Three distinct SQL syntax errors caught from worker logs:
1. **upsert_zouit** (bulk_harvest.py:940) — trailing comma after 4326
in ST_Transform(..., 3857), 4326,) → 'syntax error at or near ")"'.
Hit on 8+ quarters in 0108 block (66:41:0108055/61/69/77 etc).
Likely introduced by ruff-format auto-comma on closing paren.
2. **nspd_sync** (nspd_sync.py:120) — `:fetched_at_utc::timestamptz`
psycopg double-colon trap → 'syntax error at or near ":"'.
Crashed legacy on-demand harvest_quarter called from analyze.
3. **job_settings** (job_settings.py:224) — `:extra_config::jsonb`
same double-colon trap in admin settings UPDATE.
All converted to CAST(:x AS type) pattern (psycopg v3 safe).
40 tests still passing.
Co-authored-by: lekss361 <claudestars@proton.me>
cad_parcels.geom is geometry(Polygon, 4326) — strict schema. NSPD
occasionally returns Point geometry for parcels without detailed
boundary, causing INSERT failure:
psycopg.errors.InvalidParameterValue:
Geometry type (Point) does not match column type (Polygon)
This killed pilot v8 at quarter 66:41:0104002 — worker autoretry
exhausted, job hung at 25/50 (heartbeat stale).
Fix: filter geometry.type at Python level in upsert_parcel + upsert_zouit
(same pattern as upsert_quarter_geom_from_feature). Non-Polygon
geometry → geom=NULL, raw_props preserved.
Tables with permissive GEOMETRY schema (cad_buildings, cad_constructions,
cad_enk, cad_oncs) unaffected — they accept any geometry type.
Tests: 3 new (Point parcel → geom=NULL, Polygon happy path, LineString zouit).
Co-authored-by: lekss361 <claudestars@proton.me>
cadastre_jobs.phase_state is shared between ALL parallel Celery workers
of the same job. The early-exit triggered as soon as the FIRST worker
wrote phase=done via progress_cb — all subsequent workers (reading same
shared row) bailed without work.
Effect: pilot v7 made 5 NSPD requests for 50 quarters (=0.1/quarter).
Full ekb_full job #8: 4950 req / 2408 quarters = 2/quarter — 95% workers
early-exited without doing snapshot phase.
Root cause of 1.6% parcels coverage. Fix A/B/C only partially helped
because most workers never reached Phase 1/1.5.
Solution: delete the early-exit. Idempotency is guaranteed by
ON CONFLICT DO UPDATE in every upsert_* — re-enqueued tasks
write same rows without duplicates.
Closes part of #168 (root-cause fix for race condition revealed by #182 metrics).
Co-authored-by: lekss361 <claudestars@proton.me>
scrape_cadastre.py previously incremented requests_count only by
grid_walk_requests, so snapshot-only quarters (and the entire ekb_full
job after PR #179 skips grid-walk on broken geom) reported 0 NSPD
requests — misleading for observability.
Add HarvestResult.snapshot_requests counter (Phase 1 search + Phase 1.5
per-cat probes) and a total_requests property. Update Celery task to
write total_requests into cadastre_jobs.requests_count.
Tests added for the new property + Phase 1 snapshot_requests=1 assertion.
Co-authored-by: lekss361 <claudestars@proton.me>
72% of EKB quarters (1735/2408) have broken micro-precision geom in
cad_quarters_geom. NSPD returns valid quarter polygon in CAT_QUARTER_STATS
(36381) feature but Phase 4 was only consuming stats.
Add upsert_quarter_geom_from_feature: CTE-based UPDATE that transforms
3857 polygon to 4326 and writes only when existing geom is NULL or broken
(bbox width outside 100-10000m), verified via SQL-side sanity check.
Wire into Phase 4 with begin_nested savepoint: malformed NSPD GeoJSON
(self-intersecting ring) rollbacks only geom UPDATE, preserves
upsert_quarter_stats.
Closes part of #168 (follow-up data quality fix for #179).
