siftlode/backend/app/worker.py

"""Download worker process entry point.

Runs in a DEDICATED container (`command: python -m app.worker`, `WORKER_ENABLED=1`), separate
from the API so long yt-dlp downloads + ffmpeg postprocessing never block web requests.

Design:
  * N loop threads (download_worker_concurrency) each claim a `queued` DownloadJob with
    FOR UPDATE SKIP LOCKED and process it.
  * The shared MediaAsset is the dedup unit. A job whose asset is already `ready` finishes
    instantly (no re-download); a fresh asset is claimed pending→downloading by exactly one
    worker (a DB compare-and-set), so two jobs for the same video+format never double-download.
  * Live progress is written to the job row (short-lived sessions, throttled ~1/s) — the
    frontend polls it; the worker can't reach the API's in-process WebSocket registry.
  * Pause/cancel are cooperative: the progress hook re-reads the job status and aborts the
    yt-dlp download if it's no longer `running`.
  * Crash-safe: on startup, orphaned `running` jobs are requeued and `downloading` assets reset.
"""

import os
import shutil
import signal
import subprocess
import sys
import threading
import time
from collections.abc import Callable
from datetime import datetime, timezone

from sqlalchemy import text

from app import logs
from app import logsink
from app import sysconfig
from app.config import settings
from app.db import SessionLocal
from app.downloads import edit as editmod
from app.downloads import credentials, formats, quota, service
from app.downloads import errors as dlerrors
from app.downloads.jobdb import (
    _edit_staging_dir,
    _requeue_if_running,
    _reset_asset_pending,
    _staging_dir,
)
from app.downloads.status import STOPPED_BY_ROUTE, JobStatus
from app.models import DISPLAY_TEXT_MAX, DownloadJob, MediaAsset

log = logs.get_logger(__name__)
_stop = threading.Event()


def _handle_signal(signum, frame):
    _stop.set()


# --- claim + recovery -----------------------------------------------------------------------


def _wait_for_schema(timeout: float = 300.0) -> bool:
    """Block until the download tables exist before doing any DB work. Returns False if they never
    appeared (or we were asked to stop while waiting).

    The API applies migrations on its OWN startup, and the worker is a separate container that may
    boot first (they start in parallel). Without this, the worker would hit a missing `download_jobs`
    table and crash-loop until the API's migrations land. Poll patiently instead — and if the schema
    is still absent when the timeout runs out, say so and let the caller exit, because "continuing"
    only moved the same crash one call further down (into `_recover_orphans`), where it read as an
    unrelated failure."""
    deadline = time.monotonic() + timeout
    while not _stop.is_set():
        try:
            with SessionLocal() as db:
                db.execute(text("SELECT 1 FROM download_jobs LIMIT 1"))
            return True
        except Exception:  # noqa: BLE001 — table not migrated yet / DB not up yet
            if time.monotonic() > deadline:
                # The traceback, not `str(exc)[:120]`: this is the last thing the worker says
                # before it exits, and which driver call failed is the whole question.
                log.exception(
                    "download schema still absent after %ss — the API's migrations have not "
                    "landed; exiting so the container restarts",
                    timeout,
                )
                return False
            log.info("waiting for the database schema (API migrations)…")
            _stop.wait(3.0)
    return False


def _recover_orphans() -> None:
    """Reset state left behind by a previous crash so nothing is stuck."""
    with SessionLocal() as db:
        db.execute(text("UPDATE download_jobs SET status='queued' WHERE status='running'"))
        db.execute(text("UPDATE media_assets SET status='pending' WHERE status='downloading'"))
        db.commit()


def _claim_job() -> int | None:
    """Claim the oldest queued job whose user is under their per-user concurrency limit AND whose
    shared asset isn't already being downloaded by another worker.

    The asset exclusion (`NOT EXISTS … media_assets.status='downloading'`) stops the same-asset ping-
    pong (B-sub-3): without it, a second worker claims a queued job that shares an in-flight asset,
    finds it busy, requeues it, and immediately re-claims it — burning a worker slot behind one
    in-flight asset and starving everyone else. Excluded here, that job simply waits until its asset
    settles, so `_process_job` no longer needs to sleep on the busy/lost-race paths.

