feat(ocr): opt-in batch OCR mode via the gateway's async batch routes
Add DOCUMENT_OCR_MODE=sync|batch (default sync). In batch mode the tier-3 OCR
processor submits documents to the embedding gateway's async Batch OCR routes
(POST /v1/ocr/batch + GET /v1/ocr/batch/{job_id}, astrolabe-cloud-website#372)
for ~50% cheaper large-corpus backfill. The direct Mistral OCR path is left
untouched. Tracked on Deck #332.
Batch jobs run minutes-hours, so the OCR tier cannot block (the procrastinate
worker reclaims jobs in `doing` after INGEST_STALLED_JOB_SECONDS). Instead it
submits, records the gateway job id in a new per-tenant `batch_ocr_jobs` table
(procrastinate args are immutable across retries), and raises a BatchPending
signal that TieredEscalationStrategy turns into a same-queue deferred re-poll —
releasing the worker slot between polls. On completion the per-page markdown is
indexed like the sync path; a failure or a job past
DOCUMENT_OCR_BATCH_MAX_WAIT_SECONDS marks the document parse-failed.
Batch is opt-in and gateway-only: with the direct mistral backend, no gateway
URL, or the inline/memory pipeline (which can't defer), it falls back to sync.
One batch job per document (coalescing N docs/job is a follow-up).
- embedding/gateway_batch_client.py: submit/poll client (reuses GatewayTokenProvider).
- vector/batch_ocr_store.py + migration 008: job tracking (portable SQLite+PG).
- document_processors/escalation.py: BatchPending control-flow signal.
- document_processors/ocr.py: batch state machine + sync fallback.
- vector/processor.py: thread doc identity to the OCR tier; raise BatchPending
from the pending sentinel; propagate it as control flow (not a failure).
- vector/queue/procrastinate.py: BatchPending -> same-queue retry_in, exempt
from the transient cap (bounded by the processor's deadline).
- config + docs; tests across client/store/processor/strategy/parse-tier.
1653 unit tests pass; ruff + ty green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Chris Coutinho
co-authored by
Claude Opus 4.8
Claude Opus 4.8
parent
07ee91399b
commit
3b7e8d779b
@@ -0,0 +1,67 @@
|
||||
"""Add batch_ocr_jobs table for async batch OCR job tracking.
|
||||
|
||||
Deck #332 / embedding-gateway batch OCR (astrolabe-cloud-website#372). When
|
||||
``DOCUMENT_OCR_MODE=batch`` the OCR tier submits a document to the gateway's
|
||||
async ``POST /v1/ocr/batch`` and must re-poll ``GET /v1/ocr/batch/{job_id}``
|
||||
across procrastinate retries. procrastinate job args are immutable, so the
|
||||
gateway ``job_id`` (and submit time, for the poll deadline) are persisted here,
|
||||
keyed on the document + its content version (``etag``).
|
||||
|
||||
One row per in-flight job; the row is deleted once the job reaches a terminal
|
||||
state. Empty + unused unless batch mode is enabled (gateway-only), so OSS/SQLite
|
||||
self-hosters get an idle table and zero overhead.
|
||||
|
||||
Portable types only (Text + unix-epoch BigInteger timestamps, like the rest of
|
||||
this schema except the CP-queried usage_events) so the same migration runs on
|
||||
both self-host SQLite and cloud Postgres.
|
||||
|
||||
Revision ID: 008
|
||||
Revises: 007
|
||||
Create Date: 2026-06-15 12:00:00.000000
|
||||
"""
|
||||
|
||||
import sqlalchemy as sa
|
||||
|
||||
from alembic import op
|
||||
|
||||
revision = "008"
|
||||
down_revision = "007"
|
||||
branch_labels = None
|
||||
depends_on = None
|
||||
|
||||
|
||||
def upgrade() -> None:
|
||||
op.create_table(
|
||||
"batch_ocr_jobs",
|
||||
# Document identity (the same keys the OCR tier receives via the
|
||||
# processor ``options``). ``etag`` is the content-version key: a changed
|
||||
# document (new etag) is a new job, so a stale row never serves results
|
||||
# for the wrong content.
|
||||
sa.Column("user_id", sa.Text(), nullable=False),
|
||||
sa.Column("doc_id", sa.Text(), nullable=False),
|
||||
sa.Column("doc_type", sa.Text(), nullable=False),
|
||||
sa.Column("etag", sa.Text(), nullable=False),
|
||||
# The gateway's namespaced batch job id ("<provider>/<batch_job_id>") —
|
||||
# the only handle for polling (the gateway is stateless).
|
||||
sa.Column("job_id", sa.Text(), nullable=False),
|
||||
# Gateway-normalised status mirror (pending|succeeded|failed). Kept for
|
||||
# observability; the live decision always comes from a fresh poll.
|
||||
sa.Column("status", sa.Text(), nullable=False),
|
||||
# Unix-epoch seconds. ``submitted_at`` anchors the poll deadline
|
||||
# (DOCUMENT_OCR_BATCH_MAX_WAIT_SECONDS).
|
||||
sa.Column("submitted_at", sa.BigInteger(), nullable=False),
|
||||
sa.Column("updated_at", sa.BigInteger(), nullable=False),
|
||||
# One in-flight job per (document, content version). A resubmit for a new
|
||||
# etag inserts a new row; the superseded row is swept on resubmit.
|
||||
sa.UniqueConstraint(
|
||||
"user_id",
|
||||
"doc_id",
|
||||
"doc_type",
|
||||
"etag",
|
||||
name="uq_batch_ocr_jobs_doc",
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
def downgrade() -> None:
|
||||
op.drop_table("batch_ocr_jobs")
