pyverge

Schema evolution and migrations for versioned data. Register model versions, define migrations between them, and converge payloads to a target schema at runtime. The engine is provider-agnostic; Pydantic and JSON Schema adapters ship today.

Installation

pip install pyverge                 # core library
pip install "pyverge[cli]"          # CLI
pip install "pyverge[fastmcp]"      # MCP tool server
pip install "pyverge[faststream]"   # messaging adapter

Development: uv sync --group dev.

Quick start

from typing import Literal

import semver
from pydantic import BaseModel

from pyverge import Manager
from pyverge.migration import MigrationSettings, PydanticModelAdapter

UserManager = Manager[semver.Version].configure(
    MigrationSettings(),
    PydanticModelAdapter(),
)


@UserManager.model()
class UserV1(BaseModel):
    kind: Literal["User"] = "User"
    version: Literal["1.0.0"] = "1.0.0"
    name: str
    email: str


@UserManager.model()
class UserV2(BaseModel):
    kind: Literal["User"] = "User"
    version: Literal["2.0.0"] = "2.0.0"
    name: str
    email: str
    age: int | None = None


@UserManager.migration("User", "1.0.0", "2.0.0")
def add_age(data: dict, context=None) -> dict:
    return {**data, "age": None}


migrated = UserManager().migrate(
    {"kind": "User", "version": "1.0.0", "name": "Alice", "email": "a@b.com"}
)
# {"kind": "User", "version": "2.0.0", "name": "Alice", "email": "a@b.com", "age": None}

Async and prepared runs

result = await manager.amigrate(payload)

run = manager.delegate(payload)
first = run.run(context={"tenant_id": "tenant_a"})
second = run.run(context={"tenant_id": "tenant_b"})

migrate/amigrate execute immediately. delegate and prepare return a reusable MigrationRun. Execution policy comes from settings:

MigrationSettings(execution_mode="async", parallel_workers=4)

Migration formats

CLI

pyverge init                                       # scaffold project
pyverge doctor --manager myapp.models:UserManager  # registry health
pyverge validate data.json --manager myapp.models:UserManager \
  --schema User --version 1.0.0
pyverge migrate data.json --manager myapp.models:UserManager --target latest
pyverge migrate data.json --manager myapp.models:UserManager --check
pyverge compat --manager myapp.models:UserManager --fail-on-breaking
pyverge manifest --manager myapp.models:UserManager > manifest.json
pyverge diff --manager myapp.models:UserManager --from 1.0.0 --to 2.0.0
pyverge export --manager myapp.models:UserManager
pyverge info --manager myapp.models:UserManager

Managers are module:object_path and may be an instance, configured subclass, or zero-arg factory. Runtime schemas require a factory that builds the manager from their source.

manifest is a CLI-only contract: it emits a JGF v2 graph envelope with node schema hashes, migration hashes, and a content hash. Lifespan/startup does not write it. CI commands (doctor, validate, migrate --check, compat) return non-zero on failure. Config auto-discovery is not wired yet.

See CLI Reference.

Features

• Versioned registry — semver or ISO date versions, decorator registration
• Convergent engine — graph-driven migration of nested versioned entries
• Declarative migrations — JSON Patch or Python callables; JSONata planned
• Runtime context — shared read-only bindings across every step
• Sync/async executors — sequential or level-parallel; prepared runs
• Reflection — reconstruct missing models or propose migrations from diffs
• Compatibility — CompatibilityReport, breaking-change checks, JSON Patch rendering
• Graph manifest — JGF v2 snapshot with stable hashes for CI divergence checks
• Adapters — FastMCP tools/prompts/resources and FastStream messages
• Hooks — before/after/error callbacks, async-aware, OpenTelemetry hook

Documentation

• Getting Started
• Managers · Concepts
• Migrations · Migration Formats
• Runtime & Execution · Execution Flow
• Target Policy · Compatibility
• CLI Reference · Telemetry & Hooks
• FastMCP · FastStream
• Testing · Benchmarking
• Usage Patterns · Scenarios
• Showcases

Plan

• Additional model adapters: dataclasses, attrs, marshmallow, MessagePack
• Real-source showcase drivers: MongoDB, Kafka, RabbitMQ, MQTT
• Generated API reference once the SDK surface stabilizes