udyamlinc-api/DEPLOYMENT.md

# Deploying UdyamLinc

Two artifacts, one host.

| Artifact | Built from | Output |
|---|---|---|
| API | `UdyamLinc-Backend` | `target/udyamlinc-backend.jar` (~80 MB, executable) |
| UI | `UdyamLinc-UI` | `dist/` (static files, ~880 kB) |

The UI bundle is built with `VITE_UDYAM_API_BASE_URL=/api/v1/udyam` — a **relative**
path. It therefore contains no hostname, serves correctly from any domain, and is
same-origin with the API. That last point is the one worth keeping: same-origin means
CORS never applies, and CORS is the single largest source of "worked locally, broken on
the server" in a split deployment.

The only requirement on the web server is that it does two things:

1. serve `dist/` with an SPA fallback — unknown paths return `index.html`, because
   React Router owns `/officer/**`, `/nabard/**` and `/app/**`, not the file system;
2. pass `/api/**` through to the Spring Boot process on port 8082.

Any of nginx, Apache httpd or Tomcat can do both. Pick the section that matches your box.

---

## 1. The API (identical in all three cases)

The jar embeds Tomcat 11. It does not need an application server installed, and it does
not go in `webapps/`.

```bash
java -jar udyamlinc-backend.jar
```

Configuration is entirely by environment variable — nothing needs editing inside the jar:

```bash
export UDYAM_DB_URL=jdbc:postgresql://localhost:5432/UdyamLinc_DEMO
export UDYAM_DB_USERNAME=postgres
export UDYAM_DB_PASSWORD=<the real one>

# Set this. The default is committed to the repository, so it is public, and a
# signing key anyone can read signs anyone's token.
export UDYAM_JWT_SECRET=<a long random string>
```

`spring.jpa.hibernate.ddl-auto=update` builds the schema on first boot, so the database
only has to exist and be empty:

```sql
CREATE DATABASE "UdyamLinc_DEMO";
```

### As a systemd service

```ini
[Unit]
Description=UdyamLinc API
After=network.target postgresql.service

[Service]
User=udyamlinc
WorkingDirectory=/opt/udyamlinc
Environment=UDYAM_DB_URL=jdbc:postgresql://localhost:5432/UdyamLinc_DEMO
Environment=UDYAM_DB_USERNAME=postgres
Environment=UDYAM_DB_PASSWORD=change-me
Environment=UDYAM_JWT_SECRET=change-me
ExecStart=/usr/bin/java -jar /opt/udyamlinc/udyamlinc-backend.jar
SuccessExitStatus=143
Restart=on-failure

[Install]
WantedBy=multi-user.target
```

### First boot takes a while, and the failure mode is silent

`UdyamDataSeeder` populates 263 enterprises across 8 partner banks and 16 branches, and
it runs as **one transaction**. Nothing is visible in the database until it commits.

This has bitten us once already: the machine slept mid-seed, the transaction stayed open
overnight, and every figure read from the console during that window was the pre-seed
state — reported confidently, with no warning anywhere. **Watch for the seeder's final
log line before trusting any number on screen**, and check its timestamp rather than
just its text.

Verify the API is up:

```bash
curl -s localhost:8082/api/v1/udyam/auth/regional-offices
```

That endpoint is deliberately public — the login page reads it before anyone has signed
in — so it is a clean liveness check.

---

## 2a. Front door: nginx

`UdyamLinc-UI/nginx.conf` is ready to use as-is. Copy `dist/` to the document root and
the conf to `/etc/nginx/conf.d/`. It already carries the `/api/` proxy block, the SPA
fallback, gzip, and asset caching.

Adjust `listen 5173` to `80`/`443` for a real deployment.

