feat(generator): добавлена стартовая история через backfill

- Зачем:
  - стенду нужна повторяемая история с живым продолжением от модельной границы без дублей и разрыва визитов.
- Что:
  - добавлен backfill-режим с `GEN_MODEL_T_END`, manifest и state на `T_end`.
  - live-запуск восстанавливается из manifest без настенной дельты и проверяет совместимость state.
  - добавлены SQL-проверки формы данных, повторяемости и стыка backfill с live.
- Проверка:
  - make generator-test.
  - два чистых ClickHouse-прогона backfill дали одинаковые manifest checksums и digest.
  - reviewer gate issue 05 пройден после исправлений state/manifest.

generator/src/clickstream_generator/config.py

@@ -78,6 +78,11 @@ class Config:
             os.getenv("GEN_MODEL_T0", "2026-01-01T00:00:00+00:00")
         )
     )
+    model_t_end: datetime | None = field(
+        default_factory=lambda: _parse_optional_model_timestamp(
+            os.getenv("GEN_MODEL_T_END")
+        )
+    )
     model_timezone: str = field(
         default_factory=lambda: os.getenv("GEN_MODEL_TIMEZONE", "UTC")
     )
@@ -122,3 +127,21 @@ class Config:
             raise ValueError("GEN_MODEL_TIME_SPEED must be > 0")
         if self.run_mode not in {"live", "backfill"}:
             raise ValueError("GEN_RUN_MODE must be live or backfill")
+        if self.model_t_end is not None:
+            object.__setattr__(
+                self,
+                "model_t_end",
+                self.model_t_end.astimezone(timezone.utc),
+            )
+        if self.run_mode == "backfill":
+            if self.model_t_end is None:
+                raise ValueError("GEN_MODEL_T_END is required for backfill")
+            if self.model_t_end <= self.model_t0:
+                raise ValueError("GEN_MODEL_T_END must be after GEN_MODEL_T0")
+
+
+def _parse_optional_model_timestamp(value: str | None) -> datetime | None:
+    """Разбирает необязательную ISO-метку модельного времени."""
+    if not value:
+        return None
+    return _parse_model_timestamp(value)

generator/src/clickstream_generator/kafka_io.py

@@ -181,6 +181,84 @@ class KafkaStateManager:
             return None
 
 
+class KafkaStartupHistoryManifest:
+    """Хранение манифеста стартовой истории в Kafka compact topic."""
+
+    MANIFEST_TOPIC = "generator_startup_history_manifest"
+    MANIFEST_KEY = "default"
+
+    def __init__(self, bootstrap_servers: str):
+        self.bootstrap_servers = bootstrap_servers
+        KafkaProducerCls, _ = _kafka_importer()()
+
+        logger.info(
+            f"Connecting to Kafka for startup history manifest at {self.bootstrap_servers}"
+        )
+        self.producer = KafkaProducerCls(
+            bootstrap_servers=self.bootstrap_servers,
+            value_serializer=lambda v: json.dumps(v).encode("utf-8"),
+            key_serializer=lambda k: k.encode("utf-8") if k else None,
+            retries=3,
+            retry_backoff_ms=1000,
+        )
+        logger.info("Connected to Kafka for startup history manifest successfully")
+
+    def save(self, manifest: dict) -> None:
+        """Сохраняет манифест стартовой истории."""
+        def _do_send():
+            self.producer.send(
+                self.MANIFEST_TOPIC,
+                key=self.MANIFEST_KEY,
+                value=manifest,
+            )
+
+        _retry(_do_send, max_retries=3, base_delay=0.5)
+
+    def flush(self) -> None:
+        """Сбрасывает буфер с retry."""
+        def _do_flush():
+            self.producer.flush()
+
+        _retry(_do_flush, max_retries=3, base_delay=0.5)
+
+    def close(self) -> None:
+        """Закрывает соединение."""
+        try:
+            self.producer.close()
+        except Exception as e:
+            logger.debug(f"Error closing manifest producer (ignored): {e}")
+
+    def load(self) -> dict | None:
+        """Загружает последний манифест стартовой истории."""
+        from kafka import KafkaConsumer
+
+        logger.info(f"Loading startup history manifest from topic {self.MANIFEST_TOPIC}")
+
+        def _do_load():
+            consumer = KafkaConsumer(
+                self.MANIFEST_TOPIC,
+                bootstrap_servers=self.bootstrap_servers,
+                auto_offset_reset="earliest",
+                enable_auto_commit=False,
+                consumer_timeout_ms=5000,
+                value_deserializer=lambda v: json.loads(v.decode("utf-8")),
+            )
+
+            last_manifest = None
+            for message in consumer:
+                if message.key and message.key.decode("utf-8") == self.MANIFEST_KEY:
+                    last_manifest = message.value
+
+            consumer.close()
+            return last_manifest
+
+        try:
+            return _retry(_do_load, max_retries=3, base_delay=0.5)
+        except Exception as e:
+            logger.warning(f"Failed to load startup history manifest: {e}")
+            return None
+
+
 def ensure_topics(bootstrap_servers: str) -> None:
     """Создаёт служебные топики, если их ещё нет."""
     from kafka import KafkaAdminClient
@@ -205,8 +283,18 @@ def ensure_topics(bootstrap_servers: str) -> None:
                     "delete.retention.ms": "100",
                 },
             )
+            manifest_topic = NewTopic(
+                name=KafkaStartupHistoryManifest.MANIFEST_TOPIC,
+                num_partitions=1,
+                replication_factor=1,
+                topic_configs={
+                    "cleanup.policy": "compact",
+                    "min.cleanable.dirty.ratio": "0.1",
+                    "delete.retention.ms": "100",
+                },
+            )
 
