feat(submissions): create durable queued evaluations idempotently

This commit is contained in:
2026-10-09 06:31:15 +08:00
parent 8a01081cc9
commit 75d74908e0
9 changed files with 675 additions and 0 deletions
+7
View File
@@ -10,6 +10,7 @@ from kf_oj.api.body_limits import BodyLimitMiddleware
from kf_oj.api.v1.identity import router as identity_router
from kf_oj.api.v1.problems import router as problems_router
from kf_oj.api.v1.router import router
from kf_oj.api.v1.submissions import router as submission_router
from kf_oj.api.v1.worker_artifacts import router as worker_router
from kf_oj.core.logging import configure_logging
from kf_oj.core.settings import Settings
@@ -18,6 +19,7 @@ from kf_oj.infra.artifacts.store import ArtifactStore
from kf_oj.infra.db.session import build_engine, build_session_factory
from kf_oj.modules.identity.service import IdentityService
from kf_oj.modules.problems.service import ProblemService
from kf_oj.modules.submissions.service import SubmissionService
from kf_oj.modules.submissions.worker_access import WorkerAccess
@@ -50,6 +52,9 @@ def create_app(settings: Settings | None = None) -> FastAPI:
app.state.sessions, artifacts, settings.judge_admission
)
if artifacts is not None:
app.state.submissions = SubmissionService(
app.state.sessions, app.state.problems, artifacts, settings.judge_admission
)
app.state.worker_access = WorkerAccess(
app.state.sessions, app.state.problems, artifacts
)
@@ -58,6 +63,7 @@ def create_app(settings: Settings | None = None) -> FastAPI:
app.state.identity = None
app.state.problems = None
app.state.worker_access = None
app.state.submissions = None
if engine is not None:
await engine.dispose()
if store is not None:
@@ -84,4 +90,5 @@ def create_app(settings: Settings | None = None) -> FastAPI:
app.include_router(identity_router, prefix="/api/v1")
app.include_router(problems_router, prefix="/api/v1")
app.include_router(worker_router, prefix="/api/v1")
app.include_router(submission_router, prefix="/api/v1")
return app
+3
View File
@@ -14,6 +14,9 @@ class BodyLimitMiddleware:
await self.app(scope, receive, send)
return
limit = 8192 if scope["path"] == "/api/v1/auth/login" else 1024 * 1024
if scope["path"] == "/api/v1/submissions":
# Max256KiB UTF-8 source, worst-case JSON escaping plus bounded metadata.
limit = 6 * 256 * 1024 + 8192
headers = dict(scope.get("headers", []))
length = headers.get(b"content-length")
if length is not None:
+43
View File
@@ -0,0 +1,43 @@
from typing import Annotated
from fastapi import APIRouter, Depends, HTTPException, Request, Response
from kf_oj.api.dependencies import unsafe_principal
from kf_oj.contracts.submissions import CreateSubmission, SubmissionCreated
from kf_oj.infra.artifacts.store import ArtifactError
from kf_oj.modules.identity.public import Principal
from kf_oj.modules.submissions.public import (
IdempotencyConflict,
SubmissionBusy,
SubmissionUnavailable,
)
router = APIRouter(prefix="/submissions", tags=["submissions"])
@router.post("", status_code=201, response_model=SubmissionCreated)
async def create(
body: CreateSubmission,
request: Request,
response: Response,
principal: Annotated[Principal, Depends(unsafe_principal)],
):
if request.headers.get("content-type", "").split(";", 1)[0] != "application/json":
raise HTTPException(415, "JSON request required")
service = getattr(request.app.state, "submissions", None)
if service is None:
raise HTTPException(503, "submission service unavailable")
try:
result = await service.create(principal, body)
except IdempotencyConflict:
raise HTTPException(409, "idempotency request conflict") from None
except SubmissionBusy:
raise HTTPException(429, "pending submission capacity exhausted") from None
except SubmissionUnavailable:
raise HTTPException(404, "submission capability unavailable") from None
except ArtifactError:
raise HTTPException(503, "source storage unavailable") from None
except ValueError:
raise HTTPException(422, "submission outside admission policy") from None
response.headers["Cache-Control"] = "no-store"
return result
@@ -0,0 +1,37 @@
from datetime import datetime
from typing import Annotated, Literal
from pydantic import Field, model_validator
from kf_oj.contracts.judging import MAX_SOURCE_BYTES, Identifier, Language, WireModel
EvaluationState = Literal[
"queued", "claimed", "compiling", "running", "cancelling", "finished", "cancelled", "failed"
]
class CreateSubmission(WireModel):
problem_version_id: Identifier
language: Language
idempotency_key: Identifier
source: Annotated[
str, Field(strict=True, min_length=1, max_length=MAX_SOURCE_BYTES, repr=False)
]
@model_validator(mode="after")
def bounded_source(self):
try:
if len(self.source.encode("utf-8")) > MAX_SOURCE_BYTES:
raise ValueError("source exceeds byte policy")
except UnicodeError:
raise ValueError("source must be UTF-8") from None
return self
class SubmissionCreated(WireModel):
submission_id: Identifier
evaluation_id: Identifier
problem_version_id: Identifier
language: Language
state: EvaluationState
created_at: datetime
@@ -8,3 +8,11 @@ class WorkerAccessDenied(ValueError):
class WorkerReadBusy(ValueError):
pass
class IdempotencyConflict(ValueError):
pass
class SubmissionBusy(ValueError):
pass
@@ -0,0 +1,135 @@
"""Only initial queued evaluation/outbox writes; execution belongs to Coordinator."""
import hashlib
import json
from uuid import uuid4
from sqlalchemy import func, select, text
from kf_oj.contracts.judging import JobManifest
from kf_oj.contracts.submissions import CreateSubmission, SubmissionCreated
from kf_oj.infra.artifacts.repository import ArtifactRepository
from kf_oj.infra.messaging.models import Outbox
from kf_oj.modules.identity.public import Principal
from kf_oj.modules.problems.public import ProblemPort, ProblemUnavailable
from kf_oj.modules.submissions.models import Evaluation, Submission
from kf_oj.modules.submissions.public import (
IdempotencyConflict,
SubmissionBusy,
SubmissionUnavailable,
)
class SubmissionService:
def __init__(self, sessions, problems: ProblemPort, artifacts: ArtifactRepository, admission):
self.sessions, self.problems = sessions, problems
self.artifacts, self.admission = artifacts, admission
@staticmethod
def _response(submission, evaluation):
return SubmissionCreated(
submission_id=submission.id,
evaluation_id=evaluation.id,
problem_version_id=submission.problem_version_id,
language=submission.language,
state=evaluation.state,
created_at=submission.created_at,
)
async def create(self, principal: Principal, request: CreateSubmission) -> SubmissionCreated:
# Validate even trusted non-HTTP callers. No path or argv is accepted.
request = CreateSubmission.model_validate(request.model_dump())
source = request.source.encode("utf-8")
request_digest = hashlib.sha256(
json.dumps(
[
str(request.problem_version_id),
request.language.value,
hashlib.sha256(source).hexdigest(),
],
separators=(",", ":"),
).encode()
).hexdigest()
if self.admission is None:
raise SubmissionUnavailable("judge admission unavailable")
async with self.sessions.begin() as session:
# Bounded pending capacity and per-account idempotency serialize across API processes.
# Two-key namespace is separate from artifact/slug advisory locks.
await session.execute(text("SELECT pg_advisory_xact_lock(1262907470, 0)"))
existing = (
await session.execute(
select(Submission, Evaluation)
.join(Evaluation, Evaluation.submission_id == Submission.id)
.where(
Submission.account_id == principal.account_id,
Submission.idempotency_key == request.idempotency_key,
)
)
).one_or_none()
if existing is not None:
if existing[0].request_digest != request_digest:
raise IdempotencyConflict("idempotency key belongs to another request")
return self._response(*existing)
pending = ("queued", "claimed", "compiling", "running", "cancelling")
total = await session.scalar(
select(func.count()).select_from(Evaluation).where(Evaluation.state.in_(pending))
)
owned = await session.scalar(
select(func.count())
.select_from(Submission)
.join(Evaluation, Evaluation.submission_id == Submission.id)
.where(Submission.account_id == principal.account_id, Evaluation.state.in_(pending))
)
if total >= 64 or owned >= 8:
raise SubmissionBusy("pending submission capacity exhausted")
try:
version = await self.problems.for_submission(session, request.problem_version_id)
except ProblemUnavailable:
raise SubmissionUnavailable("problem version unavailable") from None
if request.language not in version.languages:
raise SubmissionUnavailable("language unavailable")
# Bytes fsync before metadata/ref transaction. Rollback may leave collectable bytes,
# but cannot leave an evaluation that refers to non-durable bytes.
reference = await self.artifacts.put(session, source)
submission_id, evaluation_id = uuid4(), uuid4()
manifest = JobManifest(
schema_version=1,
kind="judge.manifest",
evaluation_id=evaluation_id,
submission_id=submission_id,
problem_version_id=version.version_id,
language=request.language,
toolchain=version.toolchain,
source=reference,
cases=version.cases,
comparison="exact_bytes",
compile_limits=version.compile_limits,
run_limits=version.run_limits,
job_wall_time_us=version.job_wall_time_us,
)
self.admission.admit(manifest, 1)
submission = Submission(
id=submission_id,
account_id=principal.account_id,
problem_id=version.problem_id,
problem_version_id=version.version_id,
language=request.language.value,
source_digest=reference.digest,
idempotency_key=request.idempotency_key,
request_digest=request_digest,
)
session.add(submission)
await session.flush()
await session.execute(
Evaluation.__table__.insert().values(
id=evaluation_id,
submission_id=submission_id,
problem_version_id=version.version_id,
)
)
session.add(Outbox(id=uuid4(), evaluation_id=evaluation_id))
await session.flush()
evaluation = await session.get(Evaluation, evaluation_id)
response = self._response(submission, evaluation)
# The transaction commit must succeed before any success is returned.
return response