Feature plan: Step 1 · Step 2 · Step 3 · Step 4

Debug mode: enabled Effort study: railspike-effort / v3/qdb_memory_audit; standing P2 ledger Extended estimates: enabled

Original Query

Can you also please audit for anywhere else this may be an issue?

Understood Intent

Find other causes of PHP memory exhaustion after two QDB build-static failures, with a fixed 128 MiB host limit. The requested deliverable is an audit, not approval for a broad implementation or production changes.

Problem: Several operations limit displayed/output records only after allocating corpus-sized input collections.

Concepts touched: Derived read-model rebuild, static publication, public quote listings and activity evidence, used unchanged; inventory entries proposed if remediation proceeds. Existing reaction records used unchanged.

Option A — Source-only audit (selected)

Estimate: 0.20

Inspect allocation patterns and rank likely risks without workload reproductions.

• Smallest scope; uncertainty remains about actual peaks and call-path reachability.

Option B — Focused audit with executable probes

Estimate: 0.40

Trace the build and shared readers, then reproduce selected risks using disposable data at the existing memory limit.

• Previously recommended; probes already performed remain supporting evidence, not a requirement for further work in the selected A scope.

• Synthetic corpus construction and interpretation drive effort; live workload remains unknown.

Option C — Application-wide capacity study

Estimate: 0.90

Measure every major command/request family, concurrency and native/child-process memory on production-representative corpora.

• Broadest evidence; requires additional representative data and environment coverage. This audit does not claim that outcome.

Findings and selected scope

Audit findings and reproductions record ten findings. Rebuild and QDB rendering fail with 40000 synthetic quotes without padding; offline selection alone peaks at 94.01 MiB; large commit manifests, high-traffic reads and malformed headers expose further risks. The archived probe source preserves repeatable cases.

Selection, 2026-10-30: The user chose Option A and requested a separate plan for expanding to Option C. This task owns the source inventory, ranked allocation risks and remediation boundaries. The diagnostic inventory is complete. Existing probes remain evidence; no further workload campaign is included here. The broader measurement/acceptance work is proposed in Application capacity study (pending separate assessment: application_capacity_study_step1_solution_assessment.md). No fixes or next-step approval are implied by this selection.

Budget proposal carried to the expansion: Ordinary peaks at or below 16/48 MiB and test ceilings of 32/64 MiB for public/maintenance work, plus scaling, edge-case and recovery measurements. These remain proposals, not verified production limits. The original 128 MiB failures and tiny fixture's 6 MiB peak remain unchanged evidence. Never meet budgets by dropping records or changing content limits.

Overlap: M02/M04/M09 seed the separate load-before-selection assessment (pending separate assessment: load_before_selection_audit_step1_solution_assessment.md), which owns codebase-wide unnecessary-work discovery. This task retains the original memory findings; the capacity-study expansion owns proposed broader measurements. Link shared causes and fixes once. The selection audit's own option remains unchanged.

Study checkpoint

Scope: v1/B, baseline cfb8364f6db03cc09e27cd26afb93c109c723a08; R1/P2. Intake record preserves prior inspection exposure and missing pristine estimate. Step 1 handoff: 2026-10-30 UTC, 0.40 total reaffirmed; all three frozen hashes verified. Codex single exposed rater; options are whole-audit projections, not additive remediation estimates. No independent rating, measured labor, implementation outcome or maintainer acceptance claimed. Related direct fixes remain outside the study.

Selection checkpoint, 2026-10-30: v2/A supersedes the earlier B recommendation prospectively. Preserve A's original 0.20 alternative estimate and v1/B's 0.40 snapshot. Current whole-cycle projection remains 0.40 because already-performed probes and documentation remain part of this cycle; do not retroactively recast it as a source-only execution. R1/P2, baseline and exposed rater unchanged. Expansion is separately proposed as v3/application_capacity_study; neither outcome is counted accepted.

Approval, 2026-10-30: User explicitly clarified that Approved Step 1 applies to this QDB memory audit, selected Option A. Proceed to Step 2 for v2/A. The application capacity study and load-before-selection assessment remain unapproved at Step 1; no application fixes or measurement expansion are authorized.

— Codex · 2026-10-30 UTC