PROXY_URLS_* and other secrets were defined in .env but never loaded,
causing availability to run in slow serial mode (1 req/s) instead of
parallel mode with proxies.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
With parallel workers, threads that fetch a proxy just before escalation
can report failures after the tier has already changed — those failures
were silently counting against the new tier, immediately exhausting it
before it ever got tried (Rayobyte being skipped entirely in favour of
DataImpulse because 10 in-flight Webshare failures hit the threshold).
Fix: build a proxy_url → tier_idx reverse map at construction time and
skip the tier-level circuit breaker when the failing proxy belongs to an
already-escalated tier.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Validates each URL in PROXY_URLS_DATACENTER / PROXY_URLS_RESIDENTIAL:
logs a warning and skips any entry missing an http:// or https:// scheme
instead of passing malformed URLs that cause SSL or connection errors.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add proxy_failure_limit param to make_tiered_cycler (default 3).
Individual proxies hitting the limit are marked dead and permanently
skipped. next_proxy() auto-escalates when all proxies in the active
tier are dead. Both mechanisms coexist: per-proxy dead tracking removes
broken individuals; tier-level threshold catches systemic failure.
- proxy.py: dead_proxies set + proxy_failure_counts dict in state;
next_proxy skips dead proxies with bounded loop; record_failure/
record_success accept optional proxy_url; dead_proxy_count() added
- playtomic_tenants.py: pass proxy_url to record_success/record_failure
- playtomic_availability.py: _worker returns (proxy_url, result);
serial loops in extract + extract_recheck capture proxy_url
- test_supervisor.py: 11 new tests in TestTieredCyclerDeadProxyTracking
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously the tenants extractor flattened all proxy tiers into a single
round-robin list, bypassing the circuit breaker entirely. When the free
Webshare tier runs out of bandwidth (402), all 20 free proxies fail and
the batch crashes — the paid datacenter/residential proxies are never tried.
Changes:
- Replace make_round_robin_cycler with make_tiered_cycler (same as availability)
- Add _fetch_page_via_cycler: retries per page across tiers, records
success/failure in cycler so circuit breaker can escalate
- Fix batch_size to BATCH_SIZE=20 constant (was len(all_proxies) ≈ 22)
- Check cycler.is_exhausted() before each batch; catch RuntimeError mid-batch
and write partial results rather than crashing with nothing
- CIRCUIT_BREAKER_THRESHOLD from env (default 10), matching availability
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously fell back to len(tiers[0]) (e.g. 10 for Webshare) when
PROXY_CONCURRENCY was unset. Default is now MAX_PROXY_CONCURRENCY=200
so single-URL rotating proxies (DC/residential) run at full concurrency
without needing an explicit env var.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- playtomic_tenants: partition by YYYY/MM/DD instead of ISO week;
schedule changed from weekly to daily in workflows.toml
- playtomic_availability: _load_tenant_ids now tries 3-level glob
(*/*/*/tenants.jsonl.gz) first for daily files, falls back to
2-level for old monthly/weekly data
Alphabetical sort would rank old monthly files above daily ones
('t' > '2' in ASCII), so the explicit fallback chain is required.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Tenants extractor now partitions by ISO week (e.g. 2026/W09) instead of
month (2026/02), so each weekly run writes a fresh file rather than
skipping for the rest of the month.
_load_tenant_ids() in playtomic_availability already globs */*/tenants.jsonl.gz
and sorts reverse — 'W09' > '02' alphabetically so weekly files take priority
over old monthly ones automatically.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When DC/residential tiers have a single rotating endpoint, worker_count
defaulted to 1 (one URL = one worker). PROXY_CONCURRENCY lets you set
an explicit thread count (e.g. 100) for providers that handle concurrent
connections on a single URL.
Capped at MAX_PROXY_CONCURRENCY=200 to avoid overloading the endpoint.