-            for topic in [history_topic, state_topic]:
+            for topic in [history_topic, state_topic, manifest_topic]:
                 try:
                     admin_client.create_topics([topic])
                     logger.info(f"Created topic: {topic.name}")

generator/src/clickstream_generator/runtime.py

@@ -394,6 +394,24 @@ class TickStreamGenerator:
 
         return tick_batch
 
+    def drain_until(
+        self,
+        cutoff_at: datetime,
+        include_boundary: bool = False,
+    ) -> dict[str, list[dict]]:
+        """Выпускает созревшие события без рождения новых визитов."""
+        cutoff_time = _normalize_tick_time(cutoff_at)
+        tick_batch = _empty_batch()
+
+        self._release_due_events(
+            cutoff_time,
+            tick_batch,
+            include_boundary=include_boundary,
+        )
+        self._drop_finished_visits()
+
+        return tick_batch
+
     def _birth_visits(self, event_budget: int, tick_time: datetime) -> None:
         if len(self.active_visits) >= self.generator.config.max_active_sessions:
             self._pending_visit_births = 0.0
@@ -436,9 +454,17 @@ class TickStreamGenerator:
         self,
         tick_time: datetime,
         tick_batch: dict[str, list[dict]],
+        include_boundary: bool = True,
     ) -> None:
         for visit in self.active_visits:
-            while not visit.is_finished and visit.timestamps[visit.next_index] <= tick_time:
+            while not visit.is_finished:
+                next_timestamp = visit.timestamps[visit.next_index]
+                if include_boundary:
+                    is_due = next_timestamp <= tick_time
+                else:
+                    is_due = next_timestamp < tick_time
+                if not is_due:
+                    break
                 event_index = visit.next_index
                 for topic in TOPICS:
                     tick_batch[topic].append(visit.batch[topic][event_index])

generator/src/clickstream_generator/service.py

@@ -1,5 +1,7 @@
 """Основной сервисный цикл генератора."""
 
