test(perf): add manual-only extended coverage and DuckDB baseline

[bill-ph]

The frozen benchmarks leave common analytical query shapes unmeasured. Add 12 versioned workloads covering filtered uniques/trends, two person joins, event and actor retrieval, ordered funnels, sampled-week retention, two cohorts, and daily-person/session modeling SELECTs.

Manually select scenario=posthog_frozen_perf_extended to run all five configurations, or set coverage_target for one. Shared image builds feed isolated, sequential target jobs with separate artifacts and a pinned Trino digest. Each target gets a four-hour scenario / 270-minute job budget; failures do not cancel remaining targets. Extended queries stay out of nightly runs.

Each query runs one warmup and four measurements. Fixed populated windows, tenant-aware person deduplication, and fixture checks avoid misleadingly empty workloads. Scenario Job names respect Kubernetes limits.

Comparable dev-sized worker resources

Standard frozen perf and extended perf use three Trino workers at 7 CPU / 28 GiB each and one DuckDB worker at 21 CPU / 84 GiB. Worker requests equal limits. Trino uses a 20 GiB heap, 10 GiB per-worker query memory and 30 GiB cluster query memory. DuckDB sizing admission caps rise with its requested profile; the current headroom policy derives a 63 GiB engine memory limit. Aggregate container resources match; coordinator/support overhead and engine memory policies differ.

Extended Trino jobs deploy only the selected comparison cache mode. Bootstrap/general E2E resources retain their small defaults. Nightly keeps its original query set with the larger comparison resources.

Deployment prerequisite: charts #16434 allows node sizes that can accommodate the larger pods. Apply it before running this profile. No larger-profile benchmark has run yet.

Historical measurement

The uncached DuckDB baseline run, commit 9c9f5f305f87b7f0d41c0b54d64e5648aeb8ef83, completed 12 warmups and 48 measurements without errors: 87m06s benchmark phase, 1h42m45s workflow. Sanitized samples and methodology remain in tests/perf/baselines/coverage-v1-duckdb-uncached-2026-09-25.

That run used the old 3 CPU / 12 GiB aggregate worker budget. It is historical data, not a like-for-like comparison base for the new profile. Its runtime extrapolations are also historical; a fresh measured run is required to establish the larger-profile baseline and forecast. The subsequent all-target run hit Trino's old 2 GiB per-node query-memory limit on joins and session modeling; this sizing change addresses that constraint but does not claim those queries now pass.

Validation and limits

• Existing scenario checks, workflow actionlint, shell syntax, and rendered sizing checks pass. Existing fixture tests pass except the previously reproduced shared-catalog fixture failure. No new perftest unit tests; only existing raw-template substitutions were adapted.
• just lint still reports six pre-existing SA4023 diagnostics at three unchanged sites.
• The dependent charts change passes rendering, kubeconform and kube-linter for all nine environments.
• Sampled-day retention is a performance shape, not a production rate; modeling queries measure computation, not writes. Uncached settings do not guarantee cold OS/storage caches. Four repetitions do not establish tail latency.

[github-actions]

Test Impact Plan

Deterministic summary of how this PR changes tests, CI runners, and coverage-risk signals.

Summary

Area	Added	Changed	Deleted
Test files	8	8	0
E2E/journey files	0	0	0
Workflow files	0	1	0

Signals

• Test cases: +0 / -0
• Assertions: +1 / -1
• Skips or known failures added: 0
• Workflow continue-on-error added: 0
• Workflow path filters added: 0
• Test commands removed from justfile: 0
• E2E/journey retry lines added: 0

Coverage risk: neutral or increased

No coverage-reduction warnings detected.