feat(test): TestBenchmark drives seiload (WS-I load suite)

[bdchatham]

WS-I — complete the load suite (TestBenchmark)

Builds on the merged foundation (#428) + SDK Labels/DeletionPolicy (#429). After provisioning the chain + RPC fleet, TestBenchmark now drives seiload and asserts the chain survived the load.

Flow

1. Render the platform seiload profile (read from the seiload-profiles ConfigMap) with the per-run chain id + the fleet's EVM endpoints; write a per-run profile CM stamped sei.io/harness-run.
2. Apply seiload's own Job manifest (embedded template, parameterized — its Job spec is not constructed in Go; D3 decoupling) carrying the metrics scrape label + harness-run.
3. Wait for the Job to run the full load to completion.
4. Assert the chain stayed live (post-load WaitCaughtUp on node-0).

Assert model (deliberate)

Pass/fail = Job completion + post-load chain liveness. A throughput/regression gate belongs in telemetry — a PromQL query over the run's metrics (the conventional path, like validate-release reads Grafana) — not parsed from the report. Left as a documented follow-up; the Job already carries the metrics scrape label so it's a drop-in.

Decoupling / cleanup

• seiload runs from its own manifest; the profile is platform-owned (read from the cluster CM, not vendored).
• Per-run profile CM + seiload Job carry sei.io/harness-run; t.Cleanup deletes them on normal exit, the label-GC sweep backstops abnormal exit.

Verification

gofmt clean · go build ./... clean (no integration code in prod deps) · go vet -tags integration clean · golangci-lint 0 issues · go test -c -tags integration → TestBenchmark · skips without SEI_NODE_CLUSTER.

Next

Live smoke on harbor (demonstration) + platform PR to point the load CronJob at the new image and update the podMonitor selector to sei.io/harness-run.

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Creates real cluster Jobs/ConfigMaps and long-running load against nightly infrastructure; failures are mostly test-scoped, but misconfiguration could leave labeled resources until cleanup or future GC.

Overview
TestBenchmark now runs the full load path instead of skipping after provision: it provisions the chain + RPC fleet, then invokes runSeiload to apply load and assert liveness.

The harness renders a platform-owned profile from the cluster seiload-profiles ConfigMap (chain id + JSON-quoted EVM RPC URLs), writes a per-run profile ConfigMap labeled sei.io/harness-run, and applies a parameterized seiload Batch Job from an embedded YAML template (not built in Go). It waits for Job success (with pod log tail on failure) and checks every RPC follower is still caught up after load. Pass/fail is Job completion + post-load sync; throughput gates are explicitly deferred to PromQL/metrics.

spec and env gain SEILOAD_IMAGE, profile, duration, and commit inputs (envInt for DURATION_MINUTES). Cleanup uses t.Cleanup for the Job and profile CM; harness comments note label-GC sweep is still a pending platform backstop.

^{Reviewed by Cursor Bugbot for commit 557897e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bdchatham]

/xreview — 4 lenses (idiom · systems · k8s-dissenter · sei-network), RESOLVED

Comprehensive blinded review; spine ratified, fixes applied in the latest commit.

#	Finding	Lens	Grade	Resolution
F1	Failed seiload pod log discarded — Job condition message alone can't diagnose why	systems	advisory (3am-critical)	waitJob now tails the pod log into the fatal
F2	Job has no activeDeadlineSeconds — only the harness ctx bounds a hung seiload	systems	correctness (defense-in-depth)	self-terminating cap (durationMin+15)*60
F3	Post-load liveness checked only node-0 — a half-dead fleet ships green	sei-network	advisory	widened to every follower
F4	runLabelKey comment claims a label-GC sweep that isn't shipped	dissenter + idiom	correctness (honesty)	softened — sweep is a pending platform deliverable; cascade + t.Cleanup cover normal exit
F5	namespace re-resolved from env in runSeiload	dissenter	advisory	use the chain's resolved namespace directly
F6	runSeiload doc narrates rationale/future-plan	idiom	style	trimmed to present-state
—	chainId 713714	sei-network	COMPATIBLE	verified: seid GetEVMChainID → DefaultChainID for bench-* chains
—	CM key/mount, namespace non-divergence, pod-template GC label, DeletionPolicy, cleanup ordering	all	COMPATIBLE	conceded by the dissenter

Platform prereqs (tracked separately, NOT this PR): a harness-SA Role (configmaps/jobs/seinetworks create), the podMonitor selector flip to app.kubernetes.io/name=seiload, and the sei.io/harness-run gc-cronjob label sweep — the platform cutover + Step-2 deliverables.

Live smoke of TestBenchmark on harbor in progress (the demonstration).