+import hashlib
+import json
 import logging
 import sys
 import time
@@ -16,6 +18,7 @@ from clickstream_generator.kafka_io import (
     KafkaBatchHistory,
     KafkaPublisher,
     KafkaStateManager,
+    KafkaStartupHistoryManifest,
     ensure_topics,
 )
 from clickstream_generator.metrics import (
@@ -29,6 +32,105 @@ from clickstream_generator.runtime import TickStreamGenerator
 logger = logging.getLogger("generator")
 
 
+class _TopicManifestStats:
+    """Накопительные счётчики одного Kafka-топика для manifest."""
+
+    def __init__(self):
+        self.rows = 0
+        self.min_event_timestamp: str | None = None
+        self.max_event_timestamp: str | None = None
+        self._checksum = hashlib.sha256()
+
+    @property
+    def checksum(self) -> str:
+        return self._checksum.hexdigest()
+
+    def add(self, event: dict, event_timestamp: str | None = None) -> None:
+        self.rows += 1
+        self._checksum.update(
+            json.dumps(event, sort_keys=True, ensure_ascii=True).encode("utf-8")
+        )
+        timestamp = event_timestamp if event_timestamp is not None else event.get("event_timestamp")
+        self.add_timestamp(timestamp)
+
+    def add_timestamp(self, timestamp: str | None) -> None:
+        if timestamp is None:
+            return
+        if self.min_event_timestamp is None or timestamp < self.min_event_timestamp:
+            self.min_event_timestamp = timestamp
+        if self.max_event_timestamp is None or timestamp > self.max_event_timestamp:
+            self.max_event_timestamp = timestamp
+
+    def to_dict(self) -> dict:
+        return {
+            "rows": self.rows,
+            "min_event_timestamp": self.min_event_timestamp,
+            "max_event_timestamp": self.max_event_timestamp,
+            "checksum_sha256": self.checksum,
+        }
+
+
+class _ManifestCounters:
+    """Счётчики стартовой истории для manifest."""
+
+    def __init__(self):
+        self.topic_stats = {
+            topic: _TopicManifestStats()
+            for topic in ("browser_events", "location_events", "device_events", "geo_events")
+        }
+        self.click_ids: set[str] = set()
+        self.user_ids: set[str] = set()
+
+    def add_batch(self, batch: dict[str, list[dict]]) -> None:
+        browser_events = batch.get("browser_events", [])
+        event_timestamps = {
+            event["event_id"]: event.get("event_timestamp")
+            for event in browser_events
+            if event.get("event_id") and event.get("event_timestamp")
+        }
+        click_timestamps: dict[str, list[str]] = {}
+        for event in browser_events:
+            click_id = event.get("click_id")
+            timestamp = event.get("event_timestamp")
+            if click_id and timestamp:
+                click_timestamps.setdefault(click_id, []).append(timestamp)
+
+        for topic, events in batch.items():
+            stats = self.topic_stats[topic]
+            for event in events:
+                event_timestamp = event.get("event_timestamp")
+                if event_timestamp is None and topic == "location_events":
+                    event_timestamp = event_timestamps.get(event.get("event_id"))
+                if event_timestamp is None and topic in ("device_events", "geo_events"):
+                    timestamps = click_timestamps.get(event.get("click_id"))
+                    if timestamps:
+                        event_timestamp = min(timestamps)
+                        stats.add_timestamp(max(timestamps))
+                stats.add(event, event_timestamp=event_timestamp)
+                click_id = event.get("click_id")
+                if topic == "browser_events" and click_id:
+                    self.click_ids.add(click_id)
+                user_id = event.get("user_domain_id")
+                if topic == "device_events" and user_id:
+                    self.user_ids.add(user_id)
+
+    def to_manifest_topics(self) -> dict:
+        return {
+            topic: stats.to_dict()
+            for topic, stats in self.topic_stats.items()
+        }
+
+    def to_manifest_totals(self) -> dict:
+        browser_stats = self.topic_stats["browser_events"]
+        return {
+            "events": browser_stats.rows,
+            "visits": len(self.click_ids),
+            "users": len(self.user_ids),
+            "min_event_timestamp": browser_stats.min_event_timestamp,
+            "max_event_timestamp": browser_stats.max_event_timestamp,
+        }
+
+
 class GeneratorService:
     """Основной сервис генератора."""
 