|
||||
@@ -175,6 +175,20 @@ _DEFAULTS: dict[str, Any] = {
|
||||
# 20-60s; raise/lower per tenant. Configurable so a tenant isn't stuck with
|
||||
# the 180s default when its gateway has its own shorter ceiling.
|
||||
"document_ocr_timeout_seconds": 180.0,
|
||||
# OCR execution mode (Deck #332). "sync" (default) transcribes inline via the
|
||||
# backend's synchronous path. "batch" routes to the gateway's async Batch OCR
|
||||
# job (~50% cheaper, minutes-hours latency) for large-corpus backfill — opt-in
|
||||
# and gateway-only; with the direct mistral backend or no gateway it falls
|
||||
# back to sync. The submit->defer-poll loop runs on the per-tier procrastinate
|
||||
# path; the inline/memory pool can't defer, so batch falls back to sync there.
|
||||
"document_ocr_mode": "sync",
|
||||
# Seconds between batch-job polls (the procrastinate re-enqueue delay). Each
|
||||
# poll re-runs the tier; keep it well above a few seconds.
|
||||
"document_ocr_batch_poll_seconds": 120,
|
||||
# Hard deadline (seconds from submit) after which a still-pending batch job is
|
||||
# abandoned and the document marked parse-failed (timeout). Matches the
|
||||
# gateway's 24h Batch timeout default.
|
||||
"document_ocr_batch_max_wait_seconds": 86400,
|
||||
# Observability
|
||||
"metrics_enabled": True,
|
||||
"metrics_port": 9090,
|
||||
@@ -352,6 +366,11 @@ _dynaconf = Dynaconf(
|
||||
Validator("DOCUMENT_CHUNK_SIZE", gte=1),
|
||||
Validator("DOCUMENT_PARSE_TIMEOUT_SECONDS", gte=1),
|
||||
Validator("DOCUMENT_OCR_TIMEOUT_SECONDS", gte=1),
|
||||
Validator("DOCUMENT_OCR_MODE", is_in=("sync", "batch")),
|
||||
# Poll cadence well above a few seconds (each poll re-runs the tier);
|
||||
# deadline at least one poll interval.
|
||||
Validator("DOCUMENT_OCR_BATCH_POLL_SECONDS", gte=5),
|
||||
Validator("DOCUMENT_OCR_BATCH_MAX_WAIT_SECONDS", gte=60),
|
||||
Validator("DOCUMENT_PARSE_MEM_LIMIT_MB", gte=128),
|
||||
# 0 disables the pre-parse PDF size cap; otherwise it must be positive.
|
||||
Validator("DOCUMENT_MAX_PDF_SIZE_MB", gte=0),
|
||||
@@ -859,6 +878,13 @@ class Settings:
|
||||
# parse timeout / httpx.Timeout; per-tenant tunable so a gateway with a
|
||||
# shorter ceiling isn't masked by the 180s default.
|
||||
document_ocr_timeout_seconds: float = 180.0
|
||||
# OCR execution mode: "sync" | "batch" (Deck #332). batch is opt-in,
|
||||
# gateway-only, and used for large-corpus backfill; it falls back to sync when
|
||||
# no gateway backend resolves or the path can't defer (inline/memory pool).
|
||||
document_ocr_mode: str = "sync"
|
||||
# Batch-job poll cadence (procrastinate re-enqueue delay) and hard deadline.
|
||||
document_ocr_batch_poll_seconds: int = 120
|
||||
document_ocr_batch_max_wait_seconds: int = 86400
|
||||
# OCR escalation triggers (tier-0), per-tenant tunable. A page is OCR-worthy
|
||||
# if near-empty (< min_page_chars) OR low text-quality (< min_text_quality)
|
||||
# OR (when detect_scanned, image-analysis only runs when OCR is enabled)
|
||||
|
||||
@@ -93,3 +93,26 @@ class EscalateError(Exception):
|
||||
super().__init__(
|
||||
f"escalate {from_tier}->{to_tier} (reason={reason})",
|
||||
)
|
||||
|
||||
|
||||
class BatchPending(Exception):
|
||||
"""Raised when a tier's work is in flight on an async backend and the worker
|
||||
should poll again later (Deck #332 — batch OCR).
|
||||
|
||||
Like :class:`EscalateError` it is a **control-flow signal, NOT a failure**:
|
||||
the document's batch OCR job is still running on the gateway, so the OCR tier
|
||||
submits it (or polls an existing job) and raises this to ask the procrastinate
|
||||
retry strategy to re-run the SAME job on the SAME queue after ``retry_in``
|
||||
seconds — releasing the worker slot meanwhile so a multi-minute/hour batch
|
||||
doesn't pin a worker (and isn't reclaimed as a stalled ``doing`` job).
|
||||
|
||||
It must propagate untouched to the retry strategy: never swallowed by a broad
|
||||
``except Exception`` on the indexing path, never counted as a drop/parse
|
||||
error, and never marks the placeholder failed (the doc isn't done yet).
|
||||
Unlike ``EscalateError`` it does NOT change queue — the job stays on its own
|
||||
(``ocr``) tier queue and is simply deferred.
|
||||
"""
|
||||
|
||||
def __init__(self, *, retry_in: int) -> None:
|
||||
self.retry_in = retry_in
|
||||
super().__init__(f"batch OCR pending (retry_in={retry_in}s)")
|
||||
|
||||
@@ -18,6 +18,7 @@ Mistral (if ``MISTRAL_API_KEY``). Both return GitHub-flavoured markdown + exact
|
||||
|
||||
import base64
|
||||
import logging
|
||||
import time
|
||||
from abc import ABC, abstractmethod
|
||||
from collections.abc import Awaitable, Callable
|
||||
from typing import Any
|
||||
@@ -35,6 +36,13 @@ logger = logging.getLogger(__name__)
|
||||
# configurable via DOCUMENT_OCR_TIMEOUT_SECONDS and resolved per call.
|
||||
_OCR_CONNECT_TIMEOUT_SECONDS = 10.0
|
||||
|
||||
# Sentinel keys on a ProcessingResult.metadata that mark "batch OCR job still in
|
||||
# flight — poll again later". The processor can't raise across the registry, so
|
||||
# it returns this sentinel and ``vector/processor._parse_pdf_tier`` translates it
|
||||
# into a ``BatchPending`` control-flow raise (same site as ``EscalateError``).
|
||||
OCR_BATCH_PENDING_KEY = "ocr_batch_pending"
|
||||
OCR_BATCH_RETRY_IN_KEY = "ocr_batch_retry_in"
|
||||
|
||||
|
||||
def _pages_to_text(
|
||||
pages: list[tuple[int, str]],
|
||||
@@ -64,6 +72,24 @@ def _pages_to_text(
|
||||
return "".join(parts), boundaries
|
||||
|
||||
|
||||
def _batch_identity(
|
||||
options: dict[str, Any] | None,
|
||||
) -> tuple[str, str, str, str] | None:
|
||||
"""Extract ``(user_id, doc_id, doc_type, etag)`` from the processor options
|
||||
the per-tier path threads in, or ``None`` if identity is absent (the inline
|
||||
pool, which can't defer a poll). ``etag`` may be empty (a file with no etag is
|
||||
still one tracked job keyed on "").
