|
||
return options
|
||
}
|
||
|
||
// RunOutboxWorker 持续扫描 pending outbox；它不执行房间命令，只投递已持久化事件。
|
||
func (s *Service) RunOutboxWorker(ctx context.Context, options OutboxWorkerOptions) {
|
||
options = normalizeOutboxWorkerOptions(options)
|
||
if options.RetryStrategy != outboxRetryExponentialBackoff {
|
||
// 配置层会拒绝未知策略；这里兜底避免测试或手写构造静默改变语义。
|
||
logx.Error(ctx, "worker_stopped", fmt.Errorf("unsupported retry_strategy"), slog.String("worker", outboxWorkerName), slog.String("node_id", s.nodeID))
|
||
return
|
||
}
|
||
|
||
if err := s.ProcessPendingOutbox(ctx, options); err != nil && ctx.Err() == nil {
|
||
logx.Error(ctx, "worker_batch_failed", err, slog.String("worker", outboxWorkerName), slog.String("node_id", s.nodeID), slog.Int("batch_size", options.BatchSize))
|
||
}
|
||
|
||
ticker := time.NewTicker(options.PollInterval)
|
||
defer ticker.Stop()
|
||
|
||
for {
|
||
select {
|
||
case <-ctx.Done():
|
||
return
|
||
case <-ticker.C:
|
||
if err := s.ProcessPendingOutbox(ctx, options); err != nil && ctx.Err() == nil {
|
||
logx.Error(ctx, "worker_batch_failed", err, slog.String("worker", outboxWorkerName), slog.String("node_id", s.nodeID), slog.Int("batch_size", options.BatchSize))
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// ProcessPendingOutbox 把待投递事件补偿性地推送给异步消费者。
|
||
func (s *Service) ProcessPendingOutbox(ctx context.Context, options OutboxWorkerOptions) error {
|
||
options = normalizeOutboxWorkerOptions(options)
|
||
if options.RetryStrategy != outboxRetryExponentialBackoff {
|
||
return fmt.Errorf("unsupported outbox retry_strategy %q", options.RetryStrategy)
|
||
}
|
||
|
||
// 补偿 worker 先抢占 pending/retryable 事件，再投递，避免多机同时处理同一批。
|
||
records, err := s.repository.ClaimPendingOutbox(ctx, s.nodeID, options.BatchSize, time.Now().UTC().Add(options.PublishTimeout).UnixMilli())
|
||
if err != nil {
|
||
return err
|
||
}
|
||
|
||
for _, record := range records {
|
||
select {
|
||
case <-ctx.Done():
|
||
return ctx.Err()
|
||
default:
|
||
}
|
||
if record.RetryCount >= options.MaxRetryCount {
|
||
// 已达到最大失败次数的历史事件直接转死信，防止每轮 worker 继续打外部系统。
|
||
markCtx, markCancel := context.WithTimeout(appcode.WithContext(context.Background(), record.AppCode), options.PublishTimeout)
|
||
markErr := s.repository.MarkOutboxDead(markCtx, record.EventID, deadLetterReason(record))
|
||
markCancel()
|
||
if markErr != nil {
|
||
return markErr
|
||
}
|
||
logx.Error(ctx, "outbox_dead_letter", nil,
|
||
slog.String("worker", outboxWorkerName),
|
||
slog.String("node_id", s.nodeID),
|
||
slog.String("event_id", record.EventID),
|
||
slog.String("event_type", record.EventType),
|
||
slog.String("room_id", record.RoomID),
|
||
slog.Int("retry_count", record.RetryCount),
|
||
slog.String("last_error", trimOutboxError(record.LastError)),
|
||
)
|
||
continue
|
||
}
|
||
|
||
publishCtx, cancel := context.WithTimeout(context.Background(), options.PublishTimeout)
|
||
err := s.outboxPublisher.PublishOutboxEvent(publishCtx, record.Envelope)
|
||
cancel()
|
||
if err != nil {
|
||
// 投递失败转为 retryable，并写入指数退避时间，等待下一轮 worker 到期后再抢占。
|
||
nextRetryAtMS := time.Now().UTC().Add(outboxBackoff(record.RetryCount+1, options)).UnixMilli()
|
||
markCtx, markCancel := context.WithTimeout(appcode.WithContext(context.Background(), record.AppCode), options.PublishTimeout)
|
||
markErr := s.repository.MarkOutboxRetryable(markCtx, record.EventID, trimOutboxError(err.Error()), nextRetryAtMS)
|
||
markCancel()
|
||
if markErr != nil {
|
||
return markErr
|
||
}
|
||
|
||
logx.Warn(ctx, "outbox_publish_retryable",
|
||
slog.String("worker", outboxWorkerName),
|
||
slog.String("node_id", s.nodeID),
|
||
slog.String("event_id", record.EventID),
|
||
slog.String("event_type", record.EventType),
|
||
slog.String("room_id", record.RoomID),
|
||
slog.Int("retry_count", record.RetryCount+1),
|
||
slog.Int64("next_retry_at_ms", nextRetryAtMS),
|
||
slog.String("error", trimOutboxError(err.Error())),
|
||
)
|
||
continue
|
||
}
|
||
|
||
// 只有外部发布成功后才能标记 delivered。
|
||
markCtx, markCancel := context.WithTimeout(appcode.WithContext(context.Background(), record.AppCode), options.PublishTimeout)
|
||
markErr := s.repository.MarkOutboxDelivered(markCtx, record.EventID)
|
||
markCancel()
|
||
if markErr != nil {
|
||
return markErr
|
||
}
|
||
|
||
logx.Info(ctx, "outbox_delivered",
|
||
slog.String("worker", outboxWorkerName),
|
||
slog.String("node_id", s.nodeID),
|
||
slog.String("event_id", record.EventID),
|
||
slog.String("event_type", record.EventType),
|
||
slog.String("room_id", record.RoomID),
|
||
slog.Int("retry_count", record.RetryCount),
|
||
)
|
||
}
|
||
|
||
return nil
|
||
}
|
||
|
||
func outboxBackoff(retryCount int, options OutboxWorkerOptions) time.Duration {
|
||
if retryCount <= 1 {
|
||
return options.InitialBackoff
|
||
}
|
||
backoff := options.InitialBackoff
|
||
for i := 1; i < retryCount; i++ {
|
||
if backoff >= options.MaxBackoff/2 {
|
||
return options.MaxBackoff
|
||
}
|
||
backoff *= 2
|
||
}
|
||
if backoff > options.MaxBackoff {
|
||
return options.MaxBackoff
|
||
}
|
||
return backoff
|
||
}
|
||
|
||
func deadLetterReason(record outbox.Record) string {
|
||
if strings.TrimSpace(record.LastError) == "" {
|
||
return "outbox retry limit exceeded"
|
||
}
|
||
return trimOutboxError("outbox retry limit exceeded: " + record.LastError)
|
||
}
|
||
|
||
func trimOutboxError(value string) string {
|
||
// last_error 会写入 MySQL 和日志，先去掉首尾空白避免不可见差异影响排障。
|
||
value = strings.TrimSpace(value)
|
||
if len(value) <= 512 {
|
||
return value
|
||
}
|
||
|
||
// V1 只保留错误前缀，避免外部响应过大撑爆日志或 TEXT 字段查询成本。
|
||
return value[:512]
|
||
}
|