perf(extract): auto-detect workers from proxies, skip throttle on success, crash-safe partial JSONL

- proxy.py: delete unused make_sticky_selector()
- utils.py: add load_partial_results() + flush_partial_batch() for crash-resumable extraction
- playtomic_availability.py:
  - drop MAX_WORKERS / EXTRACT_WORKERS — worker_count = len(proxy_urls) or 1
  - skip time.sleep(THROTTLE_SECONDS) on success when proxy_url is set; keep sleeps for 429/5xx
  - replace cursor-based resumption with .partial.jsonl sidecar (flush every 50 records)
  - _fetch_venues_parallel accepts on_result callback for incremental partial-file flushing
  - mirror auto-detect worker count in extract_recheck()

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

extract/padelnomics_extract/src/padelnomics_extract/playtomic_availability.py

@@ -5,8 +5,13 @@ unauthenticated /v1/availability endpoint for each venue's next-day slots.
 This is the highest-value source: daily snapshots enable occupancy rate
 estimation, pricing benchmarking, and demand signal detection.
 
-Parallel mode: set EXTRACT_WORKERS=N and PROXY_URLS=... to fetch N venues
+Parallel mode: worker count is derived from PROXY_URLS (one worker per proxy).
-concurrently (one proxy per worker). Without proxies, runs single-threaded.
+Without proxies, runs single-threaded with per-request throttling.
+
+Crash resumption: progress is flushed to a .partial.jsonl sidecar file every
+PARTIAL_FLUSH_SIZE records. On restart the already-fetched venues are skipped
+and prior results are merged into the final file. At most PARTIAL_FLUSH_SIZE
+records (a few seconds of work with 10 workers) are lost on crash.
 
 Recheck mode: re-queries venues with slots starting within the next 90 minutes.
 Writes a separate recheck file for more accurate occupancy measurement.
@@ -29,7 +34,7 @@ import niquests
 
 from ._shared import HTTP_TIMEOUT_SECONDS, USER_AGENT, run_extractor, setup_logging
 from .proxy import load_fallback_proxy_urls, load_proxy_urls, make_tiered_cycler
-from .utils import get_last_cursor, landing_path, write_gzip_atomic
+from .utils import flush_partial_batch, landing_path, load_partial_results, write_gzip_atomic
 
 logger = setup_logging("padelnomics.extract.playtomic_availability")
 
@@ -40,7 +45,6 @@ AVAILABILITY_URL = "https://api.playtomic.io/v1/availability"
 THROTTLE_SECONDS = 1
 MAX_VENUES_PER_RUN = 20_000
 MAX_RETRIES_PER_VENUE = 2
-MAX_WORKERS = int(os.environ.get("EXTRACT_WORKERS", "1"))
 RECHECK_WINDOW_MINUTES = int(os.environ.get("RECHECK_WINDOW_MINUTES", "90"))
 CIRCUIT_BREAKER_THRESHOLD = int(os.environ.get("CIRCUIT_BREAKER_THRESHOLD", "10"))
 
@@ -49,6 +53,9 @@ CIRCUIT_BREAKER_THRESHOLD = int(os.environ.get("CIRCUIT_BREAKER_THRESHOLD", "10"
 # batch still complete.
 PARALLEL_BATCH_SIZE = 100
 
+# Flush partial results to disk every N records — lose at most this many on crash.
+PARTIAL_FLUSH_SIZE = 50
+
 # Thread-local storage for per-worker sessions
 _thread_local = threading.local()
 
@@ -84,22 +91,6 @@ def _load_tenant_ids(landing_dir: Path) -> list[str]:
     return ids
 
 
-def _parse_resume_cursor(cursor: str | None, target_date: str) -> int:
-    """Parse cursor_value to find resume index. Returns 0 if no valid cursor."""
-    if not cursor:
-        return 0
-    parts = cursor.split(":", 1)
-    if len(parts) != 2:
-        return 0
-    cursor_date, cursor_index = parts
-    if cursor_date != target_date:
-        return 0
-    try:
-        return int(cursor_index)
-    except ValueError:
-        return 0
-
-
 # ---------------------------------------------------------------------------
 # Per-venue fetch (used by both serial and parallel modes)
 # ---------------------------------------------------------------------------
@@ -149,7 +140,8 @@ def _fetch_venue_availability(
                 continue
 
             resp.raise_for_status()
-            time.sleep(THROTTLE_SECONDS)
+            if not proxy_url:
+                time.sleep(THROTTLE_SECONDS)
             return {"tenant_id": tenant_id, "slots": resp.json()}
 
         except niquests.exceptions.RequestException as e:
@@ -177,6 +169,7 @@ def _fetch_venues_parallel(
     worker_count: int,
     cycler: dict,
     fallback_urls: list[str],
+    on_result=None,
 ) -> tuple[list[dict], int]:
     """Fetch availability for multiple venues in parallel.
 
