fleettickets/README.md

# fleettickets

Field-ops **INC ticket** ingestion, geocoding, and read-schema that powers the
**Tickets** map in FleetOps. Extracted from the `tracksolid` repo into its own module
(it previously lived there as migrations 21–23 + `tools/import_tickets.py`).

- **INC** — incident / customer-fault tickets *(this pipeline is **strictly INC**)*
- **CRQ** — new-installation requests *(schema kept, but **out of scope** — not ingested here)*

## What this owns

| Piece | What |
|---|---|
| `migrations/01_tickets_schema.sql` | The `tickets` schema: `tickets.inc` / `tickets.crq` (raw-jsonb-first), `tickets.geo_clusters` + `tickets.geo_locations` gazetteers, geom-resolution trigger, and `reporting.fn_tickets_for_map` (the GeoJSON read function) |
| `migrations/02_import_meta.sql` | `tickets.import_meta` (per-dataset snapshot envelope metadata) + `fn_tickets_for_map` re-defined to expose it as `summary.freshness` (same signature — dashboard_api unchanged) |
| `migrations/03_inc_columns.sql` | Unpacks `tickets.inc.raw` into **typed STORED generated columns** (status, cluster, region, team, owner, sla_status, mttr, lat/lng, is_* booleans, and EAT→`timestamptz` timestamps via `tickets.eat_ts()`). Computed for all rows + auto-populated on every ingest; `raw` stays the source of truth |
| `migrations/04_inc_latlng.sql` | Redefines `latitude`/`longitude` to `COALESCE(feed, ST_Y/ST_X(geom))` so they're **populated from the geocoded position** (feed is always empty); precision per `geo_source` (`location` vs `cluster` centroid) |
| `migrations/05_inc_geography.sql` | Adds `geog geography(Point,4326)` (= `geom::geography`) + GiST index for **routing** — `ST_Distance`/`ST_DWithin`/KNN in real metres (nearest-vehicle, radius search) |
| `migrations/06_inc_mttr_minutes.sql` | `mttr` generated column → integer **minutes** (source is decimal hours); drops the constant `is_alarm`/`is_auto_created`/`is_auto_closed` columns (kept in `raw`). `is_actionable` retained |
| `migrations/07_inc_drop_service_type.sql` | Drops the constant `service_type` column (always `inc`; kept in `raw`) |
| `migrations/08_inc_open_sla_view.sql` | `tickets.inc_open_sla` view — open (`is_actionable`) tickets with **derived SLA** (`hours_open`, `sla_state` vs 48h; clock = `created_at_service` ∥ `first_seen_at`), plus team/cluster/`geog` for dispatch |
| `migrations/09_inc_dashboard_fn.sql` | `reporting.fn_inc_dashboard(cluster, status, window, from, to)` — one JSON payload (`window` / `open` GeoJSON / `closed` GeoJSON / `metrics` / `freshness`) powering the FleetOps live INC map. Open=live, closed=windowed (EAT calendar / custom); filters AND |
| `migrations/10_inc_history_capture.sql` | History for time-series: `tickets.closure_events` (append-only observed closures) + `tickets.inc_daily_snapshot` (per-EAT-day open backlog + flow), populated by `tickets.capture_history()` each ingest. Unlocks **backlog-over-time** |
| `import_tickets.py` | Drains the **incremental INC change stream** from the `isptickets` bucket (`automations/inc/changes/<EAT-timestamp>.csv`), upserting on `ticket_id` oldest→newest; geocodes clusters + INC locations |
| `run_migrations.py` | Applies `migrations/*.sql` in order (ledger: `tickets.schema_migrations`) |
| `shared.py` | Minimal DB/logging helpers (self-contained — no tracksolid dependency) |

## What this does NOT own (stays where it is)

- **The DB** — the `tickets` schema lives in the shared `tracksolid_db`.
- **The read-API** — `dashboard_api` (in the tracksolid stack) serves
  `GET /webhook/tickets`, which calls `reporting.fn_tickets_for_map` (defined here).
- **The frontend** — the Tickets map is a tab in the **FleetOps** SPA (`fleetops` repo).

## Data model (raw-first)

Each row is `ticket_id` + `raw` (the full source record as `jsonb`) + a derived
`geom` / `geo_source`. Everything reads from `raw`, so a change to the source schema
needs no migration. For convenient typed/indexable access, `raw` is also **unpacked
into STORED generated columns** (migration 03) — e.g. `normalized_status`, `cluster`,
`region`, `assigned_team`, `owner`, `sla_status`, `mttr`, `is_actionable`,
`created_at_service`/`closed_at` (as EAT→`timestamptz`). These stay in lock-step with
`raw` automatically (no loader change); `raw` remains the source of truth. `geom` is resolved: **feed** coords (`raw` lat/lng) → **location**
(geocoded `location_name`) → **cluster** centroid → **none**.