@@ -40,6 +142,7 @@ class GeneratorService:
         self.publisher: KafkaPublisher | None = None
         self.history: KafkaBatchHistory | None = None
         self.state_manager: KafkaStateManager | None = None
+        self.manifest_manager: KafkaStartupHistoryManifest | None = None
         self._running = False
         self._tick = 0
         self._model_time = config.model_t0
@@ -66,17 +169,46 @@ class GeneratorService:
         self.publisher = KafkaPublisher(self.config.kafka_bootstrap_servers)
         self.history = KafkaBatchHistory(self.config.kafka_bootstrap_servers)
 
+        if self.config.run_mode == "backfill" and not self.config.state_enabled:
+            raise ValueError("GEN_STATE_ENABLED must be true for backfill")
+
         if self.config.state_enabled:
             self.state_manager = KafkaStateManager(self.config.kafka_bootstrap_servers)
 
+            if self.config.run_mode == "backfill":
+                self.manifest_manager = KafkaStartupHistoryManifest(
+                    self.config.kafka_bootstrap_servers
+                )
+                logger.info("Backfill mode starts from a fresh generator state")
+                self._run_backfill()
+                self.stop()
+                return
+
             if not self.config.state_reset:
                 restored_state = self.state_manager.load()
                 if restored_state:
                     try:
-                        self._restore_live_state(
-                            restored_state,
-                            wall_now_utc=datetime.now(timezone.utc),
+                        self.manifest_manager = KafkaStartupHistoryManifest(
+                            self.config.kafka_bootstrap_servers
                         )
+                        manifest = self.manifest_manager.load()
+                        if self._is_startup_history_state(restored_state, manifest):
+                            model_t_end = self._as_aware_utc(
+                                datetime.fromisoformat(manifest["model_t_end"])
+                            )
+                            self.restore_from_startup_history(
+                                restored_state,
+                                model_t_end=model_t_end,
+                            )
+                        elif self._is_startup_history_marker(restored_state):
+                            raise ValueError(
+                                "startup-history state without matching manifest"
+                            )
+                        else:
+                            self._restore_live_state(
+                                restored_state,
+                                wall_now_utc=datetime.now(timezone.utc),
+                            )
                         logger.info(
                             f"Restored state: continuing from tick {self._tick}, "
                             f"model_time={self._model_time.isoformat()}, "
@@ -113,6 +245,8 @@ class GeneratorService:
             self.history.close()
         if self.state_manager:
             self.state_manager.close()
+        if self.manifest_manager:
+            self.manifest_manager.close()
 
     def restore_from_startup_history(
         self,
@@ -169,6 +303,54 @@ class GeneratorService:
                 "state config mismatch: " + ", ".join(mismatches)
             )
 
+    def _is_startup_history_state(self, state, manifest: dict | None) -> bool:
+        """Проверяет, что state совпадает со слепком стартовой истории."""
+        if not manifest or manifest.get("run_mode") != "backfill":
+            return False
+
+        expected_state = manifest.get("state") or {}
+        if expected_state.get("last_batch_id") != state.last_batch_id:
+            return False
+
+        try:
+            model_t_end = self._as_aware_utc(
+                datetime.fromisoformat(manifest["model_t_end"])
+            )
+            model_t0 = self._as_aware_utc(
+                datetime.fromisoformat(manifest["model_t0"])
+            )
+        except (KeyError, TypeError, ValueError):
+            return False
+
+        if self._as_aware_utc(state.model_timestamp) != model_t_end:
+            return False
+        if self._as_aware_utc(state.model_t0) != model_t0:
+            return False
+        if state.gen_seed != manifest.get("gen_seed"):
+            return False
+        if state.model_timezone != manifest.get("model_timezone"):
+            return False
+        if (
+            self.config.model_t_end is not None
+            and self.config.model_t_end != model_t_end
+        ):
+            return False
+        if model_t0 != self.config.model_t0:
+            return False
+        if manifest.get("gen_seed") != self.config.seed:
+            return False
+        if manifest.get("model_timezone") != self.config.model_timezone:
+            return False
+
+        settings = manifest.get("generation_settings") or {}
+        if abs(state.model_time_speed - settings.get("model_time_speed", -1)) > 1e-9:
+            return False
+        return self._generation_settings() == settings
+
+    def _is_startup_history_marker(self, state) -> bool:
+        """Отличает state стартовой истории от обычного live-state."""
+        return str(state.last_batch_id).startswith("startup-history-")
+
     @staticmethod
     def _as_aware_utc(value: datetime) -> datetime:
         if value.tzinfo is None:
@@ -204,6 +386,220 @@ class GeneratorService:
             logger.warning(f"Failed to save state: {e}")
             METRICS_ERRORS_TOTAL.labels(topic="state").inc()
 