|
||||
"""
|
||||
if not options:
|
||||
return None
|
||||
user_id = options.get("user_id")
|
||||
doc_id = options.get("doc_id")
|
||||
doc_type = options.get("doc_type")
|
||||
if not user_id or not doc_id or not doc_type:
|
||||
return None
|
||||
return str(user_id), str(doc_id), str(doc_type), str(options.get("etag") or "")
|
||||
|
||||
|
||||
class _OcrBackend(ABC):
|
||||
@abstractmethod
|
||||
async def ocr(
|
||||
@@ -141,6 +167,54 @@ class _MistralOcrBackend(_OcrBackend):
|
||||
return _pages_to_text(pages)
|
||||
|
||||
|
||||
def _build_gateway_token_provider(settings: Settings) -> Any:
|
||||
"""Build the M2M ``GatewayTokenProvider`` from settings, or ``None`` when no
|
||||
client-id is configured (unauthenticated gateway). Shared by the sync OCR
|
||||
backend and the batch client so the M2M-triple validation lives in one place.
|
||||
"""
|
||||
if not settings.embedding_gateway_client_id:
|
||||
return None
|
||||
# Lazy import avoids a document_processors -> embedding cycle at load.
|
||||
from ..embedding.gateway_client import GatewayTokenProvider # noqa: PLC0415
|
||||
|
||||
# Explicit (not assert -- assert is stripped under `python -O`): the M2M
|
||||
# triple is all-or-nothing.
|
||||
if not settings.embedding_gateway_token_url:
|
||||
raise ValueError(
|
||||
"EMBEDDING_GATEWAY_TOKEN_URL is required when "
|
||||
"EMBEDDING_GATEWAY_CLIENT_ID is set"
|
||||
)
|
||||
if not settings.embedding_gateway_client_secret:
|
||||
raise ValueError(
|
||||
"EMBEDDING_GATEWAY_CLIENT_SECRET is required when "
|
||||
"EMBEDDING_GATEWAY_CLIENT_ID is set"
|
||||
)
|
||||
return GatewayTokenProvider(
|
||||
token_url=settings.embedding_gateway_token_url,
|
||||
client_id=settings.embedding_gateway_client_id,
|
||||
client_secret=settings.embedding_gateway_client_secret,
|
||||
scope=settings.embedding_gateway_scope,
|
||||
)
|
||||
|
||||
|
||||
def build_gateway_batch_client(settings: Settings) -> Any:
|
||||
"""Build a ``GatewayBatchOcrClient`` when the gateway is the OCR backend, else
|
||||
``None`` (so batch mode falls back to sync for provider=mistral / no gateway).
|
||||
Batch OCR is gateway-only — Mistral's Batch API is reached *through* the
|
||||
gateway's batch routes, never directly from the pod."""
|
||||
if settings.document_ocr_provider not in ("gateway", "auto"):
|
||||
return None
|
||||
if not settings.embedding_gateway_url:
|
||||
return None
|
||||
from ..embedding.gateway_batch_client import GatewayBatchOcrClient # noqa: PLC0415
|
||||
|
||||
return GatewayBatchOcrClient(
|
||||
settings.embedding_gateway_url,
|
||||
settings.document_ocr_model,
|
||||
_build_gateway_token_provider(settings),
|
||||
)
|
||||
|
||||
|
||||
def build_ocr_backend(settings: Settings) -> _OcrBackend | None:
|
||||
"""Select an OCR backend from settings, or None when none is available."""
|
||||
provider = settings.document_ocr_provider
|
||||
@@ -148,33 +222,10 @@ def build_ocr_backend(settings: Settings) -> _OcrBackend | None:
|
||||
return None
|
||||
|
||||
if provider in ("gateway", "auto") and settings.embedding_gateway_url:
|
||||
token_provider = None
|
||||
if settings.embedding_gateway_client_id:
|
||||
# Lazy import avoids a document_processors -> embedding cycle at load.
|
||||
from ..embedding.gateway_client import ( # noqa: PLC0415
|
||||
GatewayTokenProvider,
|
||||
)
|
||||
|
||||
# Explicit (not assert -- assert is stripped under `python -O`): the
|
||||
# M2M triple is all-or-nothing.
|
||||
if not settings.embedding_gateway_token_url:
|
||||
raise ValueError(
|
||||
"EMBEDDING_GATEWAY_TOKEN_URL is required when "
|
||||
"EMBEDDING_GATEWAY_CLIENT_ID is set"
|
||||
)
|
||||
if not settings.embedding_gateway_client_secret:
|
||||
raise ValueError(
|
||||
"EMBEDDING_GATEWAY_CLIENT_SECRET is required when "
|
||||
"EMBEDDING_GATEWAY_CLIENT_ID is set"
|
||||
)
|
||||
token_provider = GatewayTokenProvider(
|
||||
token_url=settings.embedding_gateway_token_url,
|
||||
client_id=settings.embedding_gateway_client_id,
|
||||
client_secret=settings.embedding_gateway_client_secret,
|
||||
scope=settings.embedding_gateway_scope,
|
||||
)
|
||||
return _GatewayOcrBackend(
|
||||
settings.embedding_gateway_url, settings.document_ocr_model, token_provider
|
||||
settings.embedding_gateway_url,
|
||||
settings.document_ocr_model,
|
||||
_build_gateway_token_provider(settings),
|
||||
)
|
||||
|
||||
if provider in ("mistral", "auto") and settings.mistral_api_key:
|
||||
@@ -215,6 +266,13 @@ class OcrProcessor(DocumentProcessor):