Falls back to len(tiers[0]) when unset (existing behaviour).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Phase 2a: NUTS-1 regional income for Germany (16 Bundesländer via admin1→NUTS-1 mapping)
Phase 2b: EU-wide NUTS-2 via GISCO spatial join + US Census ACS state income
- All EU-27+EFTA+UK locations now auto-resolve to NUTS-2 via ST_Contains
- Germany gets sub-Bundesland (38 Regierungsbezirke) differentiation
- US gets state-level income with PPS normalisation
- Income cascade: NUTS-2 → NUTS-1 → US state → country-level
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Tests imported make_sticky_selector but it was never implemented.
Hash-based (MD5) consistent selector — same key always returns the
same proxy, distributes across the pool.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- eurostat.py: add nama_10r_2hhinc dataset config; append filter params to
request URL so server pre-filters the large cube before download
- stg_regional_income.sql: new staging model — reads nama_10r_2hhinc.json.gz,
filters to NUTS-1 codes (3-char), normalises EL→GR / UK→GB
- dim_locations.sql: add admin1_to_nuts1 VALUES CTE (16 German Bundesländer)
+ regional_income CTE; final SELECT uses COALESCE(regional, country) income
- init_landing_seeds.py: add empty seed for nama_10r_2hhinc.json.gz
Munich/Bayern now scores ~29K PPS vs Chemnitz/Sachsen ~19K PPS instead of
both inheriting the same national average (~25.5K PPS).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
878 → 4212 cities. Broadens coverage to match the granularity of
Eurostat and GeoNames data for smaller metro markets.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace single hardcoded Chrome 131 UA with:
- BOT_UA: honest padelnomics-bot UA for Overpass, Eurostat, GeoNames etc.
- _UA_POOL + ua_for_proxy(): deterministic browser UA per proxy URL so each
IP presents a consistent, distinct fingerprint across runs.
Public-API extractors (shared session, no proxy) now send BOT_UA.
Playtomic extractors (proxy-backed) each get a stable pool UA keyed on
their proxy URL hash.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds OVERPASS_MIRRORS list (overpass-api.de, kumi.systems, openstreetmap.ru)
and a post_overpass() helper in _shared.py that tries mirrors in order,
logging a warning on each failure and re-raising the last RequestException
if all mirrors fail. Both overpass.py and overpass_tennis.py now call
post_overpass() instead of hard-coding the primary URL.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Both stg_playtomic_resources and stg_playtomic_opening_hours lacked QUALIFY ROW_NUMBER()
dedup despite declaring a grain. When both tenants.json.gz (old) and tenants.jsonl.gz (new)
exist for the same month, the UNION ALL produced exactly 2× rows.
Fixes:
- stg_playtomic_resources: QUALIFY ROW_NUMBER() OVER (PARTITION BY tenant_id, resource_id)
- stg_playtomic_opening_hours: QUALIFY ROW_NUMBER() OVER (PARTITION BY tenant_id, day_of_week)
- playtomic_tenants.py: skip if old blob OR new JSONL already exists for the month,
preventing same-month dual-format writes that trigger the duplicate
Row counts after fix: ~43.8K resources, ~93.4K opening_hours (was 87.6K, 186.8K).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace single global Overpass query (150K+ elements, times out) with
10 regional bbox queries (~10-40K elements each, 150s server / 180s client).
