mongodb: measure missing_params write, storage and index-churn cost at scale #2006

Summary

Measure the write, storage and index-churn cost of the MongoDB missing_params array and idx_resources_missing_params index (#1984) at scale. The #1984 write benchmark used 1,000 resources on a single node. The 12M storage figures come from an earlier version of the implementation.

Background

• From feat(mongodb): index missing search parameters on resources #1984, on 12M resources: about 10.5 GB for the new index, plus 1.23 GB of extra resource data, about 15% of baseline. That is about 17.6 absent parameter names per resource. Every create/update rewrites all of them.
• The layout decision is settled: one array per resource (search(mongodb): accelerate :missing=true searches with indexed per-resource metadata #1908, question 1). This issue is about measuring the cost and reducing it where cheap, not about choosing a different layout.

Proposed work

• Measure at 1M+ resources, with concurrent writers and, if available, a sharded cluster:
  • Create, update and $bulk-submit throughput.
  • p50/p99 write latency.
  • Index size, and WiredTiger cache pressure on resources.
• Run both with and without the array (main vs feat(mongodb): index missing search parameters on resources #1984).
• Evaluate cheap reductions:
  • Exclude parameters that can never be absent, or that are rarely queried with :missing.
  • An opt-in setting (e.g. HFS_MONGODB_MISSING_METADATA) for deployments that do not use :missing=true.

Acceptance criteria

• Results posted on this issue, with methodology.
• A recommendation on whether an opt-out or a reduced parameter set is needed.

Refs #1908, #1984