---

## 2b. Front door: Apache httpd

Enable the modules first — on a stock install `proxy_http` and `rewrite` are usually off:

```bash
a2enmod proxy proxy_http rewrite headers
```

Virtual host:

```apache
<VirtualHost *:80>
    ServerName udyamlinc.lincpay.tech
    DocumentRoot /var/www/udyamlinc

    # ── API ─────────────────────────────────────────────
    # The backend's context-path is already /api/v1/udyam, so the prefix passes
    # through unchanged rather than being rewritten.
    ProxyPreserveHost On
    ProxyPass        /api/  http://127.0.0.1:8082/api/
    ProxyPassReverse /api/  http://127.0.0.1:8082/api/

    # Ingestion and the three-year backfill are long-running admin calls. The
    # default proxy timeout cuts them off mid-run and reports a gateway error
    # over a job that is still working.
    ProxyTimeout 300

    <Directory /var/www/udyamlinc>
        Require all granted
        Options -Indexes

        # ── SPA fallback ────────────────────────────────
        # Anything that is not a real file, not a real directory, and not the
        # API is handed to index.html for React Router to resolve. Without the
        # !^/api/ guard this rule swallows API calls and returns HTML to axios,
        # which surfaces as a JSON parse error rather than a routing problem.
        RewriteEngine On
        RewriteCond %{REQUEST_URI} !^/api/
        RewriteCond %{REQUEST_FILENAME} !-f
        RewriteCond %{REQUEST_FILENAME} !-d
        RewriteRule . /index.html [L]
    </Directory>

    # Fingerprinted assets are immutable; index.html must never be cached or a
    # deploy cannot reach browsers that already have it.
    <LocationMatch "^/assets/">
        Header set Cache-Control "public, max-age=2592000, immutable"
    </LocationMatch>
    <LocationMatch "^/index\.html$">
        Header set Cache-Control "no-cache, no-store, must-revalidate"
    </LocationMatch>
</VirtualHost>
```

---

## 2c. Front door: Tomcat

Two things are worth separating here, because "Tomcat" can mean either.

### Serving the UI from Tomcat — fine

Copy `dist/` into `$CATALINA_HOME/webapps/ROOT/` and add the SPA fallback via the
rewrite valve. In `conf/Catalina/localhost/ROOT.xml` (or the `<Context>` element):

```xml
<Context>
    <Valve className="org.apache.catalina.valves.rewrite.RewriteValve" />
</Context>
```

with `webapps/ROOT/WEB-INF/rewrite.config`:

```
RewriteCond %{REQUEST_URI} !^/api/
RewriteCond %{DOCUMENT_ROOT}%{REQUEST_URI} !-f
RewriteRule ^(.*)$ /index.html
```

The API still runs as its own `java -jar` process on 8082, and Tomcat proxies to it —
or you put httpd/nginx in front of both.

### Deploying the API *as a WAR* into Tomcat — do not do this today

It is possible but it is not a same-day change, and the constraint is hard rather than
stylistic:

- This project is **Spring Boot 4.0.1**, which is built on Jakarta EE 11 / Servlet 6.1.
  That requires **Tomcat 11**. On Tomcat 9 or 10 the application does not start —
  Tomcat 9 is `javax.*` and cannot load a `jakarta.*` application at all.
- It needs `<packaging>war</packaging>`, `spring-boot-starter-tomcat` moved to
  `provided`, and `UdyamLincApplication` extending `SpringBootServletInitializer`.
- The context path moves from the configured `/api/v1/udyam` to the WAR's deployed
  name, so the proxy rules above stop matching.

The jar already embeds Tomcat 11, so running it directly gives the same container with
none of that risk. **Recommendation: keep the jar.**

---

## Checklist before showing anyone

- [ ] `curl localhost:8082/api/v1/udyam/auth/regional-offices` returns JSON
- [ ] The seeder's final log line has appeared, with a plausible timestamp
- [ ] The site loads, and a hard refresh on a deep link such as `/nabard/geography`
      returns the app rather than a 404 — this is the SPA fallback, and it is the
      single most common thing to get wrong
- [ ] The browser network tab shows `/api/v1/udyam/...` calls returning 200, not HTML
- [ ] `UDYAM_JWT_SECRET` is not the committed default

## Restoring the demo database — do this rather than seeding fresh

Verified on 2026-08-15: a freshly seeded database and the working demo database are
**not the same demo**. The seeder writes the plan, which is 129 enterprises; the working
database carries 263, along with three years of real Agmarknet prices that only a
`DATAGOV_API_KEY` and a long backfill can produce.

| | Restored dump | Fresh boot + bootstrap |
|---|---|---|
| Enterprises | 263 | 129 |
| RED | 50 | 14 |
| Trip-wires firing | **6** | **5** — TW-05 silent |
| Open alerts | 344 | 70 |
| Real Agmarknet rows | 8,292 | 0 |

TW-05 reads a price anomaly against a three-year seasonal mean. Generated series are
calibrated to sit near the real data's typical z of 0.44, so they legitimately do not
breach a 1.5 threshold — the trip-wire is not broken on a fresh database, it simply has
nothing anomalous to find.

```bash
createdb -U postgres UdyamLinc_DEMO
pg_restore -U postgres -d UdyamLinc_DEMO --no-owner --no-privileges udyamlinc_demo.dump
```

Then start the jar. The seeder is idempotent and will leave the restored book alone.

