rx-dag¶

What it does¶

rx-dag (the service is named ndc-loader) keeps a complete local copy of the FDA's drug product databases and serves fast lookups from it:

the NDC Directory: every listed drug product and package, by National Drug Code;
Drugs@FDA: applications, approved products, submissions, marketing status and therapeutic-equivalence codes.

It replaced live calls to openFDA's NDC API, which was slow and rate-limited. Every NDC lookup in the stack is answered from rx-dag's copy.

A Go service backed by PostgreSQL. Every night at 03:00 it:

discovers and downloads the FDA bulk files;
extracts and parses them;
loads each table into a staging copy;
checks the row count: if a table shrank by more than 20% compared with the last load, the load is rejected as suspicious;
swaps the staging copy in atomically, so readers never see a half-loaded table;
records a checkpoint per table, so an interrupted load can resume.

A load can also be started by an admin request.

Table	From	Holds
`products`	NDC Directory	One row per product NDC: names, labeler, dosage form, route, active ingredients, classes
`packages`	NDC Directory	One row per package NDC: package description, marketing dates
`applications`, `drugsfda_products`	Drugs@FDA	Approval applications and their products
`submissions`, `marketing_status`, `te_codes`	Drugs@FDA	Submission history, marketing status, therapeutic-equivalence codes

Lookups accept any NDC format (with or without dashes, 10 or 11 digits) and normalize it.

All behind an API key, internal only (cash-drugs is the caller):

Lookup: one product by NDC, and its packages.
Search: full-text search across products.
openFDA-compatible: the same request and response shape as openFDA's /drug/ndc.json, so callers written for openFDA work unchanged. This is what drug-gate's NDC lookup uses through cash-drugs.
Stats and admin: dataset counts, start a load, load status with checkpoints.
Operations: health, version, Prometheus metrics, Swagger UI.

Readers always see the last good load: a new load only replaces a table after it fully loaded and passed the row-count check. A failed or suspicious night leaves yesterday's data in place.
Resumable loads: per-table checkpoints let an interrupted load continue.
Freshness downstream: cash-drugs keeps rx-dag answers for at most an hour, so a nightly reload reaches the public API within about an hour.