GTI_GERENCIADOR_FISCAL/fiscal/central/app/jobs.py

"""Job queue for Agent orchestration: create (idempotent per kind), claim one job at
a time per Agent, complete, fail. Also the central sync scheduler helpers.

Every fiscal job is READ-ONLY on the ERP (INITIAL_SYNC, SYNC_INCREMENTAL,
REPROCESS_ALL, REPROCESS_PRODUCTS); automation never authorizes Firebird writes.
"""
from datetime import datetime, timedelta, timezone

from sqlalchemy import select, update, func, text

from .config import settings
from .models import Installation, Job

FULL_KINDS = ("INITIAL_SYNC", "REPROCESS_ALL")           # read every product, reclassify all
SYNC_KINDS = FULL_KINDS + ("SYNC_INCREMENTAL",)          # jobs that count as "a sync happened"
_VALID_KINDS = set(SYNC_KINDS) | {"REPROCESS_PRODUCTS"}
_MAX_ERROR_LENGTH = 2000
STALE_RUNNING_AFTER = timedelta(hours=1)                  # an Agent that died mid-job must not block the queue forever


def _aware(value):
    if value is not None and value.tzinfo is None:
        return value.replace(tzinfo=timezone.utc)
    return value


def create_job(session, agent, kind, payload=None):
    """Create a job. Sync-type jobs are idempotent: if this Agent already has a
    PENDING/RUNNING job that covers the request, that job is returned instead
    (no duplicates, never two full jobs at once)."""
    if kind not in _VALID_KINDS:
        raise ValueError(f"unsupported job kind: {kind}")
    if kind in ("REPROCESS_ALL", "REPROCESS_PRODUCTS", "SYNC_INCREMENTAL"):
        _flag_enrichment_reevaluation(session, agent)          # "Reprocessar"/"Sincronizar" also re-evaluate the tracked, not-yet-enriched products (KB/cache first), even when an equivalent job is already queued
        session.commit()
    if kind in SYNC_KINDS:
        # A full job subsumes an incremental one, not the other way round: asking for a
        # full reprocess while only an incremental is queued must still queue the full.
        blocking = FULL_KINDS if kind in FULL_KINDS else SYNC_KINDS
        existing = session.scalar(select(Job).where(Job.agent_id == agent.id, Job.kind.in_(blocking), Job.status.in_(("PENDING", "RUNNING"))).order_by(Job.id).limit(1))
        if existing is not None:
            return existing
    job = Job(agent_id=agent.id, kind=kind, status="PENDING", payload=payload or {})
    session.add(job)
    session.commit()
    return job


def _flag_enrichment_reevaluation(session, agent):
    try:
        installation = session.scalar(select(Installation).where(Installation.agent_id == agent.id))
        if installation is not None and installation.features_configured and installation.enrichment_mode != "DISABLED":
            from .enrichment.status import mark_for_reevaluation
            mark_for_reevaluation(session, installation.id)
    except Exception:  # noqa: BLE001 - never let the enrichment bookkeeping break job creation
        pass


def recover_stale_running(session, agent, now=None):
    now = now or datetime.now(timezone.utc)
    cutoff = now - STALE_RUNNING_AFTER
    stale = session.scalars(select(Job).where(Job.agent_id == agent.id, Job.status == "RUNNING")).all()
    changed = False
    for job in stale:
        if job.started_at is not None and _aware(job.started_at) < cutoff:
            job.status, job.finished_at, job.error = "FAILED", now, "stale: Agent stopped reporting while RUNNING"
            changed = True
    if changed:
        session.commit()


