Sauerkraut: Serializing Python's Control State

As pickle serializes Python's data state, sauerkraut serializes Python's control state.

Concretely, this library equips Python with the ability to stop a running Python function, serialize it to the network or disk, and resume later on. This is useful in different contexts, such as dynamic load balancing and checkpoint/restart. Sauerkraut is designed to be lightning fast for latency-sensitive HPC applications. Internally, FlatBuffers are created to store the serialized state.

This library is still experimental, and subject to change.

Examples

Using Greenlets (Recommended)

The following example shows how to serialize a function's state using greenlets, which provides a cleaner interface:

import sauerkraut
import greenlet
import numpy as np

def fun1(c):
    g = 4
    a = np.array([1, 2, 3])
    # Switch back to parent, preserving our state
    greenlet.getcurrent().parent.switch()
    
    # When we resume, continue here
    a += 1
    print(f'c={c}, g={g}, a={a}')
    return 3

# Create and start the greenlet
f1_gr = greenlet.greenlet(fun1)
f1_gr.switch(13)

# Serialize the greenlet's state
serframe = sauerkraut.copy_frame_from_greenlet(f1_gr, serialize=True)

# Save to disk
with open('serialized_frame.bin', 'wb') as f:
    f.write(serframe)

# Read from disk and resume execution
with open('serialized_frame.bin', 'rb') as f:
    read_frame = f.read()
code = sauerkraut.deserialize_frame(read_frame)
gr = greenlet.greenlet(sauerkraut.run_frame)
retval = gr.switch(code)
print(f"Done on the parent, child returned {retval}")

Low-level API

This example demonstrates the lower-level frame copying and serialization API:

import sauerkraut as skt
calls = 0

def fun1(c):
    global calls
    calls += 1
    g = 4
    for i in range(3):
      if i == 0:
          print("Copying frame")
          # copy_frame makes a copy of the current control + data states.
          # When we later run the frame (run_frame), execution will continue
          # directly after the call.
          # Therefore, this line will return twice:
          # The first time when the frame is copied,
          # the second time when the frame is resumed.
          frm_copy = skt.copy_current_frame()
      # Because copy_current_frame will return twice, we need to differentiate
      # between the different returns to avoid getting stuck in a loop!
      if calls == 1:
          # The copied frame does not see this write to g
          g = 5
          calls += 1
          # This variable is not serialized,
          # as it's not live
          hidden_inner = 55
          return frm_copy
      else:
          # When running the copied frame, we take this branch.
          # This line will run 3 times
          print(f'calls={calls}, c={c}, g={g}')

    calls = 0
    return 3

# Create and serialize a frame
frm = fun1(13)
serframe = skt.serialize_frame(frm)

# Save to disk
with open('serialized_frame.bin', 'wb') as f:
    f.write(serframe)

# Read from disk and resume execution
with open('serialized_frame.bin', 'rb') as f:
    read_frame = f.read()
code = skt.deserialize_frame(read_frame)
skt.run_frame(code)

Installation

From PyPI (Recommended)

pip install sauerkraut

From Source

python3 -m pip install -r requirements.txt
python3 -m pip install .

Using Docker

# Build the Docker image (takes 5-10 minutes)
docker build -t sauerkraut -f Dockerfile .

# Run the container
docker run --name="sauerkraut_img" -it library/sauerkraut

# Inside the container, you can run examples:
cd /sauerkraut/examples
python3 copy_then_serialize.py

Compatibility

Sauerkraut leverages intimate knowledge of CPython internals, and as such is vulnerable to changes in the CPython API and VM. Currently, Sauerkraut supports Python 3.13 and 3.14.