`--no-owner --no-privileges` matters when the demo server's Postgres role is not
`postgres`; without them the restore fails on every `ALTER ... OWNER TO` line.

Verify:

```sql
select count(*) from enterprise;   -- expect 263
select source, count(*) from market_signal group by source;  -- expect an AGMARKNET row
```

The sequence below is the fallback for a server that has no dump to restore.

## Populating a fresh database — the demo sequence

The seeder writes the portfolio, the ledger and the forecasts. It does **not** write
loan history, UPI inflows or the no-feed price series: those are admin calls, so a fresh
database comes up without them and several screens look thinner than they should.

**One call does all of it.** After the seeder's final log line:

```bash
BASE=http://localhost:8082/api/v1/udyam
TOKEN=<a supervisor token from POST $BASE/auth/login>

curl -X POST -H "Authorization: Bearer $TOKEN" "$BASE/admin/bootstrap/demo?months=24"
```

That runs six steps in dependency order and reports each one:

| Step | What it writes |
|---|---|
| `loan-history` | 24 monthly DPD snapshots per loan, `SEEDED_HISTORY` |
| `upi-inflows` | 26 weeks per enterprise, `SIMULATED` — the Account Aggregator stand-in |
| `prices-no-feed` | milk, egg, fodder, sugarcane — the four with no API anywhere |
| `prices-all` | three years for **every remaining basket pair**, so `price_z_score` has a seasonal mean and TW-05 can fire without a data.gov.in key |
| `features` | the 15 derivations across the book |
| `scoring` | stress scores, bands, trip-wires, alerts |

It is **idempotent** — re-running is safe, and each step commits on its own, so a
failure in one leaves the others intact and the response names which failed.

The individual endpoints remain if you want to run a stage on its own:
`/admin/loan-history/generate`, `/admin/upi/simulate`,
`/admin/ingestion/simulate/no-feed`, `/admin/ingestion/simulate/all`,
`/admin/features/compute`, `/admin/scoring/run`.

Verify with the coverage endpoints, both of which report synthetic and real separately:

```bash
curl -H "Authorization: Bearer $TOKEN" "$BASE/admin/loan-history/coverage"
curl -H "Authorization: Bearer $TOKEN" "$BASE/admin/upi/coverage"
```

### All of it is labelled, and none of it can overwrite a fact

Every row written above carries its provenance in the table — `SEEDED_HISTORY` on loan
records, `SIMULATED` on UPI and price rows — and the simulators refuse to overwrite a
non-simulated row even when asked to. Real data can land beside this at any time and
stays distinguishable forever. The console surfaces the distinction rather than leaving
it to a slide.

Two consequences to state out loud rather than discover later:

- **`loan-history/coverage` reports `trainable: false`** while no real extract exists,
  and that is correct. The seeder derives DPD from the risk band and the UPI simulator
  derives inflow drift from the same band, so a model trained on this would learn the
  seeder's own rule and validate beautifully while knowing nothing. The synthetic data
  is here to exercise the pipeline and fill the console, not to train.
- **No external dependency at all.** The bootstrap needs no `DATAGOV_API_KEY`, no
  Account Aggregator consent and no bank connection: every screen fills from generated
  data. If a key *is* present, run the real ingestion first and the price simulator will
  skip any pair that already holds real rows, filling the gaps around the real series
  rather than across it.

## What is deliberately off

`udyam.ingestion.enabled` is `false`, so a demo boot never calls a government endpoint,
and `DATAGOV_API_KEY` is unset. The portfolio is seeded, which is enough for every
screen. To refresh from live sources, set the key and run, in order:

```
POST /admin/ingestion/run
POST /admin/features/compute
POST /admin/scoring/run
```

Space bulk ingestion runs a couple of minutes apart — data.gov.in meters by quota per
window, so a run started too soon after another returns 429 on every pair at once and
records the whole ingestion as failed with zero rows.