    Users already AT their concurrency limit are excluded IN THE SQL, so one user flooding the
    queue can't monopolize the fetch window and starve everyone behind them. (The old fixed
    LIMIT-50 window did exactly that: 50 queued jobs from a single capped user filled the window,
    `chosen` came back None, and every other user's work — sitting past position 50 — was never
    seen.) The per-user limit is the quota row's `max_concurrent` (else the sysconfig default);
    `unlimited` users are never capped. A final Python `at_concurrency_limit` re-check still guards
    the tiny race between the exclusion subquery and the claim (it can only ever reject, never
    over-claim), so correctness never depends on the two observations agreeing."""
    with SessionLocal() as db:
        default_mc = sysconfig.get_int(db, "download_default_max_concurrent")
        rows = db.execute(
            text(
                """
                SELECT id, user_id FROM download_jobs
                WHERE status='queued'
                  AND NOT EXISTS (
                      SELECT 1 FROM media_assets a
                      WHERE a.id = download_jobs.asset_id AND a.status = 'downloading'
                  )
                  AND user_id NOT IN (
                      SELECT r.user_id
                      FROM download_jobs r
                      LEFT JOIN download_quotas q ON q.user_id = r.user_id
                      WHERE r.status='running' AND r.user_id IS NOT NULL
                      GROUP BY r.user_id, q.max_concurrent, q.unlimited
                      HAVING COALESCE(q.unlimited, false) = false
                         AND count(*) >= GREATEST(COALESCE(q.max_concurrent, :dmc), 1)
                  )
                ORDER BY queue_pos, id
                FOR UPDATE SKIP LOCKED
                LIMIT 50
                """
            ),
            {"dmc": default_mc},
        ).fetchall()
        chosen = next((jid for jid, uid in rows if not quota.at_concurrency_limit(db, uid)), None)
        if chosen is None:
            db.commit()  # release the row locks; nothing eligible right now
            return None
        db.execute(
            text(
                "UPDATE download_jobs SET status='running', phase='starting', updated_at=now() "
                "WHERE id = :id"
            ),
            {"id": chosen},
        )
        db.commit()
        return chosen


def _claim_asset(asset_id: int) -> bool:
    """Compare-and-set pending→downloading. True if THIS worker won the right to download."""
    with SessionLocal() as db:
        row = db.execute(
            text(
                "UPDATE media_assets SET status='downloading' "
                "WHERE id = :id AND status='pending' RETURNING id"
            ),
            {"id": asset_id},
        ).fetchone()
        db.commit()
        return row is not None


# --- supervised child execution + watchdog --------------------------------------------------
# The heavy work runs in a child process so a wedged yt-dlp/ffmpeg is HARD-killable. A worker thread
# can't be killed, a C-level spin (e.g. a catastrophic regex) can't even be cooperatively cancelled,
# and one such thread starves the rest via the GIL — the incident that motivated this. The parent
# thread supervises: it kills the child (and its process group, so ffmpeg grandchildren die too) on
# a route pause/cancel, on a no-progress stall, or past an absolute wall-clock cap.
#
# POSIX-only: the kill path uses process groups (start_new_session / os.killpg / os.getpgid), so the
# worker is Linux-container-only — never run `python -m app.worker` natively on Windows.
# Edits ALSO carry a second, tighter watchdog inside run_ffmpeg (a per-ffmpeg stall timeout with a
# cleaner "ffmpeg stalled" error); this process-level supervisor is the uniform backstop for both the
# download and edit paths, and the last line of defence for anything run_ffmpeg can't see.

_CHILD_OK = 0


def _spawn_download_child(job_id: int) -> subprocess.Popen:
    """Start the child that does the actual download/edit for `job_id`, in its OWN session/process
    group so stopping it also stops any ffmpeg / helper processes yt-dlp spawned."""
    return subprocess.Popen(
        [sys.executable, "-m", "app.downloads.runner", str(job_id)],
        start_new_session=True,
    )


def _job_liveness(job_id: int) -> tuple[str | None, float | None]:
    """(status, seconds-since-last-row-update) for the job, computed with the DB clock so there is
    no parent↔DB skew. A live download rewrites its row ~1/s; a stalled one never does. Returns
    (None, None) if the row is gone."""
    with SessionLocal() as db:
        row = db.execute(
            text(
                "SELECT status, EXTRACT(EPOCH FROM (now() - updated_at)) "
                "FROM download_jobs WHERE id = :id"
            ),
            {"id": job_id},
        ).one_or_none()
    if row is None:
        return None, None
    return row[0], (float(row[1]) if row[1] is not None else None)


def _signal_group(proc: subprocess.Popen, pgid: int | None, sig: int) -> None:
    """Signal the child's whole process group (falling back to the child alone), swallowing the race
    where it has already exited."""
    try:
        if pgid is not None:
            os.killpg(pgid, sig)
        else:
            proc.send_signal(sig)
    except ProcessLookupError:
        pass  # silent-ok: the child already exited — the race this function is written around
    except OSError as exc:
        # Any OTHER refusal means the signal did not land and the child is still running, which
        # is how a "cancelled" download keeps writing to a staging tree nobody is watching.
        log.info(
            "a download child could not be signalled",
            extra=logs.event(
                "download.child.signal_failed", signal=int(sig), reason=type(exc).__name__
            ),
        )


def _stop_child(proc: subprocess.Popen, grace_s: float) -> None:
    """SIGTERM the child's process GROUP, then SIGKILL if it outlives the grace period. Shares the
    terminate→wait→kill escalation with run_ffmpeg (`editmod.escalate_kill`); the only difference is
    that we signal the whole group so ffmpeg grandchildren die too. `escalate_kill` no-ops if the
    child already exited, and `getpgid` yields None for a reaped pid, so no separate poll guard."""
    try:
        pgid: int | None = os.getpgid(proc.pid)
    # silent-ok: `getpgid` refuses for a reaped pid, which the docstring above already treats as
    # the ordinary case — the escalation then signals the child alone and no-ops if it is gone.
    except OSError:
        pgid = None
    editmod.escalate_kill(
        proc,
        terminate=lambda: _signal_group(proc, pgid, signal.SIGTERM),
        kill=lambda: _signal_group(proc, pgid, signal.SIGKILL),
        grace_s=grace_s,
        label="download child",
    )


def _stop_and_free(job_id: int, asset_id: int, proc: subprocess.Popen, grace_s: float) -> None:
    """Kill the child, free its claimed asset, and clear its staging — the common tail of every path
    that stops a child (a route pause/cancel, a shutdown, an over-cap abandon)."""
    _stop_child(proc, grace_s)
    _reset_asset_pending(asset_id)
    _cleanup_staging(asset_id)


def _fail_after_kill(
    job_id: int,
    asset_id: int,
    proc: subprocess.Popen,
    grace_s: float,
    error: str,
    error_code: str,
    cause: str | None = None,
) -> bool:
    """`_stop_and_free` + mark the job errored — for the watchdog's stall/timeout kills and the
    supervisor-error fallback. Returns whether the job failed here (`_fail_running`)."""
    _stop_and_free(job_id, asset_id, proc, grace_s)
    return _fail_running(job_id, error=error, error_code=error_code, cause=cause)


def _fail_running(
    job_id: int,
    *,
    error: str,
    error_code: str | None,
    error_detail: str | None = None,
    cause: str | None = None,
    exit_code: int | None = None,
) -> bool:
    """Fail a claimed job — only while it is still `running` (`service.finalize_if_running`), so a
    cancel/pause a route landed first is never turned back into `error`, which holds the asset the
    route already released. `error_detail` is written every time: a failure never shows an earlier
    failure's explanation.

    `cause` is the worker's half of `download.job.failed`: the failures the child never got to
    report because it was killed, never ran or died — a stall, the wall-clock cap, a non-zero exit
    (with its `exit_code`). Without it those left no line at all. A caller that logs its own line
    (a worker error, with its stack — `_report_worker_error`) passes none and uses the returned
    verdict. Logged only when the job actually failed here: a route's cancel that landed first is
    not a failure. Returns whether it failed here."""
    with SessionLocal() as db:
        failed = service.finalize_if_running(
            db,
            job_id,
            status=JobStatus.ERROR,
            error=error,
            error_code=error_code,
            error_detail=error_detail,
            phase=None,
        )
        db.commit()
    if cause is not None and failed:
        log.warning(
            "job %s failed in the worker: %s",
            job_id,
            cause,
            extra=logs.event(
                "download.job.failed",
                job_id=job_id,
                error_code=error_code,
                cause=cause,
                **({"exit_code": exit_code} if exit_code is not None else {}),
            ),
        )
    return failed


def _report_worker_error(job_id: int, exc: Exception, cause: str, fail: Callable[[], bool]) -> None:
    """A worker error on a job (the supervisor loop, or processing the job, raised): fail the job,
    then ONE line with the stack. Its event says what the error did to the job —
    `download.job.failed` when the job failed with it, `download.worker.error` when a route had
    already stopped it (a cancel that landed first is not a failure, `_fail_running`).

    The line goes out even when failing the job raises (the database is down): as
    `download.worker.error`, since the job was not failed, and the raise goes on to the caller —
    without `exc` as its context, which this line has already carried with its stack."""
    failed = False
    try:
        failed = fail()
    except BaseException as write_error:
        write_error.__suppress_context__ = True
        raise
    finally:
        log.error(
            "job %s: worker error (%s)",
            job_id,
            cause,
            exc_info=exc,
            extra=logs.event(
                "download.job.failed" if failed else "download.worker.error",
                job_id=job_id,
                cause=cause,
            ),
        )


def _run_supervised(job_id: int, asset_id: int) -> None:
    """Run the heavy download/edit for an already-claimed asset in a child process and watchdog it.

    A clean child exit means the child already wrote the terminal job/asset state (done/error, or a
    cooperative pause via the in-child progress hook) — we leave it. We step in only when the child
    can't finish or self-abort: a route paused/cancelled it (kill, free the asset, keep the route's
    status), it made no progress for `stall_timeout` or ran past `wall_clock_max` (kill, fail the
    job), the worker is shutting down (kill, requeue), or it crashed without setting state."""
    poll_s = max(1, settings.download_supervisor_poll_seconds)
    stall_s = settings.download_stall_timeout_seconds
    wall_s = settings.download_wall_clock_max_seconds
    grace_s = settings.download_kill_grace_seconds

    try:
        proc = _spawn_download_child(job_id)
    except Exception:  # noqa: BLE001 — spawn failed; don't leave the asset stuck 'downloading'
        log.exception("failed to spawn download child for job %s", job_id)
        _reset_asset_pending(asset_id)
        _requeue_if_running(job_id, None)
        return

    started = time.monotonic()
    reason: str | None = None
    try:
        while True:
            try:
                proc.wait(timeout=poll_s)
                break  # child finished on its own
            # silent-ok: the timeout is the loop's clock, not a failure — it is what gives this
            # thread a turn to check whether a route cancelled the job.
            except subprocess.TimeoutExpired:
                pass
            status, stale = _job_liveness(job_id)
            if status is None or status in STOPPED_BY_ROUTE:
                reason = "cancelled"  # a route stopped it (or the row vanished)
                break
            if status == "running":
                if stale is not None and stale > stall_s:
                    reason = "stalled"
                    break
                if _stop.is_set():
                    reason = "shutdown"
                    break
            # Absolute wall-clock cap frees the worker slot even for a FINISHED child wedged in
            # teardown. A still-running job over the cap is failed ("timeout"); a terminal one (already
            # done/error) just gets its lingering child killed, its DB status left intact ("abandon").
            if time.monotonic() - started > wall_s:
                reason = "timeout" if status == "running" else "abandon"
                break
            # else: a terminal status under the cap (done/error, or a re-queued contended asset) means
            # the child is exiting on its own — keep polling; NEVER kill a finishing child (that could
            # drop its completion notice).
    except Exception as exc:  # noqa: BLE001 — a watchdog error must NEVER leave the child running unsupervised
        # silent-ok: not silent — `_report_worker_error` is the report.
        _report_worker_error(
            job_id,
            exc,
            "supervisor_error",
            lambda: _fail_after_kill(
                job_id,
                asset_id,
                proc,
                grace_s,
                "The download was stopped after a worker error.",
                "unknown",
            ),
        )
        return

    if reason is None:
        # Child exited. Reconcile the asset we claimed: free it if the child left it non-finalized —
        # a crash/OOM, or an early return on a job deleted mid-flight — a no-op when it finalized
        # normally. Then, if it died non-zero while the job is still 'running', fail the job so it is
        # not left stuck. Guarded: the child is already dead, so a DB blip here must not propagate.
        try:
            _reset_asset_pending(asset_id)
            if proc.returncode != _CHILD_OK:
                status, _ = _job_liveness(job_id)
                if status == "running":
                    _cleanup_staging(asset_id)
                    _fail_running(
                        job_id,
                        error=f"The download process exited unexpectedly (code {proc.returncode}).",
                        error_code="unknown",
                        cause="child_exit",
                        exit_code=proc.returncode,
                    )
        except Exception:  # noqa: BLE001 — boot recovery is the backstop for a stuck 'running' job
            log.exception("post-exit reconciliation failed for job %s", job_id)
        return

    if reason == "stalled":
        _fail_after_kill(
            job_id,
            asset_id,
            proc,
            grace_s,
            f"The download stalled (no progress for {int(stall_s)}s) and was stopped.",
            "network",
            cause="stalled",
        )
        return
    if reason == "timeout":
        _fail_after_kill(
            job_id,
            asset_id,
            proc,
            grace_s,
            f"The download exceeded the {int(wall_s)}s time limit and was stopped.",
            "network",
            cause="timeout",
        )
        return
    # cancelled / abandon / shutdown: stop + free the asset, but leave the job's own DB status.
    _stop_and_free(job_id, asset_id, proc, grace_s)
    if reason == "shutdown":
        _requeue_if_running(job_id, None)  # resume on the next boot
    elif reason == "abandon":
        log.warning(
            "job %s: killed a child wedged past the wall-clock; DB status left intact",
            job_id,
            extra=logs.event("download.child.abandoned", job_id=job_id),
        )
    else:  # cancelled
        # Not `download.job.aborted`: that is the child's line, logged when it notices the pause
        # itself — and it often has, and is only tearing down when this poll comes round. This one
        # says the worker had to stop it.
        log.info(
            "job %s: child stopped (route pause/cancel)",
            job_id,
            extra=logs.event("download.child.stopped", job_id=job_id),
        )


def _cleanup_staging(asset_id: int) -> None:
    """Best-effort removal of a killed child's leftover staging dirs (its own `finally` never ran).
    The next attempt on this asset also wipes staging on start, so this is just tidiness/disk."""
    for path in (_staging_dir(asset_id), _edit_staging_dir(asset_id)):
        shutil.rmtree(path, ignore_errors=True)


# --- processing -----------------------------------------------------------------------------


def _process_job(job_id: int) -> None:
    """Dispatch a claimed job. The trivial outcomes (cache hit, dead/contended asset) are handled
    inline and cheaply; the actual heavy work (yt-dlp download / ffmpeg edit) is claimed here and
    then run in a supervised CHILD PROCESS (`_run_supervised`) so a wedged yt-dlp/ffmpeg can be
    hard-killed — a worker THREAD cannot be killed, nor (when it spins in a C call) even
    cooperatively cancelled, and one such thread starves the others via the GIL."""
    with SessionLocal() as db:
        job = db.get(DownloadJob, job_id)
        if job is None:
            return
        asset = db.get(MediaAsset, job.asset_id) if job.asset_id else None
        asset_id = asset.id if asset else None
        asset_status = asset.status if asset else None
        asset_error = asset.error if asset else None
        if asset_status == "error":
            # A job that meets an already-failed asset gets that failure's explanation — not a bare
            # "Download failed". Only the text is on the asset, so it is diagnosed again, for this
            # link, with whether a cookie goes with it asked of the same gate the download child
            # used (`credentials.usable_cookie`) — as it stands now, so the advice fits this job's Retry.
            url = service.source_url(job.source_kind, job.source_ref)
            cookie = credentials.usable_cookie(
                db, job.user_id, url, asset, formats.normalize(job.profile_snapshot)
            )
            diagnosis = dlerrors.diagnose(asset_error, url=url, cookie_attached=cookie is not None)
            del cookie  # only whether there is one matters here

    if asset is None:
        _fail_running(job_id, error="No media asset", error_code=None, cause="no_asset")
        return

    if asset_status == "ready":
        _finish_from_cache(job_id, asset_id)
        return
    if asset_status == "error":
        _fail_running(
            job_id,
            error=asset_error or "Download failed",
            error_code=diagnosis.code,
            error_detail=diagnosis.detail,
            cause="asset_failed",
        )
        return
    if asset_status == "downloading":
        # Another worker won this asset between our claim and here (a narrow race — the claim SQL
        # excludes downloading-asset jobs). Requeue and let it be reclaimed once the asset settles;
        # no sleep — the claim exclusion keeps it out of the window until then, so no busy-loop.
        _requeue_if_running(job_id, "waiting")
        return