Co-authored-by: lekss361 <claudestars@proton.me>
Switch base search_by_quarter from thematicSearchId=1 to thematic=1
(includes ZOUIT in features). Add optional category_id param that routes
ZOUIT (36940/469039/469040/469042) -> thematicSearchId=5,
ENK (39663) -> thematicSearchId=15.
Add Phase 1.5 in harvest_quarter: for each PER_CATEGORY_PROBE_CATS
category present in meta_counts (meta_total>0), run dedicated
per-category search call. Skips if category in GRID_WALK_CATS and
bbox is valid (grid-walk gives better coverage).
Curl-probe confirmed NSPD API caps at 20/category with no pagination;
categoryId param is server-ignored; thematic=1 strictly better than
thematicSearchId=1 (includes ZOUIT).
Closes part of #168 (ZOUIT/ENK coverage gap).
Co-authored-by: lekss361 <claudestars@proton.me>
* fix(cadastre): NSPD response parse — properties.category + top-level meta (#168)
* fix(cadastre): test fixtures use real NSPD shape — properties.category + top-level meta (#168)
Bot review of #177 flagged that fixtures still used legacy properties.categoryId
+ data.meta — meaning CI exercised only fallback branches, not primary paths
where the actual bug lived. Pattern of 3 consecutive "live-only" parse fixes
(#175 verify, #176 headers, #177 parse) confirmed need for test coverage.
Changes:
- test_nspd_bulk_client.py: sample_quarter_response → properties.category +
meta moved to top-level (real NSPD shape verified live)
- test_cadastre_bulk.py: 7 fixtures categoryId → category (regex replace)
- test_nspd_bulk_feature_parse_basic: primary path now exercised
Plus schema hardening per bot review:
- NSPDBulkFeature.category_id: `is not None` check (not truthy `or`)
to avoid edge case category=0; int() wrapped in try/except so
non-numeric ID (e.g. "ЗУ") doesn't crash upsert_features loop.
35/35 tests pass locally.
---------
Co-authored-by: lekss361 <claudestars@proton.me>
* fix(parcels): _parse_floors handle int (post-migration #169 schema change)
After PR #169 cad_buildings schema migration, `floors` column is INT
(was TEXT in legacy schema). Existing `_parse_floors(r.get("floors"))`
call in analyze_parcel → _neighbors_summary crashes with:
AttributeError: 'int' object has no attribute 'strip'
Fix: type union str | int | None. If int → return directly (no strip).
Preserve TEXT range parsing ("5-7" → 7) for backwards-compat with
any legacy data still in cad_buildings_old_apr26.
* test(smoke): production smoke tests for post-deploy regressions (#168)
Add tests/smoke/test_prod_smoke.py covering known regression surfaces.
Marks: prod_smoke + slow. Env: PROD_SMOKE_BASE_URL, PROD_SMOKE_ADMIN_TOKEN.
Run manually: cd backend && uv run pytest tests/smoke/ -m prod_smoke -v
* fix(cadastre): exact-match headers to nspd_lite to bypass NSPD WAF (#168)
Pilot v2 (job_id=2) failed 50/50 with HTTP 403 WAF block. Comparing
nspd_bulk_client.DEFAULT_HEADERS vs legacy nspd_lite.HEADERS (which works
on VPS IP since April 2026):
PascalCase keys → lowercase keys
Chrome/148 UA → Chrome/144 UA
No cache-control / pragma → "no-cache" both
accept-language ru first → en first
No origin → "https://nspd.gov.ru"
referer "/map" → "/map?thematic=PKK"
NSPD WAF (BotShield-class) likely fingerprints на header order + values
combined with TLS fingerprint. Matching legacy exactly minimizes deltas.
Test plan: retry pilot job after deploy, expect 0 WAF blocks for first
5 quarters.
* fix(test): exclude prod_smoke tests by default (#168)
CI ran tests/smoke/test_prod_smoke.py and they hit production
https://gendsgn.ru/api/v1/parcels/.../analyze which currently returns 500
(parse_floors regression — exactly what this PR fixes). Catch-22: PR can't
merge because smoke tests fail against pre-merge prod.
