Files
mcp-nextcloud/tests/unit/vector/test_tiered_escalation_strategy.py
T
Chris CoutinhoandClaude Opus 4.8 c21804fbbc feat(ingest): split OCR into tier2 in-cluster (GPU, gateway-only) + tier3 upstream
Insert a configurable in-cluster OCR rung into the escalation ladder (Deck #353):
a tier2-eligible doc is OCR'd on the on-demand burst GPU before falling through to
paid upstream OCR. The in-cluster backend is reached ONLY via the embedding gateway
(model prefix routes to the GPU over the tailnet) and is a config value (default
surya/surya-ocr-2, swappable to e.g. lightonocr) — never hard-coded.

Ladder: fast -> structured -> ocr-incluster -> ocr-upstream
(queues ingest-ocr-incluster / ingest-ocr-upstream).

- escalation.py: 4-tier ladder; in-cluster flag folded into the dead-letter signature.
- ocr.py: OcrProcessor(name, tier, model_setting, gateway_only); build_ocr_backend(
  ..., model=, gateway_only=) — gateway_only forces the gateway backend (never the
  direct Mistral fallback), disabling the tier with a warning if no gateway URL.
- registry.py: per-rung enable map; scanned docs target minimum="ocr-incluster";
  inline path runs the cheapest available OCR rung.
- procrastinate.py: two OCR queues; legacy ingest-ocr kept as a drain target.
- config.py: DOCUMENT_OCR_INCLUSTER_ENABLED (off) + DOCUMENT_OCR_INCLUSTER_MODEL.
- __init__.py: register the two OCR instances; vector/processor.py: pages_ocr
  metered for the upstream (paid) rung only; cli.py: new --tier choices + legacy drain.
- metrics.py: zero the legacy ingest-ocr queue gauge during rollout.
- tests: migrated to the split ladder + new tests (gateway-only forcing, per-tier
  model incl. lightonocr override, no-hard-coded-surya guard). 1792 pass; ruff + ty green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-17 23:28:29 +02:00

157 lines
6.4 KiB
Python

"""Unit tests for the per-tier escalation primitives (Deck #323).
Covers the tier-ladder helpers + EscalateError (document_processors.escalation)
and the procrastinate TieredEscalationStrategy that turns a raised exception
into a queue-hop / same-tier retry / give-up decision.
"""
from datetime import datetime, timezone
import httpx
import pytest
from procrastinate.jobs import Job
import nextcloud_mcp_server.vector.queue.procrastinate as pq
from nextcloud_mcp_server.document_processors.escalation import (
TIER_LADDER,
BatchPending,
EscalateError,
next_tier,
)
pytestmark = pytest.mark.unit
def _job(queue: str = pq.INGEST_QUEUE_FAST, attempts: int = 1) -> Job:
return Job(
id=1,
queue=queue,
task_name=pq.INGEST_TASK_NAME,
lock=None,
queueing_lock=None,
attempts=attempts,
)
class TestLadder:
def test_next_tier_ordering(self):
assert next_tier("fast") == "structured"
assert next_tier("structured") == "ocr-incluster"
assert next_tier("ocr-incluster") == "ocr-upstream"
assert next_tier("ocr-upstream") is None # terminal
assert next_tier("unknown") is None
def test_ladder_is_cheapest_first(self):
assert TIER_LADDER == ("fast", "structured", "ocr-incluster", "ocr-upstream")
def test_tier_for_queue(self):
assert pq.tier_for_queue(pq.INGEST_QUEUE_OCR_UPSTREAM) == "ocr-upstream"
assert pq.tier_for_queue(pq.INGEST_QUEUE_STRUCTURED) == "structured"
# Legacy / unknown / None all fall back to the cheapest tier.
assert pq.tier_for_queue(pq.LEGACY_INGEST_QUEUE) == "fast"
assert pq.tier_for_queue(None) == "fast"
class TestTieredEscalationStrategy:
def _strategy(self, max_transient: int = 5):
return pq.TieredEscalationStrategy(max_transient_attempts=max_transient)
def test_escalate_hops_to_target_queue(self):
exc = EscalateError(
from_tier="fast", to_tier="ocr-upstream", reason="empty_text"
)
decision = self._strategy().get_retry_decision(exception=exc, job=_job())
assert decision is not None
assert decision.queue == pq.INGEST_QUEUE_OCR_UPSTREAM
def test_escalate_to_structured(self):
exc = EscalateError(
from_tier="fast", to_tier="structured", reason="low_confidence"
)
decision = self._strategy().get_retry_decision(exception=exc, job=_job())
assert decision is not None
assert decision.queue == pq.INGEST_QUEUE_STRUCTURED
def test_escalate_unknown_tier_gives_up(self):
exc = EscalateError(from_tier="ocr", to_tier="bogus", reason="low_confidence")
decision = self._strategy().get_retry_decision(exception=exc, job=_job())
assert decision is None
def test_escalate_unwraps_exception_group(self):
exc = EscalateError(
from_tier="fast", to_tier="ocr-upstream", reason="empty_text"
)
group = ExceptionGroup("wrapped", [exc])
decision = self._strategy().get_retry_decision(exception=group, job=_job())
assert decision is not None
assert decision.queue == pq.INGEST_QUEUE_OCR_UPSTREAM
def test_transient_retries_same_queue_under_cap(self):
decision = self._strategy(max_transient=5).get_retry_decision(
exception=httpx.ConnectError("refused"), job=_job(attempts=1)
)
assert decision is not None
# Same-tier retry: no queue override (stays on its current queue).
assert decision.queue is None
assert decision.retry_at is not None
def test_transient_backoff_progression(self):
# min(4 * 2**(attempts-1), 300): 4, 8, 16, ... capped at 300s.
# procrastinate sets retry_at = utcnow() + wait at call time. Bracketing
# the call with before/after makes the assertion exact and independent of
# runner load: with before <= call_now <= after, we have
# (retry_at - after) <= wait <= (retry_at - before).
strat = self._strategy(max_transient=100)
for attempts, expected in [(1, 4), (2, 8), (3, 16), (4, 32), (20, 300)]:
before = datetime.now(timezone.utc)
decision = strat.get_retry_decision(
exception=httpx.ConnectError("x"), job=_job(attempts=attempts)
)
after = datetime.now(timezone.utc)
assert decision is not None and decision.retry_at is not None
lo = (decision.retry_at - after).total_seconds()
hi = (decision.retry_at - before).total_seconds()
assert lo <= expected <= hi, (
f"attempts={attempts}: expected={expected}s not in [{lo:.3f}, {hi:.3f}]"
)
def test_transient_gives_up_over_cap(self):
decision = self._strategy(max_transient=5).get_retry_decision(
exception=httpx.ConnectError("refused"), job=_job(attempts=5)
)
assert decision is None
def test_non_transient_error_gives_up(self):
decision = self._strategy().get_retry_decision(
exception=ValueError("permanent"), job=_job(attempts=1)
)
assert decision is None
def test_batch_pending_defers_same_queue(self):
# Batch OCR re-poll (Deck #332): same-queue deferral after retry_in.
before = datetime.now(timezone.utc)
decision = self._strategy().get_retry_decision(
exception=BatchPending(retry_in=120),
job=_job(queue=pq.INGEST_QUEUE_OCR_UPSTREAM),
)
after = datetime.now(timezone.utc)
assert decision is not None
assert decision.queue is None # stays on its own tier queue
assert decision.retry_at is not None
lo = (decision.retry_at - after).total_seconds()
hi = (decision.retry_at - before).total_seconds()
assert lo <= 120 <= hi
def test_batch_pending_exempt_from_transient_cap(self):
# A batch can take hours -> many polls; the transient cap must NOT stop it
# (the OCR processor's own deadline terminates a stuck job instead).
decision = self._strategy(max_transient=5).get_retry_decision(
exception=BatchPending(retry_in=60), job=_job(attempts=999)
)
assert decision is not None and decision.retry_at is not None
def test_batch_pending_unwraps_exception_group(self):
group = ExceptionGroup("wrapped", [BatchPending(retry_in=60)])
decision = self._strategy().get_retry_decision(exception=group, job=_job())
assert decision is not None and decision.retry_at is not None