Backend, Kubernetes, database, and infrastructure notes from real engineering work.
36 articles
A detailed retrospective on building PostgreSQL Consistency Lab: deterministic transaction failures, isolated sandboxes, managed actors, 16 real scenarios, bounded workloads, PostgreSQL observability, a synchronized dashboard, and one shared Go control plane for CLI, REST/SSE, and MCP.
Activity Monitor showed DevBerth using 69.2% CPU with no window open. This is how I followed the work through Instruments and brought the same Release build down to 0.118%.
DevBerth started as a port viewer. It turned into a native macOS runtime manager once I realized that seeing a process and safely controlling it were two very different problems.
How a research dashboard recorded each candidate’s first failing gate, then a frozen multi-horizon diagnostic confirmed no_trade before test access or model training.
How coding agents move developer leverage from typing code to defining systems, constraints, review, and responsibility.
A forensic analysis of timing, contract, direction, fold, and cost fragility, followed by MBP-1 spread, latency, and adverse-movement calibration for OHLCV candidates.
The final control-system arc: accountable reward tuning, dual-cluster comparison, bounded Kubernetes actions, semantic dashboard evidence, and honest energy limits.
How deterministic baselines, feature IC, and a strategy zoo selected no_trade, then train, concentration, and cost-stress gates rejected the first validation-positive candidates.
How model signals became separate agent proposals, conservative referee decisions, dashboard-visible state, and a live Kubernetes control loop.
How past-only contract selection and session filtering produced an MES/ES silver dataset, then a strict matrix with physically separated features, labels, metadata, and splits.
How schema validation, temporal label repair, multi-horizon XGBoost models, and explicit action thresholds established a trustworthy control signal.
How a refusal-first research philosophy became a concrete data boundary with preserved Databento jobs, hash verification, bronze conversion, and quality reports.