NewsSite/ai-worker/worker.py

import base64
import logging
import os
import socket
import threading
import time
from io import BytesIO
from pathlib import Path
from typing import Optional

import psycopg2
import torch
from diffusers import StableDiffusionPipeline

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")

DB_URL = os.environ.get("DATABASE_URL", "postgres://app:appsecret@database:5432/appdb?sslmode=disable")
DEFAULT_MODEL_PATH = Path(__file__).resolve().parent / "models" / "sd15"
MODEL_PATH = Path(DEFAULT_MODEL_PATH)
IMAGE_OUTPUT_DIR = Path(os.environ.get("IMAGE_OUTPUT_DIR") or "/app/images")
IMAGE_OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
MODEL_ID = "stabilityai/sd-turbo"
WORKER_ID = os.environ.get("WORKER_ID") or f"{socket.gethostname()}:{os.getpid()}"
BATCH_SIZE = int(os.environ.get("IMAGE_BATCH_SIZE") or "3")
CLAIM_TTL_MINUTES = int(os.environ.get("IMAGE_CLAIM_TTL_MINUTES") or "20")
pipe = None
pipe_lock = threading.Lock()


def connect():
    return psycopg2.connect(DB_URL)


def load_pipeline() -> Optional[StableDiffusionPipeline]:
    global pipe

    if pipe is not None:
        return pipe

    with pipe_lock:
        if pipe is not None:
            return pipe

        try:
            # Prüfen, ob lokal schon etwas im Modellordner liegt
            has_local_model = MODEL_PATH.is_dir() and any(MODEL_PATH.iterdir())

            if has_local_model:
                logging.info("Lade Stable Diffusion Modell lokal aus %s", MODEL_PATH)
                pipe = StableDiffusionPipeline.from_pretrained(
                    str(MODEL_PATH),
                    torch_dtype=torch.float32,
                    local_files_only=True,
                )
            else:
                logging.info("Kein lokales Modell gefunden – lade von Hugging Face (%s)", MODEL_ID)
                pipe = StableDiffusionPipeline.from_pretrained(
                    MODEL_ID,
                    torch_dtype=torch.float32
                )
                pipe.save_pretrained(MODEL_PATH)
                logging.info("Modell erfolgreich nach %s gespeichert", MODEL_PATH)

            pipe = pipe.to("cpu")
            pipe.enable_attention_slicing()
            return pipe

        except Exception as e:
            logging.error("Konnte Pipeline nicht laden: %s", e)
            pipe = None
            return None


def fetch_articles_without_image(cur, limit=BATCH_SIZE):
    cur.execute(
        """
        WITH claim AS (
            SELECT id
            FROM articles
            WHERE image IS NULL
              AND (
                image_claimed_at IS NULL
                OR image_claimed_at < now() - (%s * INTERVAL '1 minute')
              )
            ORDER BY created_at DESC
            LIMIT %s
            FOR UPDATE SKIP LOCKED
        )
        UPDATE articles AS a
        SET image_claimed_at = now(),
            image_claimed_by = %s
        FROM claim
        WHERE a.id = claim.id
        RETURNING a.id, a.title, a.article_id
        """,
        (CLAIM_TTL_MINUTES, limit, WORKER_ID),
    )
    return cur.fetchall()


def update_image(cur, article_id: int, data_uri: str):
    cur.execute(
        """
        UPDATE articles
        SET image = %s,
            image_claimed_at = NULL,
            image_claimed_by = NULL
        WHERE id = %s
        """,
        (data_uri, article_id),
    )


def release_claim(cur, article_id: int):
    cur.execute(
        """
        UPDATE articles
        SET image_claimed_at = NULL,
            image_claimed_by = NULL
        WHERE id = %s AND image IS NULL
        """,
        (article_id,),
    )


def safe_filename(name: str) -> str:
    safe = "".join(c if c.isalnum() or c in ("-", "_") else "_" for c in name)
    return safe or "image"


def generate_image(prompt: str) -> Optional[tuple[str, bytes]]:
    model = load_pipeline()
    if model is None:
        return None
    try:
        img = model(
            prompt=prompt,
            num_inference_steps=8,
            guidance_scale=7,
        ).images[0]
        buf = BytesIO()
        img.save(buf, format="PNG")
        img_bytes = buf.getvalue()
        data = base64.b64encode(img_bytes).decode("ascii")
        return f"data:image/png;base64,{data}", img_bytes
    except Exception as e:
        logging.error("Bildgenerierung fehlgeschlagen: %s", e)
        return None


def main():
    while True:
        try:
            with connect() as conn:
                with conn.cursor() as cur:
                    rows = fetch_articles_without_image(cur)
                    if not rows:
                        logging.info("keine neuen Artikel ohne Bild")
                    for aid, title, article_id in rows:
                        prompt = title or "news illustration"
                        result = generate_image(prompt)
                        if result:
                            data_uri, img_bytes = result
                            update_image(cur, aid, data_uri)
                            logging.info("Bild gesetzt für Artikel %s", aid)
                        else:
                            release_claim(cur, aid)
                conn.commit()
        except Exception as e:
            logging.error("Fehler im Worker: %s", e)

        time.sleep(30)


if __name__ == "__main__":
    main()