Posts tagged compute amp.

Agave evaluates getProgramAccounts filters inline during the scan, not via an index, so a never-matching filter holds a spawn_blocking thread for the full walk and returns an empty 37-byte array. The measured signal is pool saturation under concurrency: aggregate throughput pegs flat at ~60 req/s while per-worker rate drops 8x.

Prysm's POST /eth/v2/beacon/pool/attestations decodes an unbounded JSON list and runs a BLS G2 subgroup-check per element before any cap applies, turning a 59 MB request into 5.81 seconds of single-goroutine CPU on an exposed Beacon API.