Bounded Workflow Progress Preparation: WSL2 Linux Evidence¶

This report records the required preparation-only evidence matrix for ADR-0005 and issue #140. The authoritative measurements are in the adjacent JSON artifact.

• Collected: 2026-07-21T13:24:05Z
• Source revision: c42fb22634712e52d8aee74c86a62fea459da5e8
• Implementation digest: 2d32b2b2175706aaecd1a3348134f0e27ab61f4f489a8820c57c2fc3d8de3ce9
• JSON SHA-256: 8ea7b9e22182e9bbaf9962eedaea9dc4872893a34538f55adcbe9da23d3edf73
• JSON size: 27,200 bytes
• Result: all six required cases and the forced-termination cleanup control passed

Scope¶

The benchmark exercises the non-production SQLite spill prototype directly. Every case runs in a fresh child process, consumes topology and initial-detail generators once, prepares canonical V1 output, and removes its private workspace before the parent accepts the result.

The run did not start Ray, access a Django database, contact Kubernetes, change a runtime writer, or activate schema v3. It is evidence for the preparation contract, not for end-to-end workflow latency or production admission capacity.

Environment¶

The checkout was created by Windows Git. The outer WSL command supplied core.autocrlf=input and core.filemode=false through invocation-scoped GIT_CONFIG_* variables so WSL evaluated the same index and working-tree content as Windows Git. This did not rewrite the checkout or change repository configuration. Both source snapshots recorded dirty=false, the same revision, and the same implementation digest.

Selected resource profile¶

All cases reported the same effective settings and V1 output limits. The retained query plans were five direct scans or primary-key searches; none used a temporary B-tree or materialization.

Results¶

Memory and spill values use MiB. RSS delta subtracts the current RSS captured at the start of each fresh child process; the JSON also retains the process high-water baseline and sampler output.

Interpretation¶

Once retained output reached the V1 caps, memory stayed effectively flat while observed input grew:

• sparse RSS delta was 154.3, 156.6, and 156.6 MiB, while tracemalloc peak stayed at 64.8 MiB;
• high-edge RSS delta was 253.4, 253.2, and 252.9 MiB, while tracemalloc peak stayed at 101.2 MiB; and
• retained output stayed at 25,000 nodes and detail rows, with either 24,999 sparse edges or the 100,000-edge V1 cap.

The external spill grew with observed input, from 14.5 to 71.4 MiB for sparse cases and from 21.0 to 136.6 MiB for high-edge cases. That is the intended trade: exact duplicate and reference state scales on bounded local disk rather than as Python identity collections. Wall time also grew with observed input, so these measurements establish a memory contract and decision baseline, not a latency SLO.

Source and cleanup controls¶

• Source snapshots before and after the matrix were identical and clean at revision c42fb22634712e52d8aee74c86a62fea459da5e8.
• Every normal case reported both worker-context and parent-watchdog cleanup as removed; no scenario root remained.
• The forced-termination child published its parent-owned readiness nonce while the SQLite workspace was open, was killed with return code -9, and produced no durable candidate.
• The parent removed the forced child's workspace and scenario root, and the configured benchmark parent was empty in an immediate post-run filesystem check.

Decision and limitations¶

Together with the 76 focused canonical, failure, cancellation, and cleanup tests that passed on both Windows and Linux/WSL2 at the measured revision, this matrix satisfies ADR-0005's issue-#140 evidence gate for accepting the exact spill-backed preparation contract. It does not deliver or authorize production writer activation. Issue #141 still owns topology integration, issue #142 owns composite topology/detail lifetime and final production memory proof, and issue #79 owns live wire, mailbox, producer, concurrency, and node-local spill admission bounds.

This is one run on one WSL2 host and filesystem, without repetitions or concurrent preparers. Process RSS is a child-lifetime high-water measurement and includes Python startup; the recorded baseline makes that scope explicit. Production integration must repeat fault and capacity evidence for its configured spill root, janitor, aggregate quota, and concurrency policy.

Reproduction¶

The native-Linux form is documented in Performance. This WSL run used the following checkout-local command; the GIT_CONFIG_* values are only needed when a Windows-created checkout otherwise appears dirty to WSL Git.

cd /mnt/d/Repos/django-ray
mkdir -p /tmp/django-ray-issue140-benchmark
GIT_CONFIG_COUNT=2 \
GIT_CONFIG_KEY_0=core.autocrlf \
GIT_CONFIG_VALUE_0=input \
GIT_CONFIG_KEY_1=core.filemode \
GIT_CONFIG_VALUE_1=false \
PYTHONPATH=/mnt/d/Repos/django-ray/src:/mnt/d/Repos/django-ray \
/tmp/django-ray-issue140-venv/bin/python \
  scripts/benchmark_workflow_progress_preparation.py \
  --required-scale \
  --timeout-seconds 7200 \
  --workspace-parent /tmp/django-ray-issue140-benchmark \
  --output \
  docs/benchmarks/workflow-progress-preparation-sqlite-wsl2-linux-2026-07-21.json