[BUG] edge_num_ double-subtraction in batch_delete_edges due to missing already-deleted guard

[zhanglei1949]

Describe the bug

edge_num_ can be incorrectly decremented multiple times for already-deleted edges, causing the counter to underflow (wrap around for uint64_t), leading to incorrect edge count reporting and potential downstream failures.

Affected locations

1. SingleMutableCsr::batch_delete_edges (offset version) — src/storages/csr/mutable_csr.cc:676-697

for (const auto& edge : edge_list) {
    vid_t src = edge.first;
    if (src >= vnum) { continue; }
    auto& nbr = data[src];
    assert(edge.second == 0);
    nbr.timestamp.store(std::numeric_limits<timestamp_t>::max());
    edge_num_.fetch_sub(1, std::memory_order_relaxed);  // ← unconditional!
}

edge_num_.fetch_sub(1) is called without checking whether the edge was already deleted (timestamp already at max). If the same edge is deleted twice (e.g., through retry logic or duplicate entries), the counter is decremented twice.

2. MutableCsr::batch_delete_edges (src/dst version) — src/storages/csr/mutable_csr.cc:363-370

if (pair.second.find(write_ptr->neighbor) != pair.second.end()) {
    write_ptr->timestamp.store(std::numeric_limits<timestamp_t>::max());
    edge_num_.fetch_sub(1, std::memory_order_relaxed);  // ← unconditional!
}

Same issue — no check for already-deleted state before decrementing.

Contrast with correct implementations

• MutableCsr::batch_delete_edges (offset version, line 399-405) correctly checks before decrementing.
• SingleMutableCsr::batch_delete_edges (src/dst version, line 668-672) correctly checks before decrementing.

Only the two locations above are missing the guard.

Expected behavior

All batch_delete_edges variants should check timestamp != max before calling edge_num_.fetch_sub(1), consistent with the other implementations in the same file.

Suggested fix

For SingleMutableCsr::batch_delete_edges (offset version):

if (nbr.timestamp.load() != std::numeric_limits<timestamp_t>::max()) {
    edge_num_.fetch_sub(1, std::memory_order_relaxed);
}
nbr.timestamp.store(std::numeric_limits<timestamp_t>::max());

For MutableCsr::batch_delete_edges (src/dst version):

if (write_ptr->timestamp.load() != std::numeric_limits<timestamp_t>::max()) {
    edge_num_.fetch_sub(1, std::memory_order_relaxed);
}
write_ptr->timestamp.store(std::numeric_limits<timestamp_t>::max());

Impact

• Severity: High — edge count corruption can trigger false THROW_STORAGE_EXCEPTION in compact() (line 204-211) and Open() (line 79-87), causing storage layer failures.
• Likelihood: Medium — requires duplicate edge deletion, which can occur during transaction retries or WAL replay.

[github-actions]

🤖 Auto-triage: could not confidently classify this issue under any current v0.2 umbrella.

Reason: The issue is a bug in the CSR storage layer's batch_delete_edges method, which does not fit any specific umbrella's scope_in (not transaction, not checkpoint, not COPY, not GDS, etc.) and is a targeted bug fix rather than an umbrella/tracking issue itself.

Auto-applied: type → Bug

Maintainers: please pick a parent manually, or close if out of scope.