Broker capacity model

Solace PubSub+ inventory, workload experiments, and saturation-boundary modeling.

Run the autonomous 100-experiment grid

./run-grid run --directory data/grid-100

Run the same command to resume. Ctrl-C stops and checkpoints; ./run-grid status --directory data/grid-100 shows saved progress. Runner documentation covers the exact grid, detector, output files, timing controls, and current scope. This is a standalone program with no LLM dependency. It targets the dedicated 5K and preserves shared demo brokers.

Current status

The initial discovery pass covers five services in the Solace SEs workspace whose ownedBy value exactly matches Raphael Caillon's verified user ID. See docs/initial-findings.md and data/initial_snapshot.json.

The collector:

• accepts only explicitly exported read-only SEMP credentials;
• rejects every broker whose owner does not match --owner-id;
• performs only HTTP GET requests;
• follows SEMP cursor pagination;
• removes the temporary credential file even if collection fails;
• writes only sanitized telemetry and metadata.

Run the collector

PYTHONPATH=src python -m broker_capacity.cli /tmp/solace-semp-readonly.json \
  --owner-id usfos7cfqge \
  --output data/snapshot.json

The temporary access file is intentionally not committed and is deleted after each run.

Tests

PYTHONPATH=src python -m unittest discover -s tests

Saturation experiment campaign

The initial snapshot is discovery evidence, not a measured performance boundary. See the reviewed campaign for the 24 baseline configurations, targeted follow-ups, mixed-workload rays, broker ladder, measurement protocol and model-validation criteria. The machine-readable design is campaigns/saturation-v1.json.

The earlier Claude pilot has been stopped. Its pilot brief, pilot status, and review findings are historical context. The experiment list through 5K defines the broader research plan; the terminal runner above implements the current 100-workload SMF grid on dedicated 5K service 4qn20a1u6ny.

The earlier token-only Claude handoff is retained for reference. The terminal grid does not require that supervisor or an LLM. Local process/log state is under ignored .pilot/; no credentials are copied into reports.

The token in ignored .env now includes services:post alongside self-service discovery/management access. See verified token access; token values are never stored in reports.