@@ -184,6 +177,9 @@ def _fetch_venues_parallel(
     completes, checks the circuit breaker: if it opened and there is no
     fallback configured, stops submitting further batches.
 
+    on_result: optional callable(result: dict) invoked inside the lock for
+    each successful result — used for incremental partial-file flushing.
+
     Returns (venues_data, venues_errored).
     """
     venues_data: list[dict] = []
@@ -215,6 +211,8 @@ def _fetch_venues_parallel(
                     if result is not None:
                         venues_data.append(result)
                         cycler["record_success"]()
+                        if on_result is not None:
+                            on_result(result)
                     else:
                         venues_errored += 1
                         cycler["record_failure"]()
@@ -262,41 +260,56 @@ def extract(
         logger.info("Already have %s — skipping", dest)
         return {"files_written": 0, "files_skipped": 1, "bytes_written": 0}
 
-    # Resume from cursor if crashed mid-run
+    # Crash resumption: load already-fetched venues from partial file
-    last_cursor = get_last_cursor(conn, EXTRACTOR_NAME)
+    partial_path = dest.with_suffix(".partial.jsonl")
-    resume_index = _parse_resume_cursor(last_cursor, target_date)
+    prior_results, already_done = load_partial_results(partial_path, id_key="tenant_id")
-    if resume_index > 0:
+    if already_done:
-        logger.info("Resuming from index %d (cursor: %s)", resume_index, last_cursor)
+        logger.info("Resuming: %d venues already fetched from partial file", len(already_done))
 
-    venues_to_process = tenant_ids[:MAX_VENUES_PER_RUN]
+    all_venues_to_process = tenant_ids[:MAX_VENUES_PER_RUN]
-    if resume_index > 0:
+    venues_to_process = [tid for tid in all_venues_to_process if tid not in already_done]
-        venues_to_process = venues_to_process[resume_index:]
 
     # Set up tiered proxy cycler with circuit breaker
     proxy_urls = load_proxy_urls()
     fallback_urls = load_fallback_proxy_urls()
-    worker_count = min(MAX_WORKERS, len(proxy_urls)) if proxy_urls else 1
+    worker_count = len(proxy_urls) if proxy_urls else 1
     cycler = make_tiered_cycler(proxy_urls, fallback_urls, CIRCUIT_BREAKER_THRESHOLD)
 
     start_min_str = start_min.strftime("%Y-%m-%dT%H:%M:%S")
     start_max_str = start_max.strftime("%Y-%m-%dT%H:%M:%S")
 
+    # Partial file for incremental crash-safe progress
+    partial_file = open(partial_path, "a")  # noqa: SIM115
+    partial_lock = threading.Lock()
+    pending_batch: list[dict] = []
+
+    def _on_result(result: dict) -> None:
+        # Called inside _fetch_venues_parallel's lock — no additional locking needed.
+        # In serial mode, called single-threaded — also safe without extra locking.
+        pending_batch.append(result)
+        if len(pending_batch) >= PARTIAL_FLUSH_SIZE:
+            flush_partial_batch(partial_file, partial_lock, pending_batch)
+            pending_batch.clear()
+
+    new_venues_data: list[dict] = []
+    venues_errored = 0
+
     if worker_count > 1:
         logger.info("Parallel mode: %d workers, %d proxies", worker_count, len(proxy_urls))
-        venues_data, venues_errored = _fetch_venues_parallel(
+        new_venues_data, venues_errored = _fetch_venues_parallel(
             venues_to_process, start_min_str, start_max_str, worker_count, cycler, fallback_urls,
+            on_result=_on_result,
         )
     else:
         logger.info("Serial mode: 1 worker, %d venues", len(venues_to_process))
-        venues_data = []
-        venues_errored = 0
         for i, tenant_id in enumerate(venues_to_process):
             result = _fetch_venue_availability(
                 tenant_id, start_min_str, start_max_str, cycler["next_proxy"](),
             )
             if result is not None:
-                venues_data.append(result)
+                new_venues_data.append(result)
                 cycler["record_success"]()
+                _on_result(result)
             else:
                 venues_errored += 1
                 circuit_opened = cycler["record_failure"]()
@@ -310,7 +323,14 @@ def extract(
                     i + 1, len(venues_to_process), venues_errored,
                 )
 