- REGIONS: 10 bboxes covering all continents
- Crash recovery: working.jsonl accumulates per-region results;
already_seen_ids deduplication skips re-written elements on restart
- Overlapping bbox elements deduped by OSM id across regions
- Retry per region: up to 2 retries with 30s cooldown
- Polite 5s inter-region delay
- Skip if courts.jsonl.gz or courts.json.gz already exists for the month
stg_tennis_courts: UNION ALL transition (jsonl_elements + blob_elements)
- jsonl_elements: JSONL, explicit columns, COALESCE lat/lon with center coords
(supports both node direct lat/lon and way/relation Overpass out center)
- blob_elements: existing UNNEST(elements) pattern, unchanged
- Removed osm_type='node' filter — ways/relations now usable via center coords
- Dedup on (osm_id, extracted_date DESC) unchanged
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- availability_{date}.jsonl.gz replaces .json.gz for morning snapshots
- Each JSONL line = one venue object with date + captured_at_utc injected
- Eliminates in-memory consolidation: working.jsonl IS the final file
(compress_jsonl_atomic at end instead of write_gzip_atomic blob)
- Crash recovery unchanged: working.jsonl accumulates via flush_partial_batch
- _load_morning_availability tries .jsonl.gz first, falls back to .json.gz
- Skip check covers both formats during transition
- Recheck files stay blob format (small, infrequent)
stg_playtomic_availability: UNION ALL transition (morning_jsonl + morning_blob + recheck_blob)
- morning_jsonl: read_json JSONL, tenant_id direct column, no outer UNNEST
- morning_blob / recheck_blob: subquery + LATERAL UNNEST (unchanged semantics)
- All three produce (snapshot_date, captured_at_utc, snapshot_type, recheck_hour, tenant_id, slots_json)
- Downstream raw_resources / raw_slots CTEs unchanged
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- playtomic_tenants.py: write each tenant as a JSONL line after dedup,
compress via compress_jsonl_atomic → tenants.jsonl.gz
- playtomic_availability.py: update _load_tenant_ids() to prefer
tenants.jsonl.gz, fall back to tenants.json.gz (transition)
- stg_playtomic_venues.sql: UNION ALL jsonl+blob CTEs for transition;
JSONL reads top-level columns directly, no UNNEST(tenants) needed
- stg_playtomic_resources.sql: same UNION ALL pattern, single UNNEST
for resources in JSONL path vs double UNNEST in blob path
- stg_playtomic_opening_hours.sql: same UNION ALL pattern, opening_hours
as top-level JSON column in JSONL path
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Streams a JSONL working file to .jsonl.gz in 1MB chunks (constant memory),
atomic rename via .tmp sibling, deletes source on success. Companion to
write_gzip_atomic() for extractors that stream records incrementally.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Each slot is now rechecked once, at most 30 min before it starts.
Worst-case miss: a booking made 29 min before start.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
60-min window + hourly rechecks = each slot caught exactly once, 0-60 min
before it starts. 90-min window causes double-querying (T-90 and T-30).
Slot duration is irrelevant — it doesn't affect when the slot appears in
the window.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Data analysis of 5,115 venues with slots shows 24.8% have a 90-min minimum
slot duration. A 60-min window would miss those venues entirely with hourly
rechecks. 90 min is correct — covers 30/60/90-min minimum venues.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
With hourly rechecks and 60-min minimum slots, a 90-min window causes each
slot to be queried twice. 60-min window = each slot caught exactly once in
the recheck immediately before it starts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- playtomic_tenants.py: batch_size = len(proxy_urls) pages fired in parallel per
batch; each page gets its own session + proxy; sorted(results) ensures
deterministic done-detection; falls back to serial + THROTTLE_SECONDS when no
proxies. Expected speedup: ~2.5 min → ~15 s with 10 proxies.
- .env.dev.sops, .env.prod.sops: remove EXTRACT_WORKERS (now derived from
PROXY_URLS length)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- all.py: replace sequential loop with graphlib.TopologicalSorter + ThreadPoolExecutor
- EXTRACTORS dict declares (func, [deps]) — self-documenting dependency graph
- 8 extractors run in parallel immediately; availability starts as soon as
tenants finishes (not after all others complete)
- max_workers=len(EXTRACTORS) — all I/O-bound, no CPU contention
- playtomic_tenants.py: add proxy rotation via make_round_robin_cycler
- no throttle when PROXY_URLS set (IP rotation removes per-IP rate concern)
- keeps 2s throttle for direct runs
- _shared.py: add optional proxy_url param to run_extractor()
- any extractor can opt in to proxy support via the shared session
- overpass_tennis.py: fix query timeout (out body → out center, timeout 180 → 300)
- out center returns centroids only, not full geometry — fits within server limits
- playtomic_availability.py: fix CIRCUIT_BREAKER_THRESHOLD empty string crash
- int(os.environ.get(..., "10")) → int(os.environ.get(...) or "10")
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- export_serving.py: move `import re` to module level — was imported
inside a loop body on every iteration
- sitemap.py: add comment documenting that the in-memory TTL cache is
process-local (valid for single-worker deployment, Dockerfile --workers 1)
- playtomic_availability.py: use `or "10"` fallback for
CIRCUIT_BREAKER_THRESHOLD env var to handle empty-string case
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Assert landing_dir.is_dir() and year_month format (YYYY/MM) at the
entry point of each extract function — turning silent wrong-path bugs
into immediate AssertionError with a descriptive message.