|
||||
# doesn't each build a backend (and fetch its own M2M token). Lazy-init:
|
||||
# anyio primitives must not be created at import time.
|
||||
self._backend_lock: anyio.Lock | None = None
|
||||
# Batch-mode (Deck #332): the gateway batch client is cached like the sync
|
||||
# backend so its GatewayTokenProvider keeps its M2M-token cache across
|
||||
# documents. ``_batch_fallback_warned`` rate-limits the "can't batch,
|
||||
# using sync" warning to once per pod.
|
||||
self._batch_client_resolved = False
|
||||
self._batch_client: Any = None
|
||||
self._batch_fallback_warned = False
|
||||
|
||||
@property
|
||||
def name(self) -> str:
|
||||
@@ -239,6 +297,19 @@ class OcrProcessor(DocumentProcessor):
|
||||
) = None,
|
||||
) -> ProcessingResult:
|
||||
settings = get_settings()
|
||||
|
||||
# Batch mode (Deck #332): submit to the gateway's async Batch OCR job and
|
||||
# poll across procrastinate retries. Returns a result (incl. the
|
||||
# "pending" sentinel) when handled, or None to fall back to the
|
||||
# synchronous path below (no gateway backend, or no per-doc identity — the
|
||||
# inline/memory pool can't defer a poll).
|
||||
if settings.document_ocr_mode == "batch":
|
||||
batch_result = await self._process_batch(
|
||||
content, content_type, filename, options, settings
|
||||
)
|
||||
if batch_result is not None:
|
||||
return batch_result
|
||||
|
||||
if not self._backend_resolved:
|
||||
if self._backend_lock is None:
|
||||
self._backend_lock = anyio.Lock()
|
||||
@@ -301,6 +372,149 @@ class OcrProcessor(DocumentProcessor):
|
||||
processor=self.name,
|
||||
)
|
||||
|
||||
async def _get_batch_client(self) -> Any:
|
||||
"""Cached gateway batch client (or ``None`` when batch isn't applicable —
|
||||
provider=mistral / no gateway). Resolved once under the backend lock so the
|
||||
token provider's M2M cache survives across documents."""
|
||||
if not self._batch_client_resolved:
|
||||
if self._backend_lock is None:
|
||||
self._backend_lock = anyio.Lock()
|
||||
async with self._backend_lock:
|
||||
if not self._batch_client_resolved: # double-checked
|
||||
self._batch_client = build_gateway_batch_client(get_settings())
|
||||
self._batch_client_resolved = True
|
||||
return self._batch_client
|
||||
|
||||
def _batch_fallback(self, reason: str, filename: str | None) -> None:
|
||||
"""Warn once that batch mode is falling back to the synchronous path."""
|
||||
if not self._batch_fallback_warned:
|
||||
logger.warning(
|
||||
"DOCUMENT_OCR_MODE=batch but %s; falling back to synchronous OCR",
|
||||
reason,
|
||||
)
|
||||
self._batch_fallback_warned = True
|
||||
|
||||
async def _process_batch(
|
||||
self,
|
||||
content: bytes,
|
||||
content_type: str,
|
||||
filename: str | None,
|
||||
options: dict[str, Any] | None,
|
||||
settings: Settings,
|
||||
) -> ProcessingResult | None:
|
||||
"""Submit + poll a one-document batch OCR job.
|
||||
|
||||
Returns a :class:`ProcessingResult` when batch handled the document — the
|
||||
terminal success/failure result, or the *pending sentinel* (``success=False``
|
||||
+ ``OCR_BATCH_PENDING_KEY`` metadata) that ``_parse_pdf_tier`` turns into a
|
||||
``BatchPending`` re-poll. Returns ``None`` to fall back to synchronous OCR
|
||||
(no gateway backend, or no per-doc identity — the inline pool can't defer).
|
||||
"""
|
||||
# Per-doc identity is threaded via ``options`` only on the per-tier
|
||||
# procrastinate path; the inline/memory pool omits it and can't defer a
|
||||
# poll, so batch is inapplicable there.
|
||||
identity = _batch_identity(options)
|
||||
if identity is None:
|
||||
self._batch_fallback("no per-document identity (inline path)", filename)
|
||||
return None
|
||||
client = await self._get_batch_client()
|
||||
if client is None:
|
||||
self._batch_fallback(
|
||||
"no gateway backend (provider=mistral or EMBEDDING_GATEWAY_URL unset)",
|
||||
filename,
|
||||
)
|
||||
return None
|
||||
|
||||
# Lazy import: keep the vector/DB stack off the document_processors load
|
||||
# path (mirrors the EscalateError lazy import in vector/processor).
|
||||
from ..vector.batch_ocr_store import BatchOcrJobStore # noqa: PLC0415
|
||||
|
||||
user_id, doc_id, doc_type, etag = identity
|
||||
store = await BatchOcrJobStore.shared()
|
||||
mime = content_type.split(";")[0].strip().lower()
|
||||
poll_seconds = settings.document_ocr_batch_poll_seconds
|
||||
|
||||
job = await store.get(
|
||||
user_id=user_id, doc_id=doc_id, doc_type=doc_type, etag=etag
|
||||
)
|
||||
if job is None:
|
||||
# New submission. Drop any superseded-version rows for this doc first
|
||||
# (a re-edited file changes etag), then submit + record.
|
||||
await store.delete_stale_for_doc(
|
||||
user_id=user_id, doc_id=doc_id, doc_type=doc_type, keep_etag=etag
|
||||
)
|
||||
job_id = await client.submit(content, mime, custom_id=doc_id)
|
||||
await store.insert_pending(
|
||||
user_id=user_id,
|
||||
doc_id=doc_id,
|
||||
doc_type=doc_type,
|
||||
etag=etag,
|
||||
job_id=job_id,
|
||||
)
|
||||
logger.info(
|
||||
"batch OCR job submitted for %s (job_id=%s); deferring poll",
|
||||
filename or doc_id,
|
||||
job_id,
|
||||
)
|
||||
return self._pending(poll_seconds)
|
||||
|
||||
# Existing job — poll the gateway.
|
||||
result = await client.poll(job.job_id)
|
||||
if result.is_pending:
|
||||
elapsed = int(time.time()) - job.submitted_at
|
||||
if elapsed >= settings.document_ocr_batch_max_wait_seconds:
|
||||
await store.delete(
|
||||
user_id=user_id, doc_id=doc_id, doc_type=doc_type, etag=etag
|
||||
)
|
||||
logger.warning(
|
||||
"batch OCR job %s exceeded max wait (%ss); marking failed",
|
||||
job.job_id,
|
||||
settings.document_ocr_batch_max_wait_seconds,
|
||||
)
|
||||
return ProcessingResult(
|
||||
text="",
|
||||
metadata={"parse_failed_reason": "timeout"},
|
||||
processor=self.name,
|
||||
success=False,
|
||||
error="batch OCR timed out",
|
||||
)
|
||||
return self._pending(poll_seconds)
|
||||
|
||||
# Terminal — drop the tracking row either way.
|
||||
await store.delete(user_id=user_id, doc_id=doc_id, doc_type=doc_type, etag=etag)
|
||||
if result.is_failed:
|
||||
logger.warning(
|
||||
"batch OCR job %s failed: %s", job.job_id, result.error or "unknown"
|
||||
)
|
||||
return ProcessingResult(
|
||||
text="",
|
||||
metadata={"parse_failed_reason": "error"},
|
||||
processor=self.name,
|
||||
success=False,
|
||||
error=f"batch OCR failed: {result.error or 'unknown'}",
|
||||
)
|
||||
text, boundaries = _pages_to_text(result.pages)
|
||||
return ProcessingResult(
|
||||
text=text,
|
||||
metadata={
|
||||
"page_count": len(boundaries),
|
||||
"page_boundaries": boundaries,
|
||||
"file_size": len(content),
|
||||
},
|
||||
processor=self.name,
|
||||
)
|
||||
|
||||
def _pending(self, retry_in: int) -> ProcessingResult:
|
||||
"""The pending sentinel — ``_parse_pdf_tier`` raises ``BatchPending`` from
|
||||
it. ``success=False`` keeps it out of the index path, and the sentinel key
|
||||
keeps it out of the parse-failed path (it isn't a failure)."""
|
||||
return ProcessingResult(
|
||||
text="",
|
||||
metadata={OCR_BATCH_PENDING_KEY: True, OCR_BATCH_RETRY_IN_KEY: retry_in},
|
||||
processor=self.name,
|
||||
success=False,
|
||||
)
|
||||
|
||||
async def health_check(self) -> bool:
|
||||
# Backends are resolved lazily (and configured per tenant), so there is
|
||||
# nothing to probe here without making a billable upstream call -- the
|
||||
|
||||
@@ -0,0 +1,171 @@
|
||||
"""Client for the embedding gateway's async **batch OCR** routes (Deck #332).
|
||||
|
||||
The gateway exposes two batch routes alongside the synchronous ``POST /v1/ocr``
|
||||
(astrolabe-cloud-website#372):
|
||||
|
||||
- ``POST /v1/ocr/batch`` — submit N documents (each with a caller ``custom_id``)
|
||||
as one Mistral Batch job; returns ``202`` + a namespaced ``job_id``
|
||||
(``<provider>/<batch_job_id>``).
|
||||
- ``GET /v1/ocr/batch/{job_id}`` — poll; returns the lifecycle status and, once
|
||||
terminal, per-document results (per-page markdown, or a per-document error).
|
||||
|
||||
The gateway is a **stateless passthrough** to Mistral's Batch API — the
|
||||
``job_id`` is the only handle, so the worker persists it (see
|
||||
``vector/batch_ocr_store``) and re-polls across procrastinate retries.
|
||||
|
||||
This client submits exactly **one document per job** (the v1 unit; coalescing N
|
||||
docs/job is a follow-up). Auth + ``/v1`` base-url handling mirror the synchronous
|
||||
:class:`~nextcloud_mcp_server.embedding.gateway_client.GatewayProvider` /
|
||||
``_GatewayOcrBackend`` — same M2M :class:`GatewayTokenProvider` bearer, no
|
||||
provider keys in the pod.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import base64
|
||||
import logging
|
||||
from dataclasses import dataclass
|
||||
from typing import Any
|
||||
|
||||
import httpx
|
||||
|
||||
from .gateway_client import GatewayTokenProvider
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Connect timeout for the (cheap) submit/poll calls. These are control-plane-ish
|
||||
# requests — a submit returns immediately with a job id and a poll is a status
|
||||
# read — so they get a short, fixed timeout, NOT the document-OCR read timeout
|
||||
# (which sizes a synchronous transcription).
|
||||
_BATCH_CONNECT_TIMEOUT_SECONDS = 5.0
|
||||
_BATCH_REQUEST_TIMEOUT_SECONDS = 30.0
|
||||
|
||||
# Gateway-normalised batch lifecycle (OcrBatchStatus on the gateway side).
|
||||
_PENDING = "pending"
|
||||
_SUCCEEDED = "succeeded"
|
||||
_FAILED = "failed"
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class BatchPollResult:
|
||||
"""One poll of a batch OCR job.
|
||||
|
||||
``status`` is the gateway-normalised lifecycle (``pending`` | ``succeeded`` |
|
||||
``failed``). For a single-document job: on ``succeeded`` ``pages`` holds the
|
||||
document's per-page ``(index, markdown)`` (empty + ``error`` set if that one
|
||||
document errored inside an otherwise-successful job); on ``failed`` ``error``
|
||||
carries the job-level failure.
|
||||
"""
|
||||
|
||||
status: str
|
||||
pages: list[tuple[int, str]]
|
||||
error: str | None = None
|
||||
|
||||
@property
|
||||
def is_pending(self) -> bool:
|
||||
return self.status == _PENDING
|
||||
|
||||
@property
|
||||
def is_succeeded(self) -> bool:
|
||||
return self.status == _SUCCEEDED
|
||||
|
||||
@property
|
||||
def is_failed(self) -> bool:
|
||||
return self.status == _FAILED
|
||||
|
||||
|
||||
class GatewayBatchOcrClient:
|
||||
"""Submits + polls single-document batch OCR jobs against the gateway."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
base_url: str,
|
||||
model: str,
|
||||
token_provider: GatewayTokenProvider | None = None,
|
||||
) -> None:
|
||||
# EMBEDDING_GATEWAY_URL is a bare origin; the batch routes live under /v1
|
||||
# like the rest of the gateway API. Idempotent if already /v1-suffixed.
|
||||
base = base_url.rstrip("/")
|
||||
if not base.endswith("/v1"):
|
||||
base = f"{base}/v1"
|
||||
self._base = base
|
||||
self._model = model
|
||||
self._token_provider = token_provider
|
||||
|
||||
async def _headers(self) -> dict[str, str]:
|
||||
if self._token_provider is None:
|
||||
return {}
|
||||
return {"Authorization": f"Bearer {await self._token_provider.get_token()}"}
|
||||
|
||||
async def submit(self, content: bytes, mime_type: str, custom_id: str) -> str:
|
||||
"""Submit ``content`` as a one-document batch job; return the namespaced
|
||||
``job_id`` to persist + poll. Raises on transport / non-2xx."""
|
||||
payload = {
|
||||
"model": self._model,
|
||||
"documents": [
|
||||
{
|
||||
"custom_id": custom_id,
|
||||
"mime_type": mime_type,
|
||||
"document_b64": base64.b64encode(content).decode("ascii"),
|
||||
}
|
||||
],
|
||||
}
|
||||
async with httpx.AsyncClient(
|
||||
timeout=httpx.Timeout(
|
||||
_BATCH_REQUEST_TIMEOUT_SECONDS, connect=_BATCH_CONNECT_TIMEOUT_SECONDS
|
||||
)
|
||||
) as client:
|
||||
resp = await client.post(
|
||||
f"{self._base}/ocr/batch", json=payload, headers=await self._headers()
|
||||
)
|
||||
resp.raise_for_status()
|
||||
body = resp.json()
|
||||
job_id = body["job_id"]
|
||||
logger.info(
|
||||
"batch OCR submitted: job_id=%s custom_id=%s status=%s",
|
||||
job_id,
|
||||
custom_id,
|
||||
body.get("status"),
|
||||
)
|
||||
return job_id
|
||||
|
||||
async def poll(self, job_id: str) -> BatchPollResult:
|
||||
"""Poll a batch job. Raises on transport / non-2xx; maps a terminal job's
|
||||
single-document result into :class:`BatchPollResult`."""
|
||||
async with httpx.AsyncClient(
|
||||
timeout=httpx.Timeout(
|
||||
_BATCH_REQUEST_TIMEOUT_SECONDS, connect=_BATCH_CONNECT_TIMEOUT_SECONDS
|
||||
)
|
||||
) as client:
|
||||
resp = await client.get(
|
||||
f"{self._base}/ocr/batch/{job_id}", headers=await self._headers()
|
||||
)
|
||||
resp.raise_for_status()
|
||||
body = resp.json()
|
||||
status = body.get("status", _PENDING)
|
||||
if status != _SUCCEEDED:
|
||||
# pending: nothing to read yet. failed: surface the job-level error.
|
||||
return BatchPollResult(status=status, pages=[], error=body.get("error"))
|
||||
return _result_from_success(body)
|
||||
|
||||
|
||||
def _result_from_success(body: dict[str, Any]) -> BatchPollResult:
|
||||
"""Extract the single document's pages from a succeeded job's results.
|
||||
|
||||
Submitting one document per job means exactly one result item; defensively
|
||||
take the first. A per-document error inside a succeeded job (the document
|
||||
failed but the job didn't) surfaces as a failed poll so the caller marks the
|
||||
doc parse-failed rather than indexing empty text.
|
||||
"""
|
||||
results = body.get("results") or []
|
||||
if not results:
|
||||
return BatchPollResult(
|
||||
status=_FAILED,
|
||||
pages=[],
|
||||
error="batch job succeeded but returned no results",
|
||||
)
|
||||
item = results[0]
|
||||
if item.get("error") is not None or item.get("pages") is None:
|
||||
return BatchPollResult(status=_FAILED, pages=[], error=item.get("error"))
|
||||
pages = [(p["index"], p.get("markdown", "")) for p in item["pages"]]
|
||||
return BatchPollResult(status=_SUCCEEDED, pages=pages)
|
||||
@@ -0,0 +1,121 @@
|
||||
"""Tracking store for in-flight async batch OCR jobs (Deck #332).
|
||||
|
||||
When ``DOCUMENT_OCR_MODE=batch`` the OCR tier submits a document to the gateway's
|
||||
async batch route and then re-polls across procrastinate retries. procrastinate
|
||||
job args are immutable, so the gateway ``job_id`` (and submit time, for the poll
|
||||
deadline) live in the ``batch_ocr_jobs`` app-DB table, keyed on the document +
|
||||
its content version (``etag``).
|
||||
|
||||
Engine reuse mirrors :class:`~nextcloud_mcp_server.usage.store.UsageEventStore`:
|
||||
rather than open its own engine this store borrows the process-wide
|
||||
:class:`RefreshTokenStorage` singleton (``get_shared_storage()``) — same app DB,
|
||||
dialect handling, ``?``-placeholder shim, and the guarantee that Alembic
|
||||
migrations (incl. ``batch_ocr_jobs``) already ran.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
import time
|
||||
from dataclasses import dataclass
|
||||
|
||||
import anyio
|
||||
|
||||
from nextcloud_mcp_server.auth.storage import RefreshTokenStorage, get_shared_storage
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class BatchOcrJob:
|
||||
"""A tracked in-flight batch OCR job."""
|
||||
|
||||
job_id: str
|
||||
status: str
|
||||
submitted_at: int
|
||||
|
||||
|
||||
class BatchOcrJobStore:
|
||||
"""CRUD for the ``batch_ocr_jobs`` table (one row per in-flight job)."""
|
||||
|
||||
_shared_instance: "BatchOcrJobStore | None" = None
|
||||
_shared_lock: anyio.Lock = anyio.Lock()
|
||||
|
||||
def __init__(self, storage: RefreshTokenStorage) -> None:
|
||||
self._storage = storage
|
||||
|
||||
@classmethod
|
||||
async def shared(cls) -> "BatchOcrJobStore":
|
||||
"""Process-wide store backed by the storage singleton (mirrors
|
||||
``UsageEventStore.shared``). Tests should construct
|
||||
``BatchOcrJobStore(storage)`` directly — the cache is a process global
|
||||
with no teardown hook."""
|
||||
async with cls._shared_lock:
|
||||
if cls._shared_instance is None:
|
||||
cls._shared_instance = cls(await get_shared_storage())
|
||||
return cls._shared_instance
|
||||
|
||||
async def get(
|
||||
self, *, user_id: str, doc_id: str, doc_type: str, etag: str
|
||||
) -> BatchOcrJob | None:
|
||||
"""The in-flight job for this document+version, or ``None``."""
|
||||
async with self._storage.acquire() as db:
|
||||
async with db.execute(
|
||||
"SELECT job_id, status, submitted_at FROM batch_ocr_jobs "
|
||||
"WHERE user_id = ? AND doc_id = ? AND doc_type = ? AND etag = ?",
|
||||
(user_id, doc_id, doc_type, etag),
|
||||
) as cursor:
|
||||
row = await cursor.fetchone()
|
||||
if row is None:
|
||||
return None
|
||||
return BatchOcrJob(job_id=row[0], status=row[1], submitted_at=int(row[2]))
|
||||
|
||||
async def insert_pending(
|
||||
self,
|
||||
*,
|
||||
user_id: str,
|
||||
doc_id: str,
|
||||
doc_type: str,
|
||||
etag: str,
|
||||
job_id: str,
|
||||
submitted_at: int | None = None,
|
||||
) -> None:
|
||||
"""Record a freshly-submitted job. ``ON CONFLICT DO NOTHING`` makes a
|
||||
racing double-submit harmless (the first row wins; the loser's job id is
|
||||
abandoned and reaped by the gateway-side file purge)."""
|
||||
now = submitted_at if submitted_at is not None else int(time.time())
|
||||
async with self._storage.acquire() as db:
|
||||
await db.execute(
|
||||
"INSERT INTO batch_ocr_jobs "
|
||||
"(user_id, doc_id, doc_type, etag, job_id, status, submitted_at, updated_at) "
|
||||
"VALUES (?, ?, ?, ?, ?, ?, ?, ?) "
|
||||
"ON CONFLICT (user_id, doc_id, doc_type, etag) DO NOTHING",
|
||||
(user_id, doc_id, doc_type, etag, job_id, "pending", now, now),
|
||||
)
|
||||
await db.commit()
|
||||
|
||||
async def delete(
|
||||
self, *, user_id: str, doc_id: str, doc_type: str, etag: str
|
||||
) -> None:
|
||||
"""Drop the row once the job is terminal (succeeded or failed)."""
|
||||
async with self._storage.acquire() as db:
|
||||
await db.execute(
|
||||
"DELETE FROM batch_ocr_jobs "
|
||||
"WHERE user_id = ? AND doc_id = ? AND doc_type = ? AND etag = ?",
|
||||
(user_id, doc_id, doc_type, etag),
|
||||
)
|
||||
await db.commit()
|
||||
|
||||
async def delete_stale_for_doc(
|
||||
self, *, user_id: str, doc_id: str, doc_type: str, keep_etag: str
|
||||
) -> None:
|
||||
"""Remove superseded-version rows for a document (any etag other than the
|
||||
current one) before a resubmit, so a re-edited file doesn't leave its
|
||||
old in-flight job tracked forever."""
|
||||
async with self._storage.acquire() as db:
|
||||
await db.execute(
|
||||
"DELETE FROM batch_ocr_jobs "
|
||||
"WHERE user_id = ? AND doc_id = ? AND doc_type = ? AND etag != ?",
|
||||
(user_id, doc_id, doc_type, keep_etag),
|
||||
)
|
||||
await db.commit()
|
||||
@@ -126,6 +126,7 @@ async def _parse_pdf_tier(
|
||||
filename: str | None,
|
||||
tier: str,
|
||||
settings: Any,
|
||||
options: dict[str, Any] | None = None,
|
||||
) -> "ProcessingResult":
|
||||
"""Run a single extraction tier and apply the post-parse escalation gate.
|
||||
|
||||
@@ -141,18 +142,31 @@ async def _parse_pdf_tier(
|
||||
the "OCR is an enhancement, never worse than off" invariant: a tenant who has
|
||||
not enabled a higher tier (or has no processor for it) simply indexes the
|
||||
cheap tier's output.
|
||||
|
||||
Batch OCR (Deck #332): when the OCR tier's batch job is still in flight the
|
||||
processor returns a *pending sentinel* result; we translate it here into a
|
||||
``BatchPending`` raise (same decision point as ``EscalateError``) so the
|
||||
retry strategy re-runs this tier after a delay instead of indexing empty text.
|
||||
"""
|
||||
# Lazy import: keep the document stack (pymupdf/_isolation) off the module
|
||||
# load path; this runs only on the per-tier worker, which needs it anyway.
|
||||
from nextcloud_mcp_server.document_processors.escalation import ( # noqa: PLC0415
|
||||
BatchPending,
|
||||
EscalateError,
|
||||
)
|
||||
from nextcloud_mcp_server.document_processors.ocr import ( # noqa: PLC0415
|
||||
OCR_BATCH_PENDING_KEY,
|
||||
OCR_BATCH_RETRY_IN_KEY,
|
||||
)
|
||||
|
||||
# options / progress_callback are not threaded here -- the indexing caller
|
||||
# passes neither today, and the inline path (registry.process) omits them
|
||||
# too. Forward them if a tier processor ever needs per-call tuning (e.g. OCR
|
||||
# DPI); keeping the two paths symmetric until then.
|
||||
result = await registry.process_tier(content, content_type, filename, tier)
|
||||
# ``options`` threads per-document identity (user_id/doc_id/doc_type/etag) to
|
||||
# the OCR tier so batch mode can key its job-tracking table (Deck #332). Other
|
||||
# tiers ignore it. The inline path (registry.process) passes None.
|
||||
result = await registry.process_tier(
|
||||
content, content_type, filename, tier, options=options
|
||||
)
|
||||
if result.metadata.get(OCR_BATCH_PENDING_KEY):
|
||||
raise BatchPending(retry_in=int(result.metadata[OCR_BATCH_RETRY_IN_KEY]))
|
||||
if result.success:
|
||||
decision = registry.evaluate_escalation(
|
||||
result, content, tier, settings, filename=filename
|
||||
@@ -525,19 +539,21 @@ async def process_document(
|
||||
re-picked on the next scan; the procrastinate path (max_retries=1) caps it at
|
||||
one outer attempt (~30s) and defers. Don't stack a third retry layer here.
|
||||
"""
|
||||
# EscalateError is a control-flow signal that arises ONLY on the per-tier
|
||||
# external path (tier set). Bind the class lazily here, and only when a tier
|
||||
# is set, so the document stack is never imported at *module load* (the #877
|
||||
# invariant) nor on the delete / text-doc call paths (file processing already
|
||||
# imports it via get_registry regardless). When tier is None it can't be
|
||||
# raised, so the guards below stay inert.
|
||||
escalate_error_cls: type[BaseException] | None = None
|
||||
# EscalateError and BatchPending are control-flow signals that arise ONLY on
|
||||
# the per-tier external path (tier set). Bind them lazily here, and only when a
|
||||
# tier is set, so the document stack is never imported at *module load* (the
|
||||
# #877 invariant) nor on the delete / text-doc call paths (file processing
|
||||
# already imports them via get_registry regardless). When tier is None neither
|
||||
# can be raised, so the guards below stay inert. Bound as a tuple so the guards
|
||||
# treat both identically: propagate untouched, never record an error/drop.
|
||||
control_flow_excs: tuple[type[BaseException], ...] = ()
|
||||
if tier is not None:
|
||||
from nextcloud_mcp_server.document_processors.escalation import ( # noqa: PLC0415
|
||||
BatchPending,
|
||||
EscalateError,
|
||||
)
|
||||
|
||||
escalate_error_cls = EscalateError
|
||||
control_flow_excs = (EscalateError, BatchPending)
|
||||
|
||||
start_time = time.time()
|
||||
|
||||
@@ -632,13 +648,11 @@ async def process_document(
|
||||
return # Success
|
||||
|
||||
except Exception as e:
|
||||
# An escalation signal is control flow, not a failure:
|
||||
# propagate it untouched so the procrastinate retry strategy
|
||||
# can hop the job to the next tier's queue. Never retry it
|
||||
# A control-flow signal (escalation hop, or batch-OCR re-poll
|
||||
# deferral) is not a failure: propagate it untouched so the
|
||||
# procrastinate retry strategy handles it. Never retry it
|
||||
# in-process and never count it as a drop.
|
||||
if escalate_error_cls is not None and isinstance(
|
||||
e, escalate_error_cls
|
||||
):
|
||||
if isinstance(e, control_flow_excs):
|
||||
raise
|
||||
if attempt < max_retries - 1:
|
||||
logger.warning(
|
||||
@@ -691,10 +705,11 @@ async def process_document(
|
||||
raise
|
||||
|
||||
except Exception as e:
|
||||
# An escalation signal must reach the procrastinate retry strategy
|
||||
# un-recorded -- it is neither a processing success nor an error
|
||||
# (the hop is its own event, counted via record_document_escalation).
|
||||
if escalate_error_cls is not None and isinstance(e, escalate_error_cls):
|
||||
# A control-flow signal must reach the procrastinate retry strategy
|
||||
# un-recorded -- it is neither a processing success nor an error (an
|
||||
# escalation hop is counted via record_document_escalation; a batch
|
||||
# re-poll deferral is not an event at all).
|
||||
if isinstance(e, control_flow_excs):
|
||||
raise
|
||||
# Single processing-error call site: catches exhausted-retry
|
||||
# re-raises, delete failures, and setup errors (get_qdrant_client /
|
||||
@@ -956,6 +971,7 @@ async def _index_document(
|
||||
get_registry,
|
||||
)
|
||||
from nextcloud_mcp_server.document_processors.escalation import ( # noqa: PLC0415
|
||||
BatchPending,
|
||||
EscalateError,
|
||||
)
|
||||
|
||||
@@ -968,6 +984,15 @@ async def _index_document(
|
||||
# and the in-process/memory pool (tier is None) -- runs the inline
|
||||
# tiered pipeline (fast -> OCR escalation in one call).
|
||||
if tier is not None and _is_pdf(content_type):
|
||||
# Thread per-document identity to the OCR tier so batch mode
|
||||
# (Deck #332) can key its job-tracking table; other tiers
|
||||
# ignore it.
|
||||
ocr_options = {
|
||||
"user_id": doc_task.user_id,
|
||||
"doc_id": doc_task.doc_id,
|
||||
"doc_type": doc_task.doc_type,
|
||||
"etag": doc_task.etag or "",
|
||||
}
|
||||
result = await _parse_pdf_tier(
|
||||
registry,
|
||||
content_bytes,
|
||||
@@ -975,6 +1000,7 @@ async def _index_document(
|
||||
file_path,
|
||||
tier,
|
||||
settings,
|
||||
options=ocr_options,
|
||||
)
|
||||
else:
|
||||
result = await registry.process(
|
||||
@@ -1047,10 +1073,12 @@ async def _index_document(
|
||||
)
|
||||
else:
|
||||
logger.debug("No page_boundaries in metadata for %s", file_path)
|
||||
except EscalateError:
|
||||
# Control-flow signal (per-tier path): re-raise untouched so the
|
||||
# queue hops the job to the next tier. NOT a "failed to process"
|
||||
# error -- don't log it as one.
|
||||
except (EscalateError, BatchPending):
|
||||
# Control-flow signals (per-tier path): re-raise untouched.
|
||||
# EscalateError hops the job to the next tier; BatchPending defers
|
||||
# a re-poll on the same tier (batch OCR still in flight, Deck
|
||||
# #332). Neither is a "failed to process" error -- don't log them
|
||||
# as one.
|
||||
raise
|
||||
except Exception as e:
|
||||
logger.error("Failed to process file %s: %s", file_path, e)
|
||||
|
||||
@@ -328,13 +328,26 @@ class TieredEscalationStrategy(BaseRetryStrategy):
|
||||
def get_retry_decision(
|
||||
self, *, exception: BaseException, job: Job
|
||||
) -> RetryDecision | None:
|
||||
# Lazy import: EscalateError lives in the document stack, which the API
|
||||
# pod (it also builds this App to defer) must not load. get_retry_decision
|
||||
# runs only in the worker, where the stack is already imported.
|
||||
from ...document_processors.escalation import EscalateError # noqa: PLC0415
|
||||
# Lazy import: these live in the document stack, which the API pod (it
|
||||
# also builds this App to defer) must not load. get_retry_decision runs
|
||||
# only in the worker, where the stack is already imported.
|
||||
from ...document_processors.escalation import ( # noqa: PLC0415
|
||||
BatchPending,
|
||||
EscalateError,
|
||||
)
|
||||
|
||||
exc = _first_leaf(exception)
|
||||
|
||||
if isinstance(exc, BatchPending):
|
||||
# Batch OCR job still in flight (Deck #332): defer a re-poll on the
|
||||
# SAME queue after retry_in seconds. Deliberately exempt from the
|
||||
# transient cap below — a batch job can take minutes-hours, so the
|
||||
# poll count is unbounded here; the OCR processor's own deadline
|
||||
# (DOCUMENT_OCR_BATCH_MAX_WAIT_SECONDS) is what terminates a stuck job.
|
||||
# Releasing the worker between polls keeps the job out of `doing`, so
|
||||
# it's never stall-reclaimed.
|
||||
return RetryDecision(retry_in={"seconds": exc.retry_in})
|
||||
|
||||
if isinstance(exc, EscalateError):
|
||||
queue = TIER_QUEUES.get(exc.to_tier)
|
||||
if queue is None:
|
||||
|
||||
Reference in New Issue
Block a user