-    # Write consolidated file
+    # Final flush of any remaining partial batch
+    if pending_batch:
+        flush_partial_batch(partial_file, partial_lock, pending_batch)
+        pending_batch.clear()
+    partial_file.close()
+
+    # Consolidate prior (resumed) + new results into final file
+    venues_data = prior_results + new_venues_data
     captured_at = datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ")
     payload = json.dumps({
         "date": target_date,
@@ -321,6 +341,9 @@ def extract(
     }).encode()
 
     bytes_written = write_gzip_atomic(dest, payload)
+    if partial_path.exists():
+        partial_path.unlink()
+
     logger.info(
         "%d venues scraped (%d errors) -> %s (%s bytes)",
         len(venues_data), venues_errored, dest, f"{bytes_written:,}",
@@ -330,7 +353,7 @@ def extract(
         "files_written": 1,
         "files_skipped": 0,
         "bytes_written": bytes_written,
-        "cursor_value": f"{target_date}:{len(tenant_ids[:MAX_VENUES_PER_RUN])}",
+        "cursor_value": f"{target_date}:{len(all_venues_to_process)}",
     }
 
 
@@ -421,7 +444,7 @@ def extract_recheck(
     # Set up tiered proxy cycler with circuit breaker
     proxy_urls = load_proxy_urls()
     fallback_urls = load_fallback_proxy_urls()
-    worker_count = min(MAX_WORKERS, len(proxy_urls)) if proxy_urls else 1
+    worker_count = len(proxy_urls) if proxy_urls else 1
     cycler = make_tiered_cycler(proxy_urls, fallback_urls, CIRCUIT_BREAKER_THRESHOLD)
 
     if worker_count > 1 and len(venues_to_recheck) > 10:

extract/padelnomics_extract/src/padelnomics_extract/proxy.py

@@ -3,10 +3,6 @@
 Proxies are configured via the PROXY_URLS environment variable (comma-separated).
 When unset, all functions return None/no-op — extractors fall back to direct requests.
 
-Two routing modes:
-  round-robin — distribute requests evenly across proxies (default)
-  sticky      — same key always maps to same proxy (for session-tracked sites)
-
 Tiered proxy with circuit breaker:
   Primary tier (PROXY_URLS) is used by default — typically cheap datacenter proxies.
   Fallback tier (PROXY_URLS_FALLBACK) activates once consecutive failures >= threshold.
@@ -141,17 +137,3 @@ def make_tiered_cycler(
         "is_fallback_active": is_fallback_active,
     }
 
-
-def make_sticky_selector(proxy_urls: list[str]):
-    """Consistent-hash proxy selector — same key always maps to same proxy.
-
-    Use when the target site tracks sessions by IP (e.g. Cloudflare).
-    Returns a callable: select_proxy(key: str) -> str | None
-    """
-    if not proxy_urls:
-        return lambda key: None
-
-    def select_proxy(key: str) -> str:
-        return proxy_urls[hash(key) % len(proxy_urls)]
-
-    return select_proxy

extract/padelnomics_extract/src/padelnomics_extract/utils.py

@@ -7,7 +7,9 @@ if you add multiple data sources, extract them to a shared workspace package.
 
 import gzip
 import hashlib
+import json
 import sqlite3
+import threading
 from pathlib import Path
 
 # ---------------------------------------------------------------------------
@@ -117,6 +119,50 @@ def content_hash(data: bytes, prefix_bytes: int = 8) -> str:
     return hashlib.sha256(data).hexdigest()[:prefix_bytes]
 
 
+def load_partial_results(partial_path: Path, id_key: str) -> tuple[list[dict], set[str]]:
+    """Load already-completed records from a partial JSONL file (crash recovery).
+
+    Returns (records, seen_ids). If the file doesn't exist, returns ([], set()).
+    Gracefully handles a truncated last line from a mid-write crash.
+    """
+    records: list[dict] = []
+    seen_ids: set[str] = set()
+    if not partial_path.exists():
+        return records, seen_ids
+    with open(partial_path) as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                record = json.loads(line)
+                records.append(record)
+                rid = record.get(id_key)
+                if rid:
+                    seen_ids.add(rid)
+            except json.JSONDecodeError:
+                break  # truncated last line from crash — skip it
+    return records, seen_ids
+
+
+def flush_partial_batch(
+    partial_file,
+    lock: threading.Lock,
+    batch: list[dict],
+) -> None:
+    """Thread-safe batch write of JSON records to the partial JSONL file.
+
+    Writes all records in one lock acquisition with a single flush.
+    Call with batches of ~50 records for good I/O throughput vs crash safety tradeoff.
+    On crash, at most one batch worth of records is lost.
+    """
+    assert batch, "batch must not be empty"
+    with lock:
+        for record in batch:
+            partial_file.write(json.dumps(record, separators=(",", ":")) + "\n")
+        partial_file.flush()
+
+
 def write_gzip_atomic(path: Path, data: bytes) -> int:
     """Gzip compress data and write to path atomically via .tmp sibling.