Files changed:
- playtomic_availability.py: assert in _load_tenant_ids(), extract(),
extract_recheck()
- eurostat.py: assert in extract()
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds a two-tier proxy system for the Playtomic availability extractor:
- Primary tier (PROXY_URLS): datacenter proxies, cheap and fast
- Fallback tier (PROXY_URLS_FALLBACK): residential rotating gateway, reliable
Circuit breaker opens after CIRCUIT_BREAKER_THRESHOLD (default: 10) consecutive
failures, permanently switching to the fallback tier for the rest of the run.
No auto-recovery — avoids flapping. If circuit opens with no fallback configured,
logs an error and writes partial results rather than continuing on a dead proxy pool.
Parallel mode submits futures in PARALLEL_BATCH_SIZE=100 batches so the circuit
breaker can stop new submissions after it opens.
New env vars added to .env.dev.sops (blank defaults):
PROXY_URLS_FALLBACK — residential/rotating gateway URL
CIRCUIT_BREAKER_THRESHOLD — consecutive failures before switching (default 10)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add section 9 to data-sources-inventory.md covering live API quirks:
Eurostat SDMX city labels response shape, ONS CSV download path (observations
API 404s), US Census ACS place endpoint, GeoNames cities15000 bulk format
- Add population coverage summary table and DuckDB glob limitation note
- fix(extract): census_usa + geonames write empty placeholder when credentials
absent so SQLMesh staging models don't fail with "no files found"
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
eurostat_city_labels: API returns compact dimension JSON (category.label dict),
not SDMX 2.1 nested codelists structure. Fixed parser to read from
data["category"]["label"]. 1771 city codes fetched successfully.
ons_uk: observations endpoint (TS007A) is 404. Switched to CSV download via
/datasets/mid-year-pop-est/editions/mid-2022-england-wales — fetches ~68MB CSV,
filters to sex='all' + target year, aggregates population per LAD. 316 LADs ≥50K.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Part A: Data Layer — Sprints 1-5
Sprint 1 — Eurostat SDMX city labels (unblocks EU population):
- New extractor: eurostat_city_labels.py — fetches ESTAT/CITIES codelist
(city_code → city_name mapping) with ETag dedup
- New staging model: stg_city_labels.sql — grain city_code
- Updated dim_cities.sql — joins Eurostat population via city code lookup;
replaces hardcoded 0::BIGINT population
Sprint 2 — Market score formula v2:
- city_market_profile.sql: 30pt population (LN/1M), 25pt income PPS (/200),
30pt demand (occupancy or density), 15pt data confidence
- Moved venue_pricing_benchmarks join into base CTE so median_occupancy_rate
is available to the scoring formula
Sprint 3 — US Census ACS extractor:
- New extractor: census_usa.py — ACS 5-year place population (vintage 2023)
- New staging model: stg_population_usa.sql — grain (place_fips, ref_year)
Sprint 4 — ONS UK extractor:
- New extractor: ons_uk.py — 2021 Census LAD population via ONS beta API
- New staging model: stg_population_uk.sql — grain (lad_code, ref_year)
Sprint 5 — GeoNames global extractor:
- New extractor: geonames.py — cities15000.zip bulk download, filtered to ≥50K pop
- New staging model: stg_population_geonames.sql — grain geoname_id
- dim_cities: 5-source population cascade (Eurostat > Census > ONS > GeoNames > 0)
with case/whitespace-insensitive city name matching
Registered all 4 new CLI entrypoints in pyproject.toml and all.py.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Playtomic API ignores bbox params (min_latitude, etc.) and offset param.
Discovered that `page` param works correctly for global enumeration.
Result: 14,202 venues across 82 countries (was 100 with bbox approach).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
