Skip to content

Latest commit

 

History

History
407 lines (300 loc) · 13.4 KB

README.md

File metadata and controls

407 lines (300 loc) · 13.4 KB

TinyBaker: Composable, first-order, file-to-file transformations in Python

Python Package

TinyBaker is in beta release.

TinyBaker allows programmers to define first-order file-to-file transformations in a concise format and compose them together with clarity.

Installation via pip install tinybaker

TinyBaker Logo

The Problem

Many programs can be considered transformations between source and destination files. Training machine learning models, running predictions on dataframes, processing logs, concatenation, compilation, and many others, are examples of tasks that fundamentally pose a transformation from one set of files to another.

Since transforms aren't normally considered a first-order concept, they get wildly unwieldy to work with. Production workloads are configured separately from local transformations. Getting a local script working on production often requires lots of rework, mocking, testing, and iteration by a product team.

The Solution

Tinybaker turns file-to-file transforms into a first-order concept.

TinyBaker transforms can be configured, run, composed, hosted, and tested, all independently from their specific implementations.

The Model

The main component of TinyBaker is the base class Transform: a standalone mapping from one set of files to another.

                 ___________
---[ file1 ]--->|           |
                |           |->--[ file4 ]---
---[ file2 ]--->| Transform |
                |           |->--[ file5 ]---
---[ file3 ]--->|___________|

For example, let's say we were running predictions over a certain ML model. Such a transform might conceptually look like this:

                  ___________
---[ config ]--->|           |
                 |           |->--[ predictions ]---
---[ model ]---->|  Predict  |
                 |           |->--[ performance ]---
---[ data ]----->|___________|

TinyBaker calls the labels associated which each input / output file a tag.

                  ___________
---[ config ]--->|           |
      ^ Tag      |           |->--[ predictions ]---
---[ model ]---->|  Predict  |       ^ Tag
      ^ Tag      |           |->--[ performance ]---
---[ data ]----->|___________|       ^ Tag
      ^ Tag

We might want to configure where we store input/output files, or configure files to come from separate filesystems entirely. TinyBaker allows you to define the transform while paying attention to only the tags, even when accessing files across multiple filesystems.

                                        ___________
/path/to/config.json >----[ config ]-->|           |
                                       |           |->--[ predictions ]---> hdfs://outputs/predictions.csv
s3://path/to/model.pkl >--[ model ]--->|  Predict  |
                                       |           |->--[ performance ]---> ./performance.pkl
/path/to/data.csv >-------[ data ]---->|___________|

We can imagine a situation where we have file transformations that could theoretically compose:

                   ________________
                  |                |
---[ raw_logs ]-->| BuildDataFrame |->--[ df ]---
                  |________________|
                  
             ____________
            |            |
---[ df ]-->| BuildModel |->--[ model ]---
            |____________|

TinyBaker allows you to compose these two transformations together:

                   ___________________________
                  |                           |
---[ raw_logs ]-->| BuildDataFrame+BuildModel |->--[ model ]---
                  |___________________________|

We now only need to specify the location of 2 files-- TinyBaker handles linking the two steps together

                                 ___________________________
                                |                           |
/raw/logs.txt ---[ raw_logs ]-->| BuildDataFrame+BuildModel |->--[ model ]--- /path/to/model.pkl
                                |___________________________|

Extra file dependencies are propagated to the top level of a sequence, ensuring you'll never miss a file dependency in step 5 of 17, e.g.

