3 releases

0.0.14 Aug 8, 2021
0.0.12 Aug 4, 2021
0.0.1 Jul 25, 2021

#767 in Machine learning

MIT license

520KB
13K SLoC

Aorist

Aorist is a code-generation tool for MLOps. Its aim is to generate legible code for common repetitive tasks in data science, such as data replication, common transformations, as well as machine learning operations.

How to build

Make sure you have Rust installed, following the instructions there, then run:

cargo build
cp target/debug/libaorist.so example/aorist.so

To try it out:

./run.sh
python example/minimal.py

The command above generates Python code or R code -- you can pipe the code to a file.

Anaconda install (broken)

These instructions require Anaconda and were tested against Ubuntu Linux 20.04 LTS.

conda create -n aorist python=3.8 anaconda
conda activate aorist
pip3 install https://storage.googleapis.com/scienz-artifacts/aorist-0.0.1-cp38-cp38-linux_x86_64.whl

Overview of an Aorist universe

Let's say we are starting a new project which involves analyzing a number of large graph datasets, such as the ones provided by the SNAP project.

We will conduct our analysis in a mini data-lake, such as the Trino + MinIO solution specified by Walden.

We would like to replicate all these graphs into our data lake before we can start analyzing them. At a very high-level, this is achieved by defining a "universe", the totality of things we care about in our project. One such universe is specified below:

from snap import snap_dataset
from aorist import (
    dag,
    Universe,
    ComplianceConfig,
    HiveTableStorage,
    MinioLocation,
    StaticHiveTableLayout,
    ORCEncoding,
)
from common import DEFAULT_USERS, DEFAULT_GROUPS, DEFAULT_ENDPOINTS

universe = Universe(
    name="my_cluster",
    datasets=[
        snap_dataset,
    ],
    endpoints=DEFAULT_ENDPOINTS,
    users=DEFAULT_USERS,
    groups=DEFAULT_GROUPS,
    compliance=ComplianceConfig(
        description="""
        Testing workflow for data replication of SNAP data to
        local cluster. The SNAP dataset collection is provided
        as open data by Stanford University. The collection contains
        various social and technological network graphs, with
        reasonable and systematic efforts having been made to ensure
        the removal of all Personally Identifiable Information.
        """,
        data_about_human_subjects=True,
        contains_personally_identifiable_information=False,
    ),
)

The universe definition contains a number of things:

  • the datasets we are talking about (more about it in a bit),
  • the endpoints we have available (e.g. the fact that a MinIO server is available for storage, as opposed to HDFS or S3, etc., and where that server is available; what endpoint we should use for Presto / Trino, etc.)
  • who the users and groups are that will access the dataset,
  • some compliance annotations.

Note: Currently users, groups, and compliance annotations are supported as a proof of concept -- these concepts are not essential to an introduction so we will skip them for now.

Generating a DAG

To generate a flow that replicates our data all we have to do is run:

DIALECT = "python"
out = dag(
  universe, [
    "AllAssetsComputed",
  ], DIALECT
)

This will generate a set of Python tasks, which will do the following, for each asset (i.e., each graph) in our dataset:

  • download it from its remote location,
  • decompress it, if necessary
  • remove its header,
  • convert the file to a CSV, if necessary
  • upload the CSV data to MinIO
  • create a Hive table backing the MinIO location
  • convert the CSV-based Hive table to an ORC-based Hive table
  • drop the temporary CSV-based Hive table

This set of tasks is also known as a Directed Acyclic Graph (DAG). The same DAG can be generated as a Jupyter notebook, e.g. by setting:

DIALECT = "jupyter"

Or we can set DIALECT to "airflow" for an Airflow DAG.

Aside: what is actually going on?

What Aorist does is quite complex -- the following is an explanation of the conceptual details, but you can skip this if you'd want something a bit more concrete:

  • first, you describe the universe. This universe is actually a highly-structured hierarchy of concepts, each of which can be "constrained".
  • A constraint is something that "needs to happen". In this example all you declare that needs to happen is the constraint AllAssetsComputed. This constraint is attached to the Universe, which is a singleton object.
  • Constraints attach to specific kinds of objects -- some attach to the entire Universe, others attach to tables, etc.
  • Constraints are considered to be satisfied when their dependent constraints are satisfied. When we populate each constraint's own dependent constraints we follow a set of complex mapping rules that are nonetheless fairly intuitive (but difficult to express without a longer discussion, see the end of this document for that)
  • Programs ("recipes") are attached to this constraint graph by a Driver. The Driver decides which languages are prefered (e.g. maybe the Driver likes Bash scrips more than Presto, etc.). The driver will complain if it can't provide a solution for a particular constraint.
  • Once the recipes are attached, various minutiae are extracted from the concept hierarchy -- e.g., which endpoints to hit, actual schemas of input datasets, etc.
  • Once the various minutiae are filled in, we have a graph of Python code snippets. If these snippets are repetitive (e.g. 100 instances of the same function call but with different arguments) we compress them into for loops over parameter dictionarie.
  • We then take the compressed snippet-graph and further optimize it, for instance by pushing repeated parameters out of parameter dictionaries and into the main body of the for loop.
  • We also compute unique, maximally-descriptive names for the tasks, a combination of the constraint name and the concept's position in the hierarchy. (e.g. wine_table_has_replicated_schema). These names are shortened as much as possible while still being unique (e.g., we may shorten things to wine_schema, a less mouthful of a task name).
  • The driver then adds scaffolding for native Python, Airflow or Jupyter code generation. Other output formats (e.g. Prefect, Dagster, Makefiles, etc.) will be supported in the future.
  • Finally, the driver converts the generated Python AST to a concrete string, which it then formats as a pretty (PEP8-compliant) Python program via Python black.

Describing a dataset

Before we can turn our attention to what we would like to achieve with our data, we need to determine what the data is, to begin with. We do so via a dataset manifest, which is created using Python code.

Here's an example of how we'd create such a manifest for a canonical ML dataset (the Wine dataset, as per example/wine.py).

First, we define our attribute list:

attributes = attr_list([
    attr.Categorical("wine_class_identifier"),
    attr.PositiveFloat("alcohol"),
    attr.PositiveFloat("malic_acid"),
    attr.PositiveFloat("ash"),
    attr.PositiveFloat("alcalinity_of_ash"),
    attr.PositiveFloat("magnesium"),
    attr.PositiveFloat("total_phenols"),
    attr.PositiveFloat("non_flavanoid_phenols"),
    attr.PositiveFloat("proanthocyanins"),
    attr.PositiveFloat("color_intensity"),
    attr.PositiveFloat("hue"),
    attr.PositiveFloat("od_280__od_315_diluted_wines"),
    attr.PositiveFloat("proline"),
])

Then, we express the fact that a row corresponds to a struct with the