Files
webui4frpc/internal/cluster/ring_log.go
JianFeeeee 57950a86db fix: 集群日志重复条目 — 本地 Append 同步推进 Synced 水位
根因: Append 只推进 Seq 不推进 Synced。本地追加 seq=N 后 Synced 落后,
下一轮收到对端全量附带(含自己的 seq=N)时 ApplyDelta 视为未见过而再次追加
→ 审计日志出现同 seq 重复行(实测 .30 出现 seq=12/13/14 各两条)。

修复: Append 将 Synced 一并推进到新条目 seq——本地条目天然已同步,
对端回传的同一 seq 被 watermark 幂等跳过。

测试: TestLocalAppendNotDuplicatedByPeerFullLog

实测: 三台各触发 stop+start 制造多条事件 → 三台均 26 条 / 唯一seq 26 /
重复 0 / 水位一致 (seq=26)
2026-08-26 13:30:12 +08:00

171 lines
4.9 KiB
Go

// Cluster operation log + incremental sync via the token.
//
// Every node keeps an append-only operation log (increasing Seq). Local
// mutations (forward created/removed, node joined/left, leader change, ...)
// are appended to the node's own log. During the token's sync round (phase 2)
// the token carries the log DELTA — entries with Seq > lastSyncedSeq — so
// every other node can append-and-replay them in order, converging on the same
// full cluster picture. Because each node has the full log, a newcomer can
// pull + replay the whole log to reconstruct identical state.
package cluster
import (
"encoding/json"
"fmt"
"sort"
"sync"
"time"
)
// LogKind enumerates operation-log entry types.
const (
LogForwardAdd = "forward.add"
LogForwardRemove = "forward.remove"
LogNodeJoin = "node.join"
LogNodeLeave = "node.leave"
LogLeaderChange = "leader.change"
LogTaskClaimed = "task.claimed"
)
// LogEntry is one immutable, append-only cluster operation.
type LogEntry struct {
Seq int64 `json:"seq"`
Node string `json:"node"`
Kind string `json:"kind"`
Data json.RawMessage `json:"data,omitempty"`
At int64 `json:"at"`
}
// ClusterLog is a node's local operation log with a sync watermark.
type ClusterLog struct {
mu sync.Mutex
Seq int64 // highest local seq issued
Log []LogEntry // append-only
Synced int64 // watermark: entries <= this are known to remote nodes
}
// NewClusterLog creates an empty log with initial sequence.
func NewClusterLog() *ClusterLog {
return &ClusterLog{}
}
// Append adds an entry with the next sequence number (caller supplies kind/data).
// The entry is treated as already-synced (Synced advanced with Seq): a locally
// appended entry exists on this node and will be broadcast; the next full-log
// attachment from a peer must NOT re-apply it as a duplicate. Without this,
// a node that appends seq=N while its Synced lags behind would later receive
// its own seq=N in a peer's full attachment and ApplyDelta would re-add it
// (duplicate rows in the audit log).
func (l *ClusterLog) Append(node, kind string, data any) (LogEntry, error) {
l.mu.Lock()
defer l.mu.Unlock()
l.Seq++
raw, err := json.Marshal(data)
if err != nil {
l.Seq--
return LogEntry{}, err
}
e := LogEntry{Seq: l.Seq, Node: node, Kind: kind, Data: raw, At: time.Now().Unix()}
l.Log = append(l.Log, e)
if e.Seq > l.Synced {
l.Synced = e.Seq
}
return e, nil
}
// EntriesAfter returns entries with seq > after (delta for sync round).
func (l *ClusterLog) EntriesAfter(after int64) []LogEntry {
l.mu.Lock()
defer l.mu.Unlock()
var out []LogEntry
for _, e := range l.Log {
if e.Seq > after {
out = append(out, e)
}
}
return out
}
// ApplyDelta appends-and-replays remote delta entries in order; returns the new
// watermark. Entries with seq <= existing watermark are skipped (idempotent).
func (l *ClusterLog) ApplyDelta(delta []LogEntry) (newWatermark int64, err error) {
l.mu.Lock()
defer l.mu.Unlock()
// ensure ordered by seq
sort.Slice(delta, func(i, j int) bool { return delta[i].Seq < delta[j].Seq })
last := l.Synced
for _, e := range delta {
if e.Seq <= last {
continue // already have
}
if e.Seq != last+1 {
return last, fmt.Errorf("gap in log delta: want %d got %d", last+1, e.Seq)
}
l.Log = append(l.Log, e)
last = e.Seq
}
l.Synced = last
if last > l.Seq {
l.Seq = last
}
return last, nil
}
// Replay applies a set of log entries locally (used at join to reconstruct
// state). Same idempotent-by-seq semantics as ApplyDelta.
func (l *ClusterLog) Replay(entries []LogEntry) error {
_, err := l.ApplyDelta(entries)
return err
}
// Snapshot returns a copy of the log (for debugging / join bootstrap).
func (l *ClusterLog) Snapshot() []LogEntry {
l.mu.Lock()
defer l.mu.Unlock()
out := make([]LogEntry, len(l.Log))
copy(out, l.Log)
return out
}
// DetailOf produces a human-readable one-line summary of a log entry's
// payload, matching the frontend's detailOf formatting. Used by the audit CSV
// export and anywhere a flat text rendering of an entry is needed.
func DetailOf(e LogEntry) string {
if len(e.Data) == 0 {
return ""
}
var d map[string]any
if err := json.Unmarshal(e.Data, &d); err != nil {
return string(e.Data)
}
str := func(key string) string { s, _ := d[key].(string); return s }
switch e.Kind {
case LogForwardAdd, LogForwardRemove:
s := str("local") + " → " + str("remote")
if id := str("taskId"); id != "" {
if len(id) > 8 {
id = id[len(id)-8:]
}
s += " · " + id
}
return s
case LogNodeJoin:
if addr := str("addr"); addr != "" {
return str("node") + " @ " + addr
}
return str("node")
case LogNodeLeave:
return str("node")
case LogLeaderChange:
return "→ " + str("leader")
case LogTaskClaimed:
return fmt.Sprintf("%s→%s:%v", str("local"), str("remote"), d["port"])
default:
if len(d) > 0 {
b, _ := json.Marshal(d)
return string(b)
}
return ""
}
}