To run this example locally, install Ploomber and execute: ploomber examples -n guides/serialization

To start a free, hosted JupyterLab: binder-logo

Found an issue? Let us know.

Have questions? Ask us anything on Slack.

Serialization

Tutorial explaining how the serializer and unserializer fields in a pipeline.yaml file work.

Incremental builds allow Ploomber to skip tasks whose source code hasn’t changed; each task must save their products to disk to enable such a feature. However, there are some cases when we don’t want our pipeline to perform disk operations. For example, if we’re going to deploy our pipeline, eliminating disk operations reduces runtime considerably.

To enable a pipeline to work in both disk-based and in-memory scenarios, we can declare a serializer and unzerializer in our pipeline declaration, effectively separating our task’s logic from the read/write logic.

Note that this only applies to function tasks; other tasks are unaffected by the serializer/unserializer configuration.

Built-in pickle serialization

The easiest way to get started is to use the built-in serializer and unserializer, which use the pickle module.

Let’s see an example; the following pipeline has two tasks. The first one generates a dictionary, and the second one creates two dictionaries. Since we are using the pickle-based serialization, each dictionary is saved in the pickle binary format:

# Content of simple.yaml
serializer: ploomber.io.serializer_pickle
unserializer: ploomber.io.unserializer_pickle

tasks:
  - source: tasks.first
    product: output/one_dict

  - source: tasks.second
    product:
        another: output/another_dict
        final: output/final_dict

Let’s take a look at the task’s source code:

# Content of tasks.py
def first():
    return dict(a=1, b=2)


def second(upstream):
    first = upstream['first']
    another = dict(a=first['b'] + 1, b=first['a'] + 1)
    final = dict(a=100, b=200)
    return dict(another=another, final=final)

Since we configured a serializer and unserializer, function tasks must return their outpues instead of saving them to disk in the function’s body.

first does not have any upstream dependencies and returns a dictionary. second has the previous task as dependency and returns two dictionaries. Note that the keys in the returned dictionary must match the names of the products declared in pipeline.yaml (another, final).

Let’s now run the pipeline.

[1]:
%%sh
ploomber build --entry-point simple.yaml --force
name    Ran?      Elapsed (s)    Percentage
------  ------  -------------  ------------
first   True         0.001281       42.9433
second  True         0.001702       57.0567
Building task 'second': 100%|██████████| 2/2 [00:06<00:00,  3.39s/it]

The pickle format has important security concerns, remember only to unpickle data you trust.

Custom serialization logic

We can also define our own serialization logic, by using the @serializer, and @unserializer decorators. Let’s replicate what our pickle-based serializer/unserializer is doing as an example:

# Content of custom.py
from pathlib import Path
import pickle

from ploomber.io import serializer, unserializer


@serializer()
def my_pickle_serializer(obj, product):
    Path(product).write_bytes(pickle.dumps(obj))


@unserializer()
def my_pickle_unserializer(product):
    return pickle.loads(Path(product).read_bytes())

A @serializer function must take two arguments: the object to serializer and the product object (taken from the task declaration). The @unserializer must take a single argument (the product to unserializer), and return the unserializer object.

Let’s modify our original pipeline to use this serializer/unserializer:

# Content of custom.yaml
serializer: custom.my_pickle_serializer
unserializer: custom.my_pickle_unserializer

tasks:
  - source: tasks.first
    product: output/one_dict

  - source: tasks.second
    product:
        another: output/another_dict
        final: output/final_dict
[2]:
%%sh
ploomber build --entry-point custom.yaml --force
name    Ran?      Elapsed (s)    Percentage
------  ------  -------------  ------------
first   True         0.001216       19.4342
second  True         0.005041       80.5658
Building task 'second': 100%|██████████| 2/2 [00:07<00:00,  3.87s/it]

Custom serialization logic based on the product’s extension

Under many circumstances, there are more suitable formats than pickle. For example, we may want to store lists or dictionaries as JSON files and other files using pickle. The @serializer/@unserializer decorators use mapping as the first argument to dispatch to different functions depending on the product’s extension. Let’s see an example:

# Content of custom.py
from pathlib import Path
import pickle
import json

from ploomber.io import serializer, unserializer


def write_json(obj, product):
    Path(product).write_text(json.dumps(obj))


def read_json(product):
    return json.loads(Path(product).read_text())


@serializer({'.json': write_json})
def my_serializer(obj, product):
    Path(product).write_bytes(pickle.dumps(obj))


@unserializer({'.json': read_json})
def my_unserializer(product):
    return pickle.loads(Path(product).read_bytes())

Let’s modify our example pipeline. The product in the first task does not have an extension (output/one_dict), hence, it will use pickle-based logic. However, the tasks in the second task have a .json extension, and will be saved as JSON files.