def claim_next_job(session, agent):
    """Atomically claim the oldest PENDING job for this agent only — and only when
    processing is enabled centrally and no other job of this Agent is RUNNING
    (one fiscal job at a time per Firebird)."""
    installation = session.scalar(select(Installation).where(Installation.agent_id == agent.id))
    if installation is not None and (not installation.processing_enabled or not installation.features_configured or not installation.corrector_enabled):
        return None                    # WAITING_FEATURE_CONFIGURATION / corrector switched off centrally: nothing is handed to the Agent
    recover_stale_running(session, agent)
    if session.scalar(select(Job.id).where(Job.agent_id == agent.id, Job.status == "RUNNING").limit(1)) is not None:
        return None
    pending = session.execute(
        select(Job.id, Job.kind)
        .where(Job.agent_id == agent.id, Job.status == "PENDING")
        .order_by(Job.created_at.asc(), Job.id.asc())
        .limit(10)
    ).all()
    if not pending:
        return None
    limit = settings.heavy_job_concurrency_limit
    heavy_slot_free = True
    if limit > 0 and any(kind in FULL_KINDS for _, kind in pending):
        if session.bind is not None and session.bind.dialect.name == "postgresql":
            session.execute(text("SELECT pg_advisory_xact_lock(7305001)"))      # serialize slot accounting across concurrent pollers; released at commit
        running_heavy = session.scalar(select(func.count()).select_from(Job).where(Job.status == "RUNNING", Job.kind.in_(FULL_KINDS),
                                                                                   Job.started_at > datetime.now(timezone.utc) - STALE_RUNNING_AFTER)) or 0
        heavy_slot_free = running_heavy < limit
    # oldest job this Agent may run now: a light job is never held behind a heavy one that is waiting for a slot
    candidate = next((job_id for job_id, kind in pending if kind not in FULL_KINDS or heavy_slot_free), None)
    if candidate is None:
        session.rollback()          # releases the advisory lock; the heavy job stays PENDING (queue) and the Agent keeps polling
        return None
    now = datetime.now(timezone.utc)
    result = session.execute(
        update(Job)
        .where(Job.id == candidate, Job.status == "PENDING")
        .values(status="RUNNING", started_at=now, attempts=Job.attempts + 1)
    )
    session.commit()
    if result.rowcount != 1:
        return None  # another poller won the race; caller can poll again
    return session.get(Job, candidate)


def complete_job(session, job, result_summary):
    result = session.execute(
        update(Job)
        .where(Job.id == job.id, Job.status == "RUNNING")
        .values(status="COMPLETED", finished_at=datetime.now(timezone.utc), result_summary=result_summary)
    )
    session.commit()
    return result.rowcount == 1


def fail_job(session, job, error):
    sanitized = str(error)[:_MAX_ERROR_LENGTH]
    result = session.execute(
        update(Job)
        .where(Job.id == job.id, Job.status == "RUNNING")
        .values(status="FAILED", finished_at=datetime.now(timezone.utc), error=sanitized)
    )
    session.commit()
    return result.rowcount == 1


def schedule_automatic_sync(session, agent, installation, snapshot, now=None):
    """Called from the Agent heartbeat. Central scheduler, idempotent:
    - first heartbeat of an installation with no initial coverage and no sync job
      ever created -> exactly ONE INITIAL_SYNC;
    - otherwise, if processing is enabled and the last sync attempt is older than
      the installation's interval, and nothing is queued/running -> ONE SYNC_INCREMENTAL.
    Returns the created job or None."""
    if installation is None or not installation.processing_enabled or not installation.features_configured or not installation.corrector_enabled:
        return None
    now = now or datetime.now(timezone.utc)
    sync_jobs = session.scalars(select(Job).where(Job.agent_id == agent.id, Job.kind.in_(SYNC_KINDS)).order_by(Job.id.desc())).all()
    if any(job.status in ("PENDING", "RUNNING") for job in sync_jobs):
        return None
    never_synced = not any(job.kind == "INITIAL_SYNC" or job.kind == "REPROCESS_ALL" for job in sync_jobs) and not int(getattr(snapshot, "total_products", 0) or 0)
    if never_synced and not sync_jobs:
        return create_job(session, agent, "INITIAL_SYNC", {"reason": "INITIAL_SYNC"})
    last_attempt = max((_aware(job.finished_at or job.created_at) for job in sync_jobs if (job.finished_at or job.created_at)), default=None)
    if last_attempt is not None and (now - last_attempt).total_seconds() >= installation.sync_interval_seconds:
        return create_job(session, agent, "SYNC_INCREMENTAL", {"reason": "SCHEDULED"})
    return None