mirror of
https://github.com/hanzoai/benchmarks.git
synced 2026-10-30 14:59:11 +00:00
zeekay
feb58249a4
First hanzoai/benchmarks: a home for api.hanzo.ai / cloud unified-binary perf. cloud/shard proves the horizontal writer-scale invariant via hanzoai/ha rendezvous: 1M tenants shard dead-even across N pods (0.05% stddev @ N=3), one deterministic owner per org (no dual-writer), ~545ns routing tax, 20% remap on 4->5 scale-up. Makefile (scale/bench) + k3s integration harness scaffold.
57 lines
2.3 KiB
Bash
Executable File
57 lines
2.3 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Local N-pod cloud shard harness on k3s (installed at /usr/local/bin/k3s).
|
|
#
|
|
# Stands up the cloud unified binary as a StatefulSet with per-pod RWO PVCs and
|
|
# CLOUD_PEERS set to the 3 pod DNS names, so the org->owner shard router is LIVE
|
|
# locally — the same topology the prod cutover will use — without touching prod.
|
|
#
|
|
# SAFETY: a real per-org WRITE load test across N pods is only sound once the
|
|
# durable-tasks :9999 off-router blocker is fixed (a job on a non-owner pod would
|
|
# write the wrong PVC). Until then this harness proves ROUTING + READ scaling +
|
|
# failover; it refuses write-load unless HANZO_SHARD_WRITE_OK=1 is set explicitly.
|
|
set -euo pipefail
|
|
NS=${NS:-cloud-bench}
|
|
N=${N:-3}
|
|
IMAGE=${IMAGE:-ghcr.io/hanzoai/cloud:latest} # CI-built; we never build locally
|
|
K=${KUBECTL:-kubectl}
|
|
|
|
echo ">> ensuring a local k3s cluster"
|
|
if ! $K get nodes >/dev/null 2>&1; then
|
|
sudo k3s server --write-kubeconfig-mode 644 >/tmp/k3s.log 2>&1 &
|
|
for i in $(seq 1 30); do $K get nodes >/dev/null 2>&1 && break || sleep 1; done
|
|
fi
|
|
$K get nodes
|
|
|
|
PEERS=""; for i in $(seq 0 $((N-1))); do
|
|
PEERS+="${PEERS:+,}cloud-shard-$i@cloud-shard-$i.cloud-shard.$NS.svc:8000"
|
|
done
|
|
echo ">> N=$N CLOUD_PEERS=$PEERS"
|
|
|
|
$K create ns "$NS" --dry-run=client -o yaml | $K apply -f -
|
|
|
|
# Prefer the canary StatefulSet blue authored (universe:blue/cloud-horizontal-shard).
|
|
MANIFEST=${MANIFEST:-$HOME/work/hanzo/universe/infra/k8s/cloud/cloud-statefulset.yaml}
|
|
if [[ -f "$MANIFEST" ]]; then
|
|
echo ">> applying $MANIFEST (edit replicas=$N, CLOUD_PEERS, image as needed)"
|
|
$K -n "$NS" apply -f "$MANIFEST"
|
|
else
|
|
echo "!! canary manifest not found at $MANIFEST"
|
|
echo " checkout universe:blue/cloud-horizontal-shard, or point MANIFEST= at it."
|
|
exit 1
|
|
fi
|
|
|
|
echo ">> waiting for $N shard pods (image pull + boot)…"
|
|
$K -n "$NS" rollout status statefulset/cloud-shard --timeout=300s || true
|
|
$K -n "$NS" get pods -o wide
|
|
|
|
cat <<EOF
|
|
|
|
>> shard stack up. Drive it:
|
|
ROUTING/READ proof (safe now):
|
|
$K -n $NS port-forward svc/cloud-shard 8000:8000
|
|
# send requests with distinct X-Org-Id; confirm each org's writes land on ONE pod
|
|
# $K -n $NS logs cloud-shard-0 | grep 'org_slug'
|
|
WRITE load (only after the :9999 blocker is fixed):
|
|
HANZO_SHARD_WRITE_OK=1 ./k3s/loadgen.sh # gated on purpose
|
|
EOF
|