                   ________________
                  |                |
---[ raw_logs ]-->| BuildDataFrame |->--[ df ]---
                  |________________|
                  
                 ____________
---[ df ]------>|            |
                | BuildModel |->--[ model ]---
---[ config ]-->|____________|
            
# Goes to...

                   ___________________________
---[ raw_logs ]-->|                           |
                  | BuildDataFrame+BuildModel |->--[ model ]---
---[ config ]---->|___________________________|

In-Code Anatomy of a single transform

The following describes a minimal transform one can define in TinyBaker

from tinybaker import Transform, InputTag, OutputTag

class SampleTransform(Transform):
  # 1 tag per input file
  first_input = InputTag("first_input")
  second_input = InputTag("second_input")
  some_output = OutputTag("some_output")

  # self.script describes what actually executes when the transform task runs
  script(self):
    # Within scripts, one can operate on tags as if they're FileRefs
    with self.first_input.open() as f:
      do_something_with(f)
    with self.second_input.open() as f:
      do_something_else_with(f)

    # and output or something
    with self.some_output.open() as f:
      write_something_to(f)

This would then be executed via:

SampleTransform(
  input_paths={"first_input": "path/to/input1", "second_input"= "path/to/input2"}
  output_paths={"some_output": "path/to/write/output"}
).run()

Real-world example of a single transform

For a real-world example, consider training an ML model. This is a transformation from the two files some/path/train.csv and some/path/test.csv to a pickled ML model another/path/some_model.pkl and statistics. With tinybaker, you can specify this individual configurable step as follows:

# train_step.py
from tinybaker import Transform, cli, InputTag, OutputTag
import pandas as pd
from some_cool_ml_library import train_model, test_model

class TrainModelStep(Transform):
  train_csv = InputTag("train_csv")
  test_csv = InputTag("test_csv")
  pickled_model = OutputTag("pickled_model")
  results = OutputTag("results")

  def script():
    # Read from files
    with self.train_csv.open() as f:
      train_data = pd.read_csv(f)
    with self.test_csv.open() as f:
      test_data = pd.read_csv(f)

    # Run computations
    X = train_data.drop(["label"])
    Y = train_data[["label"]]
    [model, train_results] = train_model(X, Y)
    test_results = test_model(model, test_data)

    # Write to output files
    with self.results.open() as f:
      results = train_results.formatted_summary() + test_results.formatted_summary()
      f.write(results)
    with self.pickled_model.openbin() as f:
      pickle.dump(f, model)

if __name__ == "__main__":
  cli(SampleTransform)

Operating over multiple filesystems

Since TinyBaker uses fsspec as its filesystem, TinyBaker can use any filesystem that fsspec supports. For example, you can use s3 via setting the protocol of files to s3://

This makes building test suites for transforms very easy: test suites can operate off of local data, but production jobs can run off of s3 data.

Validation

TinyBaker performs simple validation, such as raising early if input files are missing, or erroring if fully-qualified file paths form a cycle.

Combining several build steps

We can compose several build steps together using the methods merge and sequence.

from tinybaker import Transform, sequence

class CleanLogs(Transform):
  raw_logfile = InputTag("raw_logfile")
  cleaned_logfile = OutputTag("cleaned_logfile")
  # ...

class BuildDataframe(Transform):
  cleaned_logfile = InputTag("cleaned_logfile")
  dataframe = OutputTag("dataframe")
  # ...

class BuildLabels(Transform):
  cleaned_logfile = InputTag("cleaned_logfile")
  labels = OutputTag("labels")
  # ...

class TrainModelFromDataframe(Transform):
  dataframe = InputTag("dataframe")
  labels = InputTag("labels")
  trained_model = OutputTag("trained_model")
  # ...


TrainFromRawLogs = sequence(
  CleanLogs,
  merge(BuildDataframe, BuildLabels),
  TrainModelFromDataframe
)

task = TrainFromRawLogs(
  input_paths={"raw_logfile": "/path/to/raw.log"},
  output_paths={"trained_model": "/path/to/model.pkl"}
)

task.run()

Inputs and outputs are hooked up via tag names, e.g. if step 1 outputs tag "foo", and step 2 takes tag "foo" as inputs, TinyBaker will be automatically hook them together.

Propagation of inputs and outputs

Let's say task 3 of 4 in a sequence of tasks requires tag foo, but no previous step generates tag foo, then this dependency will be propagated to the top level; the sequence as a whole will have a dependency on tag foo.