Source coordinates are empty in the feed, so geocoding is required:
- `--geocode-clusters` — one coordinate per cluster (coarse fallback).
- `--geocode-locations` — precise per-location for **actionable INC** tickets: strips the
  network codes from `location_name` (e.g. `NW_`, `ADR_MNT_`, `FDT<n>`, `SDUS`), geocodes
  the real place via a **keyed** provider (LocationIQ / OpenCage), and **rejects any result
  >25 km from the cluster centroid** (wrong-city guard). Results cache in
  `tickets.geo_locations`.

### Columns on `tickets.inc`

| Column | Type | Notes |
|---|---|---|
| `ticket_id` | text (PK) | e.g. `WOT0715527` |
| `raw` | jsonb | full source record — the source of truth |
| `normalized_status` · `raw_status` | text | use `normalized_status` for filtering (canonical) |
| `bucket` | text | lifecycle: `closed` / `pending` |
| `is_actionable` | boolean | the open/closed flag (open = `true`) |
| `cluster` · `region` · `location_name` | text | `region` lowercased; `cluster` feeds the gazetteer |
| `assigned_team` · `owner` | text | closure attribution dimensions |
| `sla_status` | text | source `Compliant`/`Breached` — **only meaningful once closed** |
| `mttr` | numeric | **minutes** (source is decimal hours); null until closed |
| `created_at_service` · `scheduled_at` · `closed_at` · `first_seen_at` · `last_seen_at` · `source_created_at` · `source_updated_at` | timestamptz | EAT→UTC via `tickets.eat_ts()`. **lifecycle** = `created_at_service`→`closed_at`; **export bookkeeping** = `first_seen_at`/`last_seen_at`/`source_*` |
| `latitude` · `longitude` | double precision | `COALESCE(feed, geocoded)` — populated from `geom` |
| `geom` | geometry(Point,4326) | display / the map |
| `geog` | geography(Point,4326) | **routing** — metres-accurate distance (GiST indexed) |
| `geo_source` | text | precision: `feed` / `location` / `cluster` / `none` |
| `ingested_at` | timestamptz | when we last upserted this row |

Dropped from the unpacked columns (still in `raw`): `service_type`, `is_alarm`,
`is_auto_created`, `is_auto_closed` (all single-cardinality), plus the ingest-time
drops below. **`reporting.fn_tickets_for_map`** reads from `raw` and serves the map;
**`tickets.inc_open_sla`** is the open-ticket SLA view for dashboards/dispatch.

## Setup

```bash
uv sync
cp .env.example .env        # fill in DATABASE_URL, RUSTFS_*, GEOCODER_*
python run_migrations.py    # apply the schema (idempotent)
```

## Run

```bash
# drain the incremental INC change stream (every new file oldest→newest, then archive)
python import_tickets.py --from-bucket --apply

# geocode (needs GEOCODER_API_KEY)
python import_tickets.py --geocode-clusters  --apply   # coarse, once
python import_tickets.py --geocode-locations --apply   # precise, actionable INC

# from a local CSV instead of the bucket (dev)
python import_tickets.py --inc-csv 2026-06-15T17-00-00.csv --apply
```

Dry-run is the default (omit `--apply`). `import_tickets.py --from-bucket` talks to S3
via **boto3** using the `RUSTFS_*` env (path-style addressing; no aws-CLI dependency).

## Deploy (Coolify)

The repo ships a [`Dockerfile`](Dockerfile) — a small batch worker with no web server.
Coolify builds it and keeps the container alive (`CMD tail -f /dev/null`); the ingest
runs as a **Scheduled Task**, not a system crontab:

- **Command:** `python import_tickets.py --from-bucket --apply`
- **Frequency:** `15 7-19 * * *` (`:15` past each hour, **07:15–19:15 EAT**). This
  Coolify instance runs scheduled tasks in **EAT (Africa/Nairobi)**, so no UTC
  conversion is needed.
- **Env vars** (Coolify → Environment Variables): `DATABASE_URL` (internal DB host),
  `RUSTFS_*` (now the `isptickets` bucket credentials), `GEOCODER_*`.

The watermark makes a run with no new change files a cheap no-op.

For a plain host/VM instead of Coolify, [`run_ingest.sh`](run_ingest.sh) loads `.env`
and runs the ingest; schedule it with a crontab line
(`CRON_TZ=Africa/Nairobi` / `15 7-19 * * *`).

### Bucket cutover (one-time reseed)

When the source provider moves the feed to a new bucket (e.g. `tickets` → `isptickets`),
the stored watermark holds a key from the *old* bucket's stream, whose timestamp may be
newer than the new bucket's first file — which would otherwise be skipped. Point the
`RUSTFS_*` creds + `TICKETS_BUCKET` at the new bucket, then drain it once with `--reseed`,
which ignores the stored watermark and ingests **every** file in `changes/` oldest→newest:

```bash
python import_tickets.py --from-bucket --reseed          # dry-run first
python import_tickets.py --from-bucket --reseed --apply  # commit + archive
```

Upserts are idempotent (`ticket_id` PK, rows never deleted) and the new stream's periodic
full-state re-emissions re-assert current state, so this is non-destructive and converges
even across the cutover gap. After it, the watermark is current — resume normal
`--from-bucket --apply` runs (no `--reseed`). The old bucket is left untouched.