Fix: add `addopts = ["-m", "not prod_smoke"]` so default pytest excludes
them. Run manually post-deploy with: pytest -m prod_smoke -v
---------
Co-authored-by: lekss361 <claudestars@proton.me>
Pilot run #1 failed on all 50 quarters with:
NspdBulkError: [SSL: CERTIFICATE_VERIFY_FAILED] self-signed certificate
in certificate chain
NSPD prod chain on Beget VPS contains an internal/self-signed CA →
httpx default verify=True trips. Reuse pattern from legacy
app.services.scrapers.nspd_lite (uses ssl._create_unverified_context()
for same reason since April 2026).
Risk profile: NSPD public gov data, no sensitive payload outbound,
admin token не уходит к NSPD endpoints. MITM risk minimal.
Co-authored-by: lekss361 <claudestars@proton.me>
SQLAlchemy text() parser confuses ':window_interval' (named param) followed
by '::interval' (cast operator) — exactly как было в PR #152 с :weights::jsonb.
Result: psycopg видит literal ':window_interval' в SQL → syntax error →
exception caught by velocity catch → return None → UI shows null.
## Fix
':window_interval::interval' → 'CAST(:window_interval AS interval)' (2 мест:
sales_rows query + _get_ekb_median percentile).
Pre-push code-reviewer должен был catch — добавим в feedback что для каждого
text() SQL grep ':[a-z]*::' before push.
Refs: PR #158 deploy verify, e2e velocity = null root cause
Co-authored-by: lekss361 <claudestars@proton.me>
Production POST /parcels/{cad}/analyze падает с 500:
sqlalchemy.exc.InternalError: (psycopg.errors.InFailedSqlTransaction)
current transaction is aborted, commands ignored until end of transaction block
## Root cause
`compute_velocity()` SQL fails (probably на cad's без conkurrentов / sparse
sale_graph data) → exception caught в try/except → НО db.rollback() отсутствует
→ transaction остаётся в aborted state.
Следующая query (_geotech_risk на line 828) пытается выполниться → крашится
с InFailedSqlTransaction.
## Fix
Wrap velocity в `with db.begin_nested()` — SAVEPOINT pattern (consistent
with PR #124 pzz_loader fix). Failure внутри savepoint:
- Rollbacks ТОЛЬКО savepoint
- Outer transaction остаётся clean
- Subsequent queries (_geotech_risk и пр.) работают
Pattern matches feedback_subagent_delegation audit recommendation для
in-loop / per-section exception handling.
## Impact
POST /parcels/{cad}/analyze больше не 500 при velocity failure. Возвращает
`velocity: null` + остальные fields normal.
Refs: user report 2026-05-15 InFailedSqlTransaction
Co-authored-by: lekss361 <claudestars@proton.me>
Per code review audit (May 14): real SQL injection vector в `enqueue_geo_job`.
Manual `chr(39)*2` quote escape НЕ injection-safe — не защищает от backslash
escape, Unicode quote tricks. Если cad_num когда-либо придёт из user input —
exploitable.
## Change
`backend/app/workers/tasks/nspd_geo.py:298-310` — заменил f-string VALUES
concatenation на SQLAlchemy `text(...)` + list of dicts (psycopg v3 handles
batch executemany под капотом).
Pattern consistent с остальными INSERT'ами в этом же файле (_log,
_start_job, _save_quarter, _save_parcel, _save_building).
## Preserved
- ON CONFLICT (job_id, cad_num, thematic_id) DO NOTHING — verbatim
- Return value (int(job_id))
- Surrounding tx/commit/close logic
- DB schema, function signature, caller interface — unchanged
## Tests
- No test_nspd_geo.py существует — skipped
- ruff check passed
## Vault
`fixes/Bug_Nspd_Geo_Sql_Injection_May14.md` — created (status: resolved).
No backward compat / migration concerns — pure in-place query rewrite.
Co-authored-by: lekss361 <claudestars@proton.me>
Forward-fix после incident May 14 (см. vault events/Incident_Prod_Down_May14_2026).
PR #117 был полностью reverted PR #119 (Caddy depends_on frontend: service_healthy
+ failing wget --spider / → Caddy не стартовал → vsё прод down ~25 мин).