Additionally, if task 3 of 4 generates a tag bar, but no further step requires bar, then the sequence exposes "bar" as an output.

expose_intermediates

If you need to expose intermediate files within a sequence, you can use the keywork arg expose_intermediates to additionally output the listed intermediate tags, e.g.

sequence([A, B, C], expose_intermediates={"some_intermediate", "some_other_intermediate"})

Renaming Tags

Right now, since association of files from one step to the next is based on tags, we may end up in a situation where we want to rename tags. If we want to change the tag names, we can use map_tags to change them.

from tinybaker import map_tags

MappedStep = map_tags(
  SomeStep,
  input_mapping={"old_input_name": "new_input_name"},
  output_mapping={"old_output_name": "new_output_name"})

CLI

TinyBaker can instantly turn a transform into a CLI:

from tinybaker import Transform, cli

class MNISTPipeline(Transform):
  # as defined in tests/slow/test_real_world.py
  # [...]

if __name__ == "__main__":
  cli(MNISTPipeline)

The above would yield the below when run:

$ python ./mnist_pipeline_transform.py --help
usage: test_real_world.py [-h] --raw_train_images RAW_TRAIN_IMAGES --raw_test_images
                          RAW_TEST_IMAGES --accuracy ACCURACY --model MODEL
                          [--version] [--overwrite]

Execute a MNISTPipeline transform

optional arguments:
  -h, --help            show this help message and exit
  --raw_train_images RAW_TRAIN_IMAGES
                        Path for output tag raw_train_images
  --raw_test_images RAW_TEST_IMAGES
                        Path for output tag raw_test_images
  --accuracy ACCURACY   Path for output tag accuracy
  --model MODEL         Path for output tag model
  --version             show program's version number and exit
  --overwrite           Whether to overwrite any existing output files

No need to write argument parsers -- TinyBaker knows what arguments the transform needs and builds a CLI around it.

Filesets

If a step operates over a dynamic set of files (e.g. logs from n different days), you can use the filesets interface to specify that. Tags that begin with the prefix fileset:: are interpreted to be filesets rather than just files.

If a sequence includes a fileset as an intermediate, then TinyBaker expects the developer to specify the paths of the intermediate, via expose_intermediates. This is a relatively fundamental restriction of the platform, as TinyBaker expects that all paths are specified before script execution.

Example

A concat task can be done as follows:

class Concat(Transform):
    files = InputTag("fileset::files")
    concatted = InputTag("concatted")

    def script(self):
        content = ""
        for ref in self.files:
            with ref.open() as f:
                content = content + f.read()

        with self.concatted.open() as f:
            f.write(content)

Concat(
    input_paths={
        "fileset::files": ["./tests/__data__/foo.txt", "./tests/__data__/bar.txt"],
    },
    output_paths={"concatted": "/tmp/concatted"},
    overwrite=True,
).run()

Experimental API: File-style Transform Definitions

Transforms can be specified in a script-like format:

# train_model.py
from tinybaker import InputTag, OutputTag, cli

train_csv = InputTag("train_csv")
test_csv = InputTag("test_csv")

results = OutputTag("results")
pickled_model = OutputTag("pickled_model")

def script():
  # Read from files
  with train_csv.open() as f:
    train_data = pd.read_csv(f)

  with test_csv.open() as f:
    test_data = pd.read_csv(f)

  # Run computations
  X = train_data.drop(["label"])
  Y = train_data[["label"]]
  [model, train_results] = train_model(X, Y)
  test_results = test_model(model, test_data)

  # Write to output files
  with results.open() as f:
    results = train_results.formatted_summary() + test_results.formatted_summary()
    f.write(results)
  with pickled_model.openbin() as f:
    pickle.dump(f, model)


if __name__ == "__main__":
  # We can still define a cli under this format.
  cli(locals())

These can be converted to transforms via:

from tinybaker import Transform
from . import train_model

TrainModelTransform = Transform.from_namespace(train_model)

# This can be consumed just like any other job.
job = TrainModelTransform(input_files={...}, output_files={...})
job.run()

Contributing

Please contribute!