+    def _run_backfill(self) -> None:
+        """Проматывает стартовую историю без сна до GEN_MODEL_T_END."""
+        if self.config.model_t_end is None:
+            raise ValueError("GEN_MODEL_T_END is required for backfill")
+        if not self.publisher:
+            raise RuntimeError("publisher is not initialized")
+        if not self.history:
+            raise RuntimeError("history is not initialized")
+
+        logger.info(
+            "Running backfill from %s to %s",
+            self.config.model_t0.isoformat(),
+            self.config.model_t_end.isoformat(),
+        )
+        counters = _ManifestCounters()
+
+        while self._model_time < self.config.model_t_end:
+            self._tick += 1
+            batch_id = f"backfill-{self._tick:08d}"
+            model_time = self._model_time
+            started_at = datetime.now(timezone.utc)
+
+            events_count = self.generator._calculate_events_count(now=model_time)
+            batch = self.stream.generate_tick(
+                events_count,
+                tick_started_at=model_time,
+            )
+            total_sent, sent_counts, status = self._publish_batch(batch)
+            self._raise_on_backfill_publish_error(batch_id, status, sent_counts)
+            counters.add_batch(batch)
+            self._write_batch_history(
+                batch_id=batch_id,
+                started_at=started_at,
+                sent_counts=sent_counts,
+                total_sent=total_sent,
+                status=status,
+                error_message=None if status == "success" else "Backfill publish error",
+            )
+            self._advance_model_time()
+
+        final_batch = self.stream.drain_until(
+            self.config.model_t_end,
+            include_boundary=False,
+        )
+        final_sent = 0
+        final_status = "success"
+        if any(final_batch.values()):
+            batch_id = f"backfill-{self._tick + 1:08d}-final"
+            started_at = datetime.now(timezone.utc)
+            final_sent, sent_counts, final_status = self._publish_batch(final_batch)
+            self._raise_on_backfill_publish_error(
+                batch_id,
+                final_status,
+                sent_counts,
+            )
+            counters.add_batch(final_batch)
+            self._write_batch_history(
+                batch_id=batch_id,
+                started_at=started_at,
+                sent_counts=sent_counts,
+                total_sent=final_sent,
+                status=final_status,
+                error_message=None if final_status == "success" else "Backfill publish error",
+            )
+
+        if self.publisher:
+            self.publisher.flush()
+
+        self._model_time = self.config.model_t_end
+        state_batch_id = self._startup_state_batch_id(counters)
+        state = self.stream.to_state(
+            tick=self._tick,
+            rng_state=self.generator.rng.getstate(),
+            last_batch_id=state_batch_id,
+            last_timestamp=self.config.model_t_end,
+            model_timestamp=self.config.model_t_end,
+            wall_timestamp=datetime.now(timezone.utc),
+            model_time_speed=self.config.model_time_speed,
+            model_timezone=self.config.model_timezone,
+            model_t0=self.config.model_t0,
+            gen_seed=self.config.seed,
+        )
+
+        manifest = self._build_startup_history_manifest(
+            counters=counters,
+            state=state,
+        )
+        if self.manifest_manager:
+            self.manifest_manager.save(manifest)
+            self.manifest_manager.flush()
+
+        if self.state_manager and self.config.state_enabled:
+            self.state_manager.save(state)
+            self.state_manager.flush()
+
+        logger.info(
+            "Backfill completed: events=%s, visits=%s, users=%s, final_sent=%s",
+            manifest["totals"]["events"],
+            manifest["totals"]["visits"],
+            manifest["totals"]["users"],
+            final_sent,
+        )
+
+    def _raise_on_backfill_publish_error(
+        self,
+        batch_id: str,
+        status: str,
+        sent_counts: dict[str, dict[str, int]],
+    ) -> None:
+        """Останавливает backfill до записи state/manifest при ошибке Kafka."""
