Warm-standby cache survival on promotion is dead code: OnLeaderStart unsubscribes before reading the consistent ts → every promotion falls back to full cold recovery

[liunyl]

The warm-standby fast-promotion optimization (keep the in-memory cache, replay only from the standby's consistent ts) never engages, because OnLeaderStart unsubscribes all sequence groups before it reads the per-shard consistent ts that the optimization depends on — and the read filters on subscribed_, so it always returns UINT64_MAX.

Evidence

OnLeaderStart unsubscribes every sequence group early (tx_service/src/fault/cc_node.cpp:323-340):

WaitableCc sub_cc([](CcShard &ccs) {
    for (auto &[grp_id, grp_info] : ccs.GetStandbysequenceGrps())
        if (grp_info.subscribed_) grp_info.Unsubscribe();   // subscribed_ = false
    return true;
}, core_cnt);
... sub_cc.Wait();

Later in the same function it reads the consistent ts to decide whether the cache survives (cc_node.cpp:401-425):

uint64_t last_consistent_ts = UINT64_MAX;
WaitableCc get_consistent_ts_cc([&](CcShard &ccs) {
    uint64_t shard_ts = ccs.MinLastStandbyConsistentTs();   // see below
    last_consistent_ts = std::min(last_consistent_ts, shard_ts);
    return true;
}, ...);
...
if (last_consistent_ts != UINT64_MAX && last_consistent_ts != 0) {
    cache_survivied = true;
    replay_start_ts = last_consistent_ts + 1;
}

But MinLastStandbyConsistentTs only considers subscribed groups (tx_service/src/cc/cc_shard.cpp:3539-3549):

uint64_t min_ts = UINT64_MAX;
for (auto &seq_grp : standby_sequence_grps_)
    if (seq_grp.second.subscribed_)
        min_ts = std::min(seq_grp.second.last_standby_consistent_ts_, min_ts);
return min_ts;

Since every group was just unsubscribed, this returns UINT64_MAX on every shard, last_consistent_ts stays UINT64_MAX, cache_survivied stays false, and promotion takes ClearCcNodeGroupData(); replay_start_ts = 0 (cc_node.cpp:432-436).

Impact

Every standby promotion clears the warm cache and does a full cold recovery (KV + replay from 0), silently defeating the warm-standby fast-failover feature — slow promotions, large recovery I/O. (Whether the full-clear floor also interacts with WAL truncation to lose data on non-shared storage should be checked as a follow-on; the dead-optimization itself is confirmed.)

Fix: capture the consistent ts before unsubscribing (or have MinLastStandbyConsistentTs read a snapshot taken pre-unsubscribe).

---

Found during a code audit (PR #493), in the priority standby area. Verified against source at the cited lines.

🤖 Found with Claude Code