Read transaction stays open indefinitely under high write load (renewingRefCount stuck)

[kriszyp]

Under high write throughput in a sharded cluster, read transactions are not closing and report errors every second until manual intervention:

[error]: Read transaction detected that has been open too long (over one minute)
Txn {
  address: 140427120637888,
  timerTracked: true,
  refCount: 1,
  renewingRefCount: 1,  ← stuck
  notCurrent: true,
  openTimer: 58
}

The renewingRefCount: 1, notCurrent: true combination suggests the transaction renewal cycle is failing to complete — the transaction knows it's stale but can't close itself because of a pending renewer.

Reproduction context

• 5-node sharded cluster, 50k req/s write throughput
• 1 KB record size with blobs
• Seen on 4.6.0-alpha.3

Related

• HarperFast/harper PR #304 — "make read txn timeout configurable and set default to 1min" (in review) adds a configurable timeout, but doesn't appear to address the stuck-renewer root cause.

Acceptance criteria

• Read transactions close or are forcibly recycled even when renewingRefCount stays positive.
• Under high write load, the "open too long" error storm doesn't recur.

🤖 Filed by Claude on behalf of Kris.