Studio 2.2.7 capacity validation¶
Date: 2026-09-15. Measurements use the candidate source over real loopback HTTP on the development host. The benchmark uses a fixed daytime clock and private temporary storage; it does not consume production workspace slots.
| Measurement | Observed |
|---|---|
| Open empty workspaces | 256 |
| Live waiting-room tickets | 1,024 |
| Process RSS with both | 225.97 MiB |
| Public background threads | 2 (sampler and shared dispatcher) |
| Idle CPU, percent of one core | 0.50% |
| Workspace creation median / p95 | 2.45 / 6.00 ms |
| Workspace creation maximum, including cold initialization | 790.26 ms |
| Status reads, 100 requests at concurrency 8, median / p95 | 11.21 / 20.19 ms |
| Status reads maximum | 22.76 ms |
| Admission beyond the waiting-room limit | HTTP 429, bounded queue |
The full Python regression was running separately during the benchmark. These are empty-workspace admission and status measurements, not a guarantee of 256 simultaneous training jobs or an Internet latency SLA. The production deployment retains one compute slot, two heavy API operations, a 2 GiB shared tmpfs, 256 MiB per workspace and its existing CPU/RAM/container limits.
Machine-readable measurement. Reproduce from the repository with:
Regression coverage¶
The 59-test focused public boundary/GPU/telemetry/admission run passed, including actual ASGI requests for FIFO admission, lost-response replay, network/global limits, queue credential isolation, cancellation racing with admission, abandoned-ticket expiry, exact noon cleanup and reopening, low-storage waiting, and 32 idle workspaces without per-visitor threads. Existing real CPU training and private GPU lease/cancellation regressions also passed.
The 239 frontend tests passed, including waiting-position rendering, automatic admission, cancellation and rejection of a late session response after leaving the queue. A busy admission boundary is retried automatically even before a queue position has been allocated. Type checking, lint and regenerated API types passed.
Chromium connected to the real candidate HTTP API under a controlled two-slot policy. Five independent browser contexts verified FIFO positions, automatic entry after another user confirmed End workspace, cancellation, preservation of the ticket across refresh, and recovery after a dropped network request. Desktop, 390 px and 360 px screenshots had no document overflow or JavaScript errors. See browser evidence.
The new pinned GPU image reports OpenDPD 2.2.7 and PyTorch 2.14.0+cu132 and passed a real CUDA tensor operation on the host GPU. The full local Python regression passed 963 tests; a fresh 59-test boundary run also passed after the final scheduler guards. Distribution wheel/sdist metadata passed validation. Production workflow evidence is added to the release record after verification. Synthetic PA/DPD runs validate the software workflow; they are not measured RF hardware results.