Overview

The query_golden_join_optimization_plan_stability resmoke suite uses the
default internalJoinPlanSamplingSize=1000 for the join optimizer's
cardinality-estimation sampler. With the deterministic sequential sample of
1000 docs, several queries in the plan_stability_tpch_fuzzed.js input set
have outer-side filters with selectivity below 0.2% — the sampler misses all
matching rows, returns CE=0, and the cost terms collapse to 0 across the
join graph. The join enumerator then falls back to its deterministic
ordering (first-tried-INLJ-when-an-index-exists) and produces a plan that
isn't a cost-based choice.

This causes brittle baselines: the recorded "good plan" is in fact an
under-sampling artifact, and any cost-model change that nudges CE estimates
even slightly (e.g. SERVER-123070's kMinCE clamp) produces visible
regressions against the baselines, even when the new plans are actually
closer to ground truth.

Proposed Change

Set internalJoinPlanSamplingSize: 10000 in
buildscripts/resmokeconfig/suites/query_golden_join_optimization_plan_stability.yml
and regenerate the golden baselines on master.

This is a 10x increase. At sample 10000:
* part (20 K rows): 50% sample — every realistic outer-side filter is reliably estimated.
* orders (150 K rows): ~6.7% sample — even the 0.02% selectivity filters in idx 211 land 1-2 sample matches in expectation.
* lineitem (~600 K rows): ~1.7% — adequate for any non-pathological filter.

Rationale

A sweep of internalJoinPlanSamplingSize on clean master against the
"idx 1" query (5-way TPC-H join with a 0.145%-selective part filter)
shows the optimizer's plan choice converges quickly and stably:

The 10000-sample plan matches what the cost model produces with exact
cardinalities. The default-sample-size baseline is the only one that
doesn't.

Verification

End-to-end test on SERVER-123070 (the branch this finding came out of):* Bumped suite YAML to internalJoinPlanSamplingSize: 10000 on clean master.

• Regenerated baselines for plan_stability_tpch_fuzzed and plan_stability_subjoin_cardinality.md. plan_stability_tpch_official was unchanged (its queries were already adequately sampled at 1000).
• Copied YAML and baselines to the SERVER-123070 branch, re-ran all three tests: 0-line diff against new baselines, all pass.

Cost

Suite wall-clock at j=1:

About +55 s overhead, dominated by plan_stability_tpch_fuzzed which
runs 222 commands.

Scope of Work

• buildscripts/resmokeconfig/suites/query_golden_join_optimization_plan_stability.yml — add internalJoinPlanSamplingSize: 10000 (with comment explaining why).
• jstests/query_golden/expected_output/plan_stability_tpch_fuzzed — regenerate.
• jstests/query_golden/expected_output/plan_stability_subjoin_cardinality.md — regenerate.

Acceptance Criteria

• Suite YAML carries the explicit sample-size override with a comment.
• Both regenerated baselines committed.
• All three tests in the suite pass at the new baselines.
• SERVER-123070's plan_stability join-opt diff disappears once that branch rebases onto this change.

Background

See full analysis attached as a comment: per-query sample-size sweep, why
the default 1000-sample plan is an under-sampling artifact, and the
relationship to SERVER-123070.