Этот PR делает то же что планировал #117, но корректно:
### Frontend healthcheck — TCP probe через node (НЕ HTTP)
```yaml
test: ["CMD", "node", "-e",
"require('net').createConnection({port:3000,host:'127.0.0.1'})
.once('connect',function(){process.exit(0)})
.once('error',function(){process.exit(1)})"]
start_period: 60s
retries: 6
interval: 15s
```
- НЕ зависит от HTTP status (200/404/308 — все OK если порт слушает)
- `node` гарантированно в alpine image (запускает server.js)
- `127.0.0.1` — никакой DNS ambiguity (busybox alpine localhost IPv6 vs IPv4)
- 60s + 6×15s = 150s window — с запасом на Next.js standalone cold start
### Caddy depends_on — backend healthy, frontend started (НЕ healthy)
```yaml
caddy:
depends_on:
backend: { condition: service_healthy } # backend имеет /health endpoint
frontend: { condition: service_started } # safety net — НЕ блокирует Caddy
```
Lesson from incident: frontend health flaky (Next.js cold start variable).
Strict service_healthy для frontend → одна misconfig = full outage обоих доменов
(gendsgn.ru + obsidian.gendsgn.ru через тот же Caddy на shared network).
### backend/worker/beat redis: service_started → service_healthy
Redis уже имеет `redis-cli ping` healthcheck. Tightening безопасно — был
artifact pre-healthcheck Redis.
### Pin rosreestr2coord>=5.0.0 (lock уже 5.3.3)
v5 убрала delay параметр; код адаптирован (rate limit на worker уровне).
Защита от silent downgrade при `uv lock --upgrade`.
### Что НЕ повторено из #117
- Caddy depends_on frontend: service_healthy — заменено на service_started
- wget --spider / — заменено на node TCP probe
### Vault обновлён
- `events/Incident_Prod_Down_May14_2026.md` — постмортем + timeline + lessons
- Update Changelog в `domains/infra/infra-MOC.md` (если есть)
Closes incident. Re-applies #117 goals safely.
Co-authored-by: lekss361 <claudestars@proton.me>
* feat(site-finder): auto-fetch cadastre geometry on-demand (#93)
Когда пользователь вводит cad-номер которого нет в БД (cad_quarters_geom /
cad_buildings / cad_parcels_geom), вместо 404 «Загрузи через NSPD geo» (dead-
end для non-admin) — теперь backend автоматически инициирует NSPD fetch,
frontend показывает loading state с polling /fetch-status каждые 2с.
Backend:
- Новый модуль app/services/site_finder/cadastre_fetch.py с helpers:
find_or_enqueue_fetch (atomic check + enqueue с дедупликацией по cad),
fetch_status (smart polling endpoint — отличает not_in_nspd от failed),
detect_thematic_id (3-сегм quarter / 4-сегм parcel / 5-сегм building),
validate_cad_format.
- Reuse: enqueue_geo_job + process_nspd_geo_job (workers/tasks/nspd_geo).
source_kind='auto_on_demand' отличает от bulk; rate_ms=200, priority=9.
- POST /parcels/{cad}/analyze graceful fallback: inline await до 25s
(fast path), затем 202 + job_id + eta_seconds для polling, либо
400/404/503 в зависимости от статуса (с Retry-After 60s на 503).
- GET /parcels/{cad}/fetch-status новый endpoint для polling.
Frontend:
- useSiteAnalysis расширен: fetchingState + cancel(). POST analyze + при 202
polling каждые 2с (max 60 итераций = 2 мин cap). status=ready → re-trigger
analyze; not_in_nspd/failed/invalid_format → typed errors.
- apiFetchWithStatus<T> + HTTPError в lib/api.ts — status-aware variant
для 200 vs 202.
- FetchingState.tsx: animated spinner, progress bar (linear до etaSeconds,
asymptote после), elapsed counter, cancel button. Светло-голубая
scheme отличается от обычного pending skeleton.
- site-finder/page.tsx: FetchingState когда fetchingState активен; обычный
pending skeleton — только при первичном analyze без 202.
