A landmark day for platform observability and data pipeline reliability. We shipped a fully wired Jobs monitoring page that surfaces all 13+ tracked job types, broke through a major ETL performance ceiling on the SynPUF dataset (17 hours → 14 minutes for drug_era builds), and closed out a cohort generation audit that uncovered eight discrete bugs across the SQL builders, API layer, and frontend.
A massive day on Parthenon with 193 commits landing across the platform. The headlining work: a near-complete publication/manuscript workflow that takes study analyses all the way to a formatted, auto-numbered document preview, plus a long-overdue rename of the R Analytics Runtime to Darkstar — the name it's been running under in Docker all along.
Today was one of those satisfying days where two major workstreams converged: we pushed the Ares data quality module from skeleton to a fully featured analytics suite with four distinct intelligence phases, and we permanently fixed a cluster of compounding bugs that had been making Achilles characterization runs fragile on large real-world datasets. Both efforts move Parthenon meaningfully closer to being a production-grade OHDSI research platform.
After months of a perpetually red CI pipeline, today marks a turning point for Parthenon: 92 commits, a full-spectrum codebase review, a complete AtlanticHealth patient synthesis pipeline, and — most satisfying of all — every CI job green. Here's how we got there.
Eighteen Docker services. Three environment files. A reverse proxy with auto-TLS. Database admin GUI. Container management dashboard. Enterprise SSO. And if you want the full stack? One command:
python3 install.py --with-infrastructure
This is the story of how we built Acropolis — the infrastructure layer that turns Parthenon from a research application into a production platform — and what we learned when we decided to ship it inside the same repository.
Every platform has a weak link. For Parthenon, it was the R container.
PHP handled 200 concurrent API requests without breaking a sweat. Python served AI inference with async workers. PostgreSQL managed million-row queries across six schemas. Redis cached sessions at sub-millisecond latency. And then there was R — single-threaded, fragile, running bare Rscript as PID 1 with no supervision, no timeouts, and a health check that lied.
This is the story of how we tore it down and built Darkstar — a production-grade R analytics engine that runs OHDSI HADES analyses concurrently, recovers from crashes automatically, and executes 35% faster than the container it replaced.
A big architectural day on Parthenon: we shipped the new Workbench launcher experience and drafted the formal plan to collapse our multi-database mess into a single, schema-isolated parthenon database. A noticeably cleaner codebase on the other side.
A massive 116-commit push today centered almost entirely on maturing the Evidence Investigation workbench — from retiring the old FinnGen UI to hardening the investigation experience with proper navigation, KPI metrics, URL-synced state, and ARIA accessibility. We also landed multi-dataset support in Morpheus and set the stage for one of the most requested features on the roadmap: volcano plots powered by the newly-renamed Darkstar R runtime.
The R runtime was the single most fragile component in the entire Parthenon stack. Every other service — PHP, Python AI, Solr, Redis, PostgreSQL — could handle concurrent requests gracefully. The R container could not. A single CohortMethod estimation on 1 million patients takes 5-30 minutes. During that time, the entire R process was locked — health checks timed out, status queries hung, and any other analysis request queued behind it with no feedback. This devlog covers the six-phase hardening effort that replaced the entire R runtime infrastructure in a single day.
What started as "Abby's responses are slow" turned into an 18-hour deep dive that touched every layer of the AI stack — from GPU driver backends to embedding model race conditions to the fundamental question of why a 4-billion-parameter medical LLM was confidently inventing researcher names. By the end, Abby went from 15-25 second hallucinated responses to 2-5 second grounded answers backed by 167,000 vectors of medical knowledge — and we found that an 8-megabyte systemd memory lock was silently killing 25% of all GPU inference requests.