Where do you want to store your MCMC draws? In memory? On disk? Or in a database running in a datacenter?

No matter where you want to put them, or which PPL generates them: McBackend takes care of your MCMC samples.

Quickstart

The mcbackend package consists of three parts:

Part 1: A schema for MCMC run & chain metadata

No matter which programming language your favorite PPL is written in, the ProtocolBuffers from McBackend can be used to generate code in languages like C++, C#, Python and many more to represent commonly used metadata about MCMC runs, chains and model variables.

The definitions in protobufs/meta.proto are designed to maximize compatibility with ArviZ objects, making it easy to transform MCMC draws stored according to the McBackend schema to xarray.DataTree objects for plotting & analysis.

Part 2: A storage backend interface

The draws and stats created by MCMC sampling algorithms at runtime need to be stored somewhere.

This "somewhere" is called the storage backend in PPLs/MCMC frameworks like PyMC or emcee.

Most storage backends must be initialized with metadata about the model variables so they can, for example, pre-allocated memory for the draws and stats they're about to receive. After then receiving thousands of draws and stats they must then provide methods by which the draws/stats can be retrieved.

The mcbackend.core module has classes such as Backend, Run, and Chain to define these interfaces for any storage backend, no matter if it's an in-memory, filesystem or database storage. Albeit this implementation is currently Python-only, the interface signature should be portable to e.g. C++.

Via mcbackend.backends the McBackend package then provides backend implementations. Currently you may choose from:

backend = mcbackend.NumPyBackend()
backend = mcbackend.ClickHouseBackend( client=clickhouse_driver.Client("localhost") )

# All that matters:
isinstance(backend, mcbackend.Backend)
# >>> True

Part 3: PPL adapters

Anything that is a Backend can be wrapped by an adapter that makes it compatible with your favorite PPL.

In the example below, a ClickHouseBackend is initialized to store MCMC draws from a PyMC model in a ClickHouse database. See below for how to run it in Docker.

import clickhouse_driver
import mcbackend
import pymc as pm

# 1. Create _any_ kind of backend
ch_client = clickhouse_driver.Client("localhost")
backend = mcbackend.ClickHouseBackend(ch_client)

with pm.Model():
    # 2. Create your model
    ...
    # 3. Hit the inference button ™ while passing the backend!
    pm.sample(trace=backend)

In case of PyMC the adapter lives in the PyMC codebase since version 5.1.1, so all you need to do is pass any mcbackend.Backend via the pm.sample(trace=...) parameter!

Instead of using PyMC's built-in NumPy backend, the MCMC draws now end up in ClickHouse.

Retrieving the draws & stats

Continuing the example from above we can now retrieve draws from the backend.

Note that since this example wrote the draws to ClickHouse, we could run the code below on another machine, and even while the above model is still sampling!

backend = mcbackend.ClickHouseBackend(ch_client)

# Fetch the run from the database (downloads just metadata)
run = backend.get_run(trace.run_id)

# Get all draws from a chain
chain = run.get_chains()[0]
chain.get_draws("my favorite variable")
# >>> array([ ... ])

# Convert everything to an inference data structure
idata = run.to_inferencedata()
print(idata)
# >>> DataTree:
# >>> 	> posterior
# >>> 	> sample_stats
# >>> 	> observed_data
# >>> 	> constant_data
# >>>
# >>> Warmup iterations saved (warmup_*).

Contributing what's next

McBackend just started and is looking for contributions. For example:

• Schema discussion: Which metadata is needed? (related: PyMC #5160)
• Interface discussion: How should Backend/Run/Chain evolve?
• Python Backends for disk storage (HDF5, *.proto, ...)
• C++ Backend/Run/Chain interfaces
• C++ ClickHouse backend (via clickhouse-cpp)

As the schema and API stabilizes a mid-term goal might be to replace PyMC BaseTrace/MultiTrace entirely to rely on mcbackend.

Getting rid of MultiTrace was a long-term goal behind making pm.sample(return_inferencedata=True) the default.

Development

First clone the repository and set up a development environment containing the protobuf compiler.

mamba create -n mcb python=3.13 grpcio-tools protobuf -y
activate mcb
pip install -r requirements-dev.txt
pip install --pre "betterproto[compiler]"
pip install -e .

To compile the *.proto files for languages other than Python, check the ProtocolBuffers documentation.

The following script compiles them for Python using the betterproto compiler plugin to get nice-looking dataclasses. It also copies the generated files to the right place in mcbackend.

python protobufs/generate.py
pre-commit run --all

To run the tests:

pytest -v

Some tests need a ClickHouse database server running locally. To start one in Docker:

docker run --detach --rm --name mcbackend-db -p 9000:9000 --ulimit nofile=262144:262144 clickhouse/clickhouse-server