Edge cases (per #93 acceptance):
- cad валидный в НСПД, fetch <25s → inline 200 OK
- cad валидный, fetch >25s → 202 + frontend polls → ready → analyze
- cad валидный, не в НСПД → 404 с понятным сообщением + формат hint
- cad invalid format → 400 + формат hint
- NSPD rate-limited / failed → 503 + Retry-After 60s
- Параллельные запросы на тот же cad → один job, оба клиента poll'ят (дедуп
через find_active_on_demand_job).
Closes#93.
* fix(site-finder): address PR #95 auto-review minor feedback
Backend (cadastre_fetch.py):
1. (race condition) Advisory lock `pg_try_advisory_xact_lock(hashtext(cad_num))`
обёрнут вокруг шагов "check active job → enqueue" в find_or_enqueue_fetch.
Lock transaction-scoped, released at COMMIT. Параллельные запросы на тот
же cad: первый получает lock и enqueue; второй lock=false → re-check
active job (уже виден после первого COMMIT) → возвращает тот же job_id.
Docstring обновлён, упоминание SELECT FOR UPDATE удалено.
Backend (parcels.py):
3. (threadpool exhaustion) _INLINE_FETCH_WAIT_S снижен 25 → 15s с подробным
комментарием: tradeoff про default Starlette anyio threadpool (40 slots)
и concurrent burst saturation. 15s баланс: НСПД avg 5-15s для quarter,
~70% fast path; остальные 30% получают 202 + polling без блока.
Data (87_on_demand_indexes.sql):
2. (missing index) New migration:
- `nspd_geo_targets_cad_num_idx ON nspd_geo_targets(cad_num)` — для
find_active/recent_completed_job (existing UNIQUE composite не покрывает
WHERE cad_num=:c).
- `nspd_geo_jobs_source_status_idx ON nspd_geo_jobs(source_kind, status)`
composite для filter auto_on_demand + queued/running.
Idempotent (CREATE INDEX IF NOT EXISTS), не блокер при текущем размере,
критично при росте on-demand traffic.
Frontend (useSiteAnalysis.ts):
4. (UI flicker) setFetchingState(null) перемещён ПОСЛЕ `await second
apiFetch`. Иначе между clear и resolve есть момент когда mutation
isPending=true + fetchingState=null → пустой экран ~1 RTT.
NOT addressed (rebuttal):
5. (Tailwind convention) — проверил: в проекте нет ни globals.css, ни
@tailwind directives. ВСЕ существующие site-finder components используют
inline styles (ConfidenceBadge, GeotechRiskBlock, ScoreBreakdownPanel etc).
Tailwind в deps но не wired up. Keep inline styles для consistency.
6. (animate-spin) — требует Tailwind globals (см. #5). `<style jsx>`
keyframes — built-in Next.js, работает.
Per auto-review on 2252236.
---------
Co-authored-by: lekss361 <claudestars@proton.me>
* feat(site-finder): D4 pipeline 24mo — future competition (#36)
Backend (parcels.py):
- Запрос к domrf_kn_objects в радиусе 5км с ready_dt BETWEEN NOW() AND NOW()+24mo.
- _aggregate_pipeline() — сводка: objects_count, flats_total, by_class (эконом/
комфорт/бизнес/...), by_quarter (хронологически, для UI bar), severity
(low <500 / medium <3000 / high) per spec, top_objects (десятка по flat_count desc).
- Поле analyze.pipeline_24mo. Backward-compat — optional.
Frontend:
- Pipeline24moBlock.tsx — severity badge + 3 summary numbers (объектов, квартир,
горизонт/радиус), by-class chips, гистограмма bar по кварталам сдачи
(нормирована на max), разворачиваемый top-N список с классом + датой сдачи.
- Добавлен в MarketTab выше "Market trend".
- TS типы: Pipeline24mo, PipelineObject, PipelineQuarterSlot.
Closes#36. Relates to #19 (Конкурентный 360 — закрыт ранее в #36 scope).
* fix(site-finder): address PR #90 auto-review feedback
Must-fix (3):
1. distance_m falsy guard: `if obj.get("distance_m") is not None` вместо
`if obj.get("distance_m")` — centroid-on-building даёт 0.0 (falsy float),
raw Decimal иначе упал бы в JSON serialization.