## Notes

- The n8n export writes an **incremental CDC change stream** to
  `automations/inc/changes/<EAT-timestamp>.csv`: a full-state baseline followed by files
  holding only the rows that changed (with periodic full-state re-emissions). No `latest`
  pointer, no metadata envelope. The loader drains **every not-yet-processed file
  oldest→newest** — taking only the newest would drop intermediate deltas.
- **Watermark:** the newest file already applied is recorded in
  `tickets.import_meta.metadata.source_max_key`; runs skip anything at/older than it, so
  reruns are cheap no-ops. `--reseed` ignores it for a one-time bucket cutover.
- **Upsert on `ticket_id`** (PRIMARY KEY) — duplication is impossible; rows are never
  deleted, so closed-ticket history accumulates. On success each file is **moved** to
  `automations/inc/processed/`.
- **Cleaning at ingest:** drop `is_alarm=true` rows + the `EXPORT STOPPED…` sentinel; drop
  `week_start`/`week_end`, `source_s3_*`/`source_snapshot_id`, `department`/`source_type`;
  normalize `region` → lowercase and `raw_status` → UPPERCASE. `service_type` and `bucket`
  (a `closed`/`pending` flag) are kept.
- `tickets.import_meta` captures snapshot freshness (surfaced as `summary.freshness` by
  `fn_tickets_for_map`).
- The curated/geocoded coordinates are written `verified = false` — review
  `tickets.geo_clusters` / `tickets.geo_locations` and flip `verified` once checked.

## Querying

```sql
-- map payload (GeoJSON + summary, incl. summary.freshness) — what dashboard_api serves
SELECT reporting.fn_tickets_for_map();              -- open-only by default
SELECT reporting.fn_tickets_for_map(p_open_only := false);   -- all geocoded tickets

-- open tickets by SLA (derived) + by cluster — via the view
SELECT sla_state, count(*) FROM tickets.inc_open_sla GROUP BY 1;
SELECT cluster, count(*), round(avg(hours_open),1) AS avg_hrs
FROM tickets.inc_open_sla GROUP BY 1 ORDER BY 2 DESC;

-- closures / creations per day (EAT)
SELECT (closed_at AT TIME ZONE 'Africa/Nairobi')::date AS d, count(*)
FROM tickets.inc WHERE closed_at IS NOT NULL GROUP BY 1 ORDER BY 1 DESC;

-- open-backlog-over-time (accrues from first capture; one row per EAT day)
SELECT snapshot_date, open_total, open_breached, closed_today
FROM tickets.inc_daily_snapshot ORDER BY snapshot_date DESC;

-- nearest open tickets to a vehicle (lng, lat) — metres, index-accelerated KNN
SELECT ticket_id, cluster, hours_open,
       round(ST_Distance(geog, ST_SetSRID(ST_MakePoint(:lng,:lat),4326)::geography))::int AS metres
FROM tickets.inc_open_sla
ORDER BY geog <-> ST_SetSRID(ST_MakePoint(:lng,:lat),4326)::geography
LIMIT 10;
```

## Data-quality & SLA notes

Findings to keep in mind (see the PRD for detail):

- **Source `sla_status` is only meaningful for *closed* tickets.** It reads
  `Compliant` for essentially all *open* tickets, so for open work use the **derived**
  state in `tickets.inc_open_sla` (`now() − created_at_service` vs the contract's 48h).
- **`created_at_service` is missing on ~30% of rows** (incl. most open ones); the SLA
  view falls back to `first_seen_at` and flags it via `sla_clock_source`.
- **`mttr` is not wall-clock** `closed_at − created_at_service` and the source's
  `Breached`/`Compliant` does **not** match a plain 48h threshold — pin the contract's
  exact SLA definition before trusting cross-field SLA math.
- **Content lag:** the feed's *file* timestamps are current, but the ticket *content*
  trails ~2 days (the underlying `…wm_task.xlsx` source), so creation/closure dates
  run a couple of days behind wall-clock.
- **History:** `tickets.inc` is current-state (upsert). Closure/creation/MTTR
  *event* series work directly off `closed_at`/`created_at_service`. **Backlog-over-time**
  now accrues via `tickets.inc_daily_snapshot` (one row per EAT day, written by
  `tickets.capture_history()` each ingest); observed closures log to
  `tickets.closure_events`. Past backlog can't be reconstructed — the series builds
  from the first capture onward.

## Status / roadmap

Live: INC ingestion deployed on Coolify (hourly `15 7-19 * * *` EAT), schema +
generated columns + geocoding + the `inc_open_sla` view in `tracksolid_db`.
Next (Phase 2): time-series analytics (closure rate, MTTR/SLA trends), then FleetNow
vehicle **dispatch** off `geog`, and **team closure attribution**. **CRQ** is a
separate future project that will reuse this machinery against `automations/crq/`.