Research

Benchmark Results

Benchmark Results

Astra vs etcd snapshots

Phase 4: 150 IOPS survival, wal_low_linger tuned run

  • Date: 2026-02-27
  • Status: PASS
  • Workload: 50k write scenario under blkio shaping
  • Astra: 7267.42 RPS, 47.52 ms p99, 50k/50k writes, no OOM
  • etcd: 980.17 RPS on the same run shape
  • Notes: This was the first public-quality proof that Astra could beat etcd decisively under the constrained-disk scenario that originally blocked the project.

Phase 6: K3s external datastore benchmark

  • Date: 2026-02-28
  • Status: PASS with cardinality caveat
  • Workload: kwok-driven K3s churn at ~17k pod realization
  • Astra: PUT 335.25 ms p99, GET 312.17 ms p99, LIST 2952.64 ms p99, 1001 nodes / 17506 pods / 10001 configmaps
  • etcd: PUT 158.11 ms p99, GET 143.06 ms p99, LIST 2728.32 ms p99, 1001 nodes / 16113 pods / 10001 configmaps
  • Notes: Astra was stable but materially slower than etcd for this early control-plane workload shape.

Phase 9: RC2 K3s acceptance matrix

  • Date: 2026-03-03
  • Status: PASS
  • Workload: 30k pod cardinality, post-warmup acceptance run
  • Astra: PUT 151.13 ms p99, GET 117.06 ms p99, LIST 268.02 ms p99, peak disk IOPS 30.0, iowait p99 5.0%
  • etcd: PUT 106.01 ms p99, GET 108.01 ms p99, LIST 259.00 ms p99, peak disk IOPS 22.0, iowait p99 16.5%
  • Notes: This run closed the biggest K3s acceptance gaps: GET decoupling, LIST parity, iowait harmony, and 30k pod realization.

Feature validations

Phase 10: QoS, governor, ecosystem, and gray-failure suite

  • Date: 2026-03-03
  • Status: PASS
  • Summary: Gateway storm reached 13447.49 RPS with 0.0253% error rate; Patroni lease updates stayed at 33.02 ms p99; gray-failure delta gate passed with stable leader behavior.

Phase 11: Migration DX and Omni fleet cutover

  • Date: 2026-03-03
  • Status: PASS
  • Summary: 5 sources converged with 5/5 tenant parity matches; 5/5 real Omni external-etcd boots passed with no key loss.

Phase 12: Directive closure run

  • Date: 2026-03-05
  • Status: PASS
  • Summary: 50k watch fanout passed at 4995.20 write RPS and 150 ms lag p99; semantic cache showed 1.299x RPS gain and 1.292x p99 improvement; multi-raft prefix sharding reached 2.009x scale gain.

Rebuild this page

uv run --project refs/scripts python refs/scripts/public/generate_public_reports.py