+        if status == "success":
+            return
+
+        raise RuntimeError(
+            f"Backfill publish failed for batch {batch_id}: "
+            f"status={status}, sent_counts={sent_counts}"
+        )
+
+    def _publish_batch(
+        self,
+        batch: dict[str, list[dict]],
+    ) -> tuple[int, dict[str, dict[str, int]], str]:
+        """Публикует batch и возвращает счётчики отправки."""
+        total_sent = 0
+        total_errors = 0
+        sent_counts = {}
+
+        for topic, events in batch.items():
+            if events:
+                sent, errors = self.publisher.publish(topic, events)
+                sent_counts[topic] = {"sent": sent, "errors": errors}
+                total_sent += sent
+                total_errors += errors
+
+        if total_errors == 0:
+            status = "success"
+        elif total_sent > 0:
+            status = "partial"
+        else:
+            status = "error"
+
+        return total_sent, sent_counts, status
+
+    def _write_batch_history(
+        self,
+        batch_id: str,
+        started_at: datetime,
+        sent_counts: dict[str, dict[str, int]],
+        total_sent: int,
+        status: str,
+        error_message: str | None,
+    ) -> None:
+        """Пишет служебную историю batch."""
+        if not self.history:
+            return
+        try:
+            record = BatchRecord(
+                batch_id=batch_id,
+                started_at=started_at,
+                finished_at=datetime.now(timezone.utc),
+                sent_total=total_sent,
+                sent_browser=sent_counts.get("browser_events", {}).get("sent", 0),
+                sent_location=sent_counts.get("location_events", {}).get("sent", 0),
+                sent_device=sent_counts.get("device_events", {}).get("sent", 0),
+                sent_geo=sent_counts.get("geo_events", {}).get("sent", 0),
+                status=status,
+                error_message=error_message,
+            )
+            self.history.add(record)
+            self.history.flush()
+        except Exception as hist_err:
+            logger.warning(f"Failed to write batch history: {hist_err}")
+            METRICS_ERRORS_TOTAL.labels(topic="history").inc()
+
+    def _startup_state_batch_id(self, counters) -> str:
+        digest = counters.topic_stats["browser_events"].checksum[:12]
+        return f"startup-history-{digest}"
+
+    def _build_startup_history_manifest(self, counters, state) -> dict:
+        return {
+            "manifest_version": "1.0",
+            "generated_at": datetime.now(timezone.utc).isoformat(),
+            "gen_seed": self.config.seed,
+            "model_t0": self.config.model_t0.isoformat(),
+            "model_t_end": self.config.model_t_end.isoformat(),
+            "model_timezone": self.config.model_timezone,
+            "run_mode": "backfill",
+            "generation_settings": self._generation_settings(),
+            "state_version": state.version,
+            "state": {
+                "topic": KafkaStateManager.STATE_TOPIC,
+                "key": KafkaStateManager.STATE_KEY,
+                "last_batch_id": state.last_batch_id,
+                "model_timestamp": state.model_timestamp.isoformat(),
+            },
+            "topics": counters.to_manifest_topics(),
+            "totals": counters.to_manifest_totals(),
+        }
+
+    def _generation_settings(self) -> dict:
+        return {
+            "tick_seconds": self.config.tick_seconds,
+            "lambda_base_per_min": self.config.lambda_base_per_min,
+            "jitter_pct": self.config.jitter_pct,
+            "min_events_per_tick": self.config.min_events_per_tick,
+            "max_events_per_tick": self.config.max_events_per_tick,
+            "max_session_events": self.config.max_session_events,
+            "max_active_sessions": self.config.max_active_sessions,
+            "population_max": self.config.population_max,
+            "p_new_user": self.config.p_new_user,
+            "min_return_minutes": self.config.min_return_minutes,
+            "model_time_speed": self.config.model_time_speed,
+        }
+
     def _main_loop(self):
         """Основной цикл тиков."""
         while self._running: