The RPC probe flapped while every other knots check stayed green. Each burst
lined up with a block: validation on the saturated HDD holds the node busy
while Fulcrum, mempool and DATUM all fetch the block, and bitcoind rejected
calls with "work queue depth exceeded" (93 times on 2026-10-18).
- 403: the probe runs every minute, so failure-threshold 4 rides out a
1-3 minute burst while a real outage still alerts within ~5 minutes.
Probes now take an optional failure_threshold over a shared probe_alert.
- rpcworkqueue=64 and rpcthreads=8, so bursts queue instead of failing.
Profiling later showed the stalls are per-input UTXO lookups at ~5 ms each -
HDD seek latency - so this treats the symptom. The fix is chainstate on SSD.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
counterweight
2a80f0e3ca