# Content of with-json.yaml
serializer: custom.my_serializer
unserializer: custom.my_unserializer

tasks:
  - source: tasks.first
    product: output/one_dict

  - source: tasks.second
    product:
        another: output/another_dict.json
        final: output/final_dict.json
[3]:
%%sh
ploomber build --entry-point with-json.yaml --force
name    Ran?      Elapsed (s)    Percentage
------  ------  -------------  ------------
first   True         0.001193       38.5834
second  True         0.001899       61.4166
Building task 'second': 100%|██████████| 2/2 [00:06<00:00,  3.26s/it]

Let’s print the .json files to verify they’re not pickle files:

[4]:
%%sh
cat output/another_dict.json
{"a": 3, "b": 2}
[5]:
%%sh
cat output/final_dict.json
{"a": 100, "b": 200}

Using a fallback format

Since it’s common to have a fallback serialization format, the decorators have a fallback argument that, when enabled, uses the pickle module when the product’s extension does not match any of the registered ones in the first argument.

The example works the same as the previous one, except we don’t have to write our pickle-based logic.

fallback can also take the joblib or cloudpickle values. They’re similar to the pickle format but have some advantages. For example, joblib produces smaller files when the serialized object contains many NumPy arrays, while cloudpickle supports serialization of some objects that the pickle module doesn’t. To use fallback='joblib' or fallback='cloudpickle' the corresponding module must be installed.

# Content of custom.py

from ploomber.io import serializer, unserializer


@serializer({'.json': write_json}, fallback=True)
def my_fallback_serializer(obj, product):
    pass


@unserializer({'.json': read_json}, fallback=True)
def my_fallback_unserializer(product):
    pass
# Content of fallback.yaml
serializer: custom.my_fallback_serializer
unserializer: custom.my_fallback_unserializer

tasks:
  - source: tasks.first
    product: output/one_dict

  - source: tasks.second
    product:
        another: output/another_dict.json
        final: output/final_dict.json
[6]:
%%sh
ploomber build --entry-point fallback.yaml --force
name    Ran?      Elapsed (s)    Percentage
------  ------  -------------  ------------
first   True         0.002278       56.8505
second  True         0.001729       43.1495
Building task 'second': 100%|██████████| 2/2 [00:06<00:00,  3.45s/it]

Let’s print the JSON files to verify their contents:

[7]:
%%sh
cat output/another_dict.json
{"a": 3, "b": 2}
[8]:
%%sh
cat output/final_dict.json
{"a": 100, "b": 200}

Using default serializers

Ploomber comes with a few convenient serialization functions to write more succint serializers. We can request the use of such default serializers using the defaults argument, which takes a list of extensions:

# Content of custom.py

from ploomber.io import serializer, unserializer


@serializer(fallback=True, defaults=['.json'])
def my_defaults_serializer(obj, product):
    pass


@unserializer(fallback=True, defaults=['.json'])
def my_defaults_unserializer(product):
    pass

Here we’re asking to dispatch .json products and use pickle for all other extensions, the same as we did for the previous examples, except this time, we don’t have to pass the mapping argument to the decorators.

defaults support:

  1. .json: the returned object must be JSON-serializable (e.g., a list or a dictionary)

  2. .txt: the returned object must be a string

  3. .csv: the returned object must be a pandas.DataFrame

  4. .parquet: the returned object must be a pandas.DataFrame, and a parquet library should be installed (such as pyarrow).

# Content of defaults.yaml
serializer: custom.my_defaults_serializer
unserializer: custom.my_defaults_unserializer

tasks:
  - source: tasks.first
    product: output/one_dict

  - source: tasks.second
    product:
        another: output/another_dict.json
        final: output/final_dict.json
[9]:
%%sh
ploomber build --entry-point defaults.yaml --force
name    Ran?      Elapsed (s)    Percentage
------  ------  -------------  ------------
first   True         0.001148       39.1675
second  True         0.001783       60.8325
Building task 'second': 100%|██████████| 2/2 [00:07<00:00,  3.64s/it]

Let’s print the JSON files to verify their contents:

[10]:
%%sh
cat output/another_dict.json
{"a": 3, "b": 2}
[11]:
%%sh
cat output/final_dict.json
{"a": 100, "b": 200}

Wrapping up

Configuring a serializer and unserializer in your pipeline.yaml is optional, but it helps you quickly generate a fully in-memory pipeline for serving predictions.

If you want to learn more about in-memory pipelines, check out the following guide.

For a complete example showing how to manage a training and a serving pipeline, and deploy it as a Flask API, click here.