Case study 01 · PostgreSQL
When an index changes the plan but not the outcome.
A local-first performance qualification story about rejecting a plausible fix, measuring a stronger alternative, and refusing to convert local evidence into a production claim.
Synthetic dataDockerized PostgreSQLThree randomized repetitionsNo production access
STAR-E narrative
01Situation
A recurring reporting query was slow.
A synthetic PostgreSQL workload repeatedly joined and aggregated completed-order data. The request path showed high reporting latency and appeared to invite an index-first remedy.
02Task
Select a remedy that survived measurement.
The goal was not to make the execution plan look better. It was to compare plausible remedies against reporting latency, mixed-workload behavior, writes, refresh behavior, and recovery concerns.
03Action
Preserve the failed candidate and continue.
The index candidate changed the plan but did not materially improve p95. The workflow recorded that negative result, compared additional candidates, and repeatedly qualified a monthly reporting read model.
04Result
The read model won locally.
Across three randomized local repetitions, the read model produced a 99.09% median reporting improvement and a 97.05% median mixed-workload improvement.
05Evidence
The result remained bounded.
The evidence supported accept_for_local_demo_only. Representative production workload, capacity, refresh ownership, cost, failure testing, security review, rollback planning, and a controlled change window were absent.
Candidate record
CandidateObserved resultDecisionScope
Index
Execution plan changed; p95 did not materially improve.
Rejected
Local lab
Monthly reporting read model
Major latency improvement with refresh and stale-data recovery checks.
Accepted
Local demonstration only
Production rollout
Representative workload, cost, ownership, and change-window evidence absent.
Not approved
Out of scope
What this demonstrates
A failed fix can be valuable evidence.
- Baseline and measurement method remained consistent.
- Candidate remedies were compared rather than assumed.
- Negative results were preserved in the decision record.
- Refresh and stale-data recovery entered the qualification.
- The accepted claim matched the available environment.
What this does not prove
Local qualification is not production approval.
- Production workload safety
- Capacity and cost under real demand
- Operational ownership and supportability
- Security and change authorization
- Rollback behavior in the target environment
Behind the case
Cloud DBRE Reference Stack
The local reference implementation connects evidence collection, deterministic analysis, candidate qualification, change artifacts, verification, and human review.