    # asset_status == "pending": try to win the download / edit.
    if not _claim_asset(asset_id):
        # Lost the pending->downloading race; requeue (same reasoning — the exclusion prevents a spin).
        _requeue_if_running(job_id, "waiting")
        return

    # The asset is now ours ('downloading'): run the heavy work in a supervised child process.
    _run_supervised(job_id, asset_id)


def _finish_from_cache(job_id: int, asset_id: int) -> None:
    with SessionLocal() as db:
        job = db.get(DownloadJob, job_id)
        asset = db.get(MediaAsset, asset_id)
        if job and asset:
            fields = {"status": JobStatus.DONE, "progress": 100, "phase": None}
            if not job.display_name and asset.title:
                # display_name is VARCHAR(255); a title can be longer (e.g. a recipe in the title).
                fields["display_name"] = asset.title[:DISPLAY_TEXT_MAX]
            asset.last_access_at = datetime.now(timezone.utc)
            # Atomic finalize: only marks the job done while it is still running, so a concurrent
            # cancel/pause is never resurrected (see service.finalize_if_running).
            service.finalize_if_running(db, job_id, **fields)
        db.commit()


# --- loop -----------------------------------------------------------------------------------


def _worker_loop(idx: int) -> None:
    while not _stop.is_set():
        try:
            job_id = _claim_job()
        except Exception:  # noqa: BLE001
            log.exception("claim failed")
            _stop.wait(settings.download_poll_seconds)
            continue
        if job_id is None:
            _stop.wait(settings.download_poll_seconds)
            continue
        # Tag every line this job produces (including the crash below) with the job id: a worker runs
        # N jobs at once on N threads, so without it the interleaved output is unreadable. The bind
        # wraps the handler too — bound inside the `try`, it was already reset by the time the crash
        # line was logged.
        with logs.bind(job_id=job_id):
            try:
                _process_job(job_id)
            except Exception as exc:  # noqa: BLE001
                try:
                    _report_worker_error(
                        job_id,
                        exc,
                        "crash",
                        lambda: _fail_running(job_id, error="Worker error", error_code=None),
                    )
                # silent-ok: not silent — `_report_worker_error` logged the crash before this
                # re-raise. Failing the job hit the database, and letting that out would end this
                # worker thread for good; boot recovery settles a job left `running`.
                except Exception:  # noqa: BLE001
                    log.exception(
                        "could not fail job %s after its crash",
                        job_id,
                        extra=logs.event("download.worker.fail_write_failed", job_id=job_id),
                    )


def main() -> None:
    logs.configure_logging("worker")
    signal.signal(signal.SIGTERM, _handle_signal)
    signal.signal(signal.SIGINT, _handle_signal)

    if not settings.worker_enabled:
        log.info("WORKER_ENABLED is off; worker exiting (nothing to do).")
        return

    # The API may still be applying migrations when we boot. If they never arrive, exit instead of
    # pressing on into a guaranteed crash — with a FAILURE status, so the container comes back under
    # any restart policy (`on-failure` too, not just the `unless-stopped` our composes ship) and the
    # exit code doesn't claim success for a worker that never started working.
    if not _wait_for_schema():
        if _stop.is_set():
            return  # asked to shut down while waiting — that IS a clean exit
        raise SystemExit(1)
    # After the wait, not before: the DB is reachable now. The wait checks `download_jobs`, not
    # `log_events`, so on the deploy that first adds the table the worker may start before the API
    # has migrated — those few seconds of batches fail and say so on stderr; nothing else breaks.
    logsink.start()
    _recover_orphans()
    n = max(1, settings.download_worker_concurrency)
    log.info("Download worker started (concurrency=%d, root=%s).", n, settings.download_root)
    threads = [threading.Thread(target=_worker_loop, args=(i,), daemon=True) for i in range(n)]
    for t in threads:
        t.start()
    while not _stop.is_set():
        _stop.wait(1.0)
    log.info("Download worker stopping…")
    for t in threads:
        t.join(timeout=5)


if __name__ == "__main__":
    main()