Skip to content

Prove Cloudflare WordPress fleet performance, resilience, and cost #2079

Description

@chubes4

Parent: #1970

Goal

Establish production readiness for an elastic preview fleet through dimension-based performance, concurrency, failure, multi-region, and cost evidence.

Workloads

  • 100-site and 1,000-site allocation/provisioning campaigns
  • concurrent artifacts of varied page count, decoded size, asset volume, and runtime complexity
  • anonymous publication reads from multiple regions and cache states
  • warm/cold admin, REST, edit, media, and subsequent-import traffic
  • D1/R2/queue latency, throttling, duplicate delivery, Worker eviction, timeout, and partial-failure injection
  • long-running soak with expiration, deletion, reclamation, and allocation reuse

Required evidence

  • p50/p95/p99 latency by API stage and public/dynamic read path
  • allocation and ready-site throughput under bounded concurrency
  • PHP-WASM boot/execute/persist memory and CPU high-water marks
  • queue age, retries, dead letters, publication lag, and recovery time
  • D1 rows/queries, R2 requests/bytes/storage, Worker CPU, and estimated cost per site lifecycle
  • isolation and invariant violations, not only aggregate success counts

Acceptance criteria

  • Publish explicit SLOs and pass/fail budgets before running the acceptance campaign.
  • No cross-site state, cache, credential, operation, or publication leakage occurs.
  • Failure injection proves committed state and previous publications remain recoverable.
  • Throughput does not degrade linearly with configured site count.
  • Public cached-read latency remains independent of PHP-WASM and coordinator load.
  • Known gaps are reported by semantic dimension: lifecycle, state transitions, resources, compatibility, and regional behavior.
  • Evidence is reviewer-resolvable and linked to this tracker.
  • No production deployment or load test is performed without explicit authorization.

AI assistance

OpenAI GPT-5.6 Sol via OpenCode translated the current two-site evidence and measured latency gaps into this fleet acceptance contract with Chris Huber.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions