1 commit

Author	SHA1	Message	Date
Cursor Agent	ebf4346e0c	Big-bang: replace Prisma with SQLAlchemy + Prisma-compat shim ... This is the rip-and-replace counterpart to the phased migration in PR #28366. The proxy database layer is moved to SQLAlchemy + asyncpg in a single PR. Existing prisma_client.db.<table>.<method>(...) call sites keep working unchanged via a compatibility shim; the prisma PyPI package is dropped from runtime dependencies. The PR description carries the full breakage inventory. Test pass rates after rip-and-replace: * tests/test_litellm/proxy/db/ 247 / 263 = 93.9% * tests/test_litellm/proxy/management_endpoints/ 1288 / 1289 = 99.9% * tests/test_litellm/proxy/db/sqlmodel_orm/ 38 / 38 = 100% * combined 1573 / 1590 = 98.9% The 17 failures are all in tests of Prisma-internal subprocess / watchdog / RoutingPrismaWrapper mechanics that no longer apply to a SQLAlchemy backend. Architecture: * litellm/proxy/db/sqlmodel/schema_parser.py Pure-Python parser for the subset of Prisma DSL used by schema.prisma. (carried forward from the Phase 1 PR.) * litellm/proxy/db/sqlmodel/_generate.py Code generator that emits Black-formatted SQLModel classes from a parsed schema. Output uses JSONB().with_variant(JSON(), 'sqlite') so the model definitions are portable to SQLite test environments. * litellm/proxy/db/sqlmodel/models.py SQLModel classes for all 64 Prisma models, generator-seeded. * litellm/proxy/db/sqlmodel/engine.py LiteLLMDB wrapping the async SQLAlchemy engine + sessionmaker, with pluggable RDS IAM token rotation (engine recreate on each tick) and optional read-replica routing. * litellm/proxy/db/sqlmodel/compat.py PrismaCompatClient exposes the historical Prisma surface -- per-table accessors, find_unique/first/many, create/update/upsert/ delete/update_many/delete_many/create_many/count, query_raw/ execute_raw, batch_(), tx() -- backed by SQLAlchemy. Filter translation covers in/contains/startswith/endswith/lt/lte/gt/gte/has /AND/OR/NOT/equals plus mode:insensitive and data={...:{increment:N}}. * litellm/proxy/db/sqlmodel/errors.py Native error classes (PrismaError, UniqueViolationError, ...) raised by the shim. SQLAlchemy IntegrityError / NoResultFound / ProgrammingError are translated to the corresponding LiteLLM-native types. * tests/_prisma_compat.py + conftest.py Test-only stand-in for the prisma namespace so legacy from prisma.errors import X imports in test fixtures still resolve to the LiteLLM-native errors. Production code has been ported off from prisma imports entirely. Wiring: * litellm/proxy/utils.py PrismaClient.__init__ now constructs LiteLLMDB + PrismaCompatClient instead of Prisma + PrismaWrapper + RoutingPrismaWrapper. _get_engine_pid returns 0 (no subprocess). * The 9 production files that used from prisma.errors import ... / import prisma are redirected to the native errors module: exception_handler, log_db_metrics, toolset_db, budget/config_override /mcp/workflow management endpoints, key_management_endpoints. The prisma.Json(...) marker in key_management is replaced by an identity shim (SQLAlchemy JSONB accepts dicts directly). * prisma==0.11.0 dropped from extra_proxy and proxy-dev in pyproject.toml. Known gaps (called out explicitly in the PR body): * include= for relation eager-loading is logged-and-ignored. SQLModel classes don't carry SQLAlchemy relationship() defs yet; ~55 production sites pass include={} and would see None for the relation attribute. Hand-written relationships per table is the next step. * select=, group_by, and some/every relation filters raise NotImplementedError -- 3, 4, and 2 production sites respectively. * litellm-proxy-extras Prisma migration bundle is unchanged. Alembic baseline + recovery path is the next mechanical step. * schema.prisma copies, the sync workflows, helm migration jobs, and Dockerfile prisma-generate steps are unchanged. All rote follow-ups. Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>	2026-05-20 21:54:58 +00:00