2. SQL plan note добавлен про seq scan ~3000 строк OK; при росте — нужен
GIST/index на (latitude, longitude) — отдельный issue для database-expert
(будет создан separately).
3. obj_class NULL bug помечен в docstring _aggregate_pipeline с reference на
fixes/Bug_Kn_API_Obj_Class_Always_Null_OPEN. D6/#38 — fix плановый.
Cleanup (3 из 5):
4. CLASS_LABEL.null dead key убран — JSON null приходит as absent key, не
"null" string.
6. Magic numbers вынесены: PIPELINE_RADIUS_M=5000, PIPELINE_HORIZON_MONTHS=24,
PIPELINE_SEVERITY_MEDIUM_THRESHOLD=500, PIPELINE_SEVERITY_HIGH_THRESHOLD=3000,
PIPELINE_TOP_OBJECTS_LIMIT=10. SQL query теперь через f-string подставляет
их (защищённое от injection — это int литералы).
8. obj.ready_dt formatting через fmtMonth() с new Date + toLocaleDateString —
robust к datetime suffix vs date-only, fallback к substring(0,7) при NaN.
Не сделано (defer):
5. Async 3 HTTP calls (pre-existing pattern, нужен ThreadPoolExecutor refactor
отдельным PR — затрагивает weather/air_quality fetch architecture).
7. ST_GeomFromText дважды — CSE справляется на этом масштабе.
Per auto-review on ade511b.
* fix(site-finder): address PR #90 auto-review minor feedback
1. TS PipelineObject.distance_m — `number | null` для отражения defensive
Python guard (`if obj.get("distance_m") is not None`). Comment объясняет
почему.
2. Pipeline SQL: `text(f"...")` → `text("...")` + parameters. radius_m и
horizon_months через `:param` placeholders + `cast(:horizon_months || ' months'
AS interval)`. Consistency с остальными SQL в файле, plus защита от
accidental injection при будущих изменениях.
3. top_objects: explicit field selection вместо `dict(r) for r in rows`.
Раньше leak'ило все колонки из CTE `SELECT *` (latitude/longitude/
snapshot_date/region_cd/dev_id) в API response. Теперь только nominated
fields: obj_id, comm_name, dev_name, obj_class, flat_count, ready_dt,
distance_m. Schema clean.
Per auto-review on 4e431bf.
---------
Co-authored-by: lekss361 <claudestars@proton.me>
Backend (parcels.py):
- POI scoring loop teper строит score_breakdown_detailed: per-factor list с
verbal explain (через _verbal_for_poi helper) и группировкой
(_POI_GROUP: Социалка / Торговля / Парки / Транспорт / Шум/трамвай / Локация).
- center_bonus добавлен как synthetic factor группы "Локация" с weight=1.0
(decay не применяется — bonus IS the value).
- factor key включает enumerate idx — prevents React key collision когда
два POI одной категории совпадают по округлённому расстоянию.
- Skip факторов с |contribution| < 0.01 (POI > 1км) — UI шуму не нужен.
- abs_total fallback на 1.0 — защита от division-by-zero для empty factors.
- Top-3 positives/negatives: explicit ascending sort для негативов
("most-negative first" очевидно из кода).
- score_by_group: stacked-bar данные с count + contribution_pct.
- group_totals type: dict[str, dict[str, float | int]] (count это int).
Frontend:
- Новый ScoreBreakdownPanel.tsx: stacked-bar по группам с tooltip + legend,
топ-3 плюса (▲ зелёный) / топ-3 минуса (▼ красный) с verbal, отдельная
строка "Снижают балл — Шум/трамвай: ..." для negative groups, разворачиваемая
таблица всех факторов (sticky thead, scrollable).
- Интегрирован в OverviewTab под секцией "Балл".
- TS типы: FactorContribution, ScoreGroupTotal.
Closes#47.
NB: branch создан заново из-за rebase mess (см. PR #87 comments). Логически
эквивалентно но history clean.
Co-authored-by: lekss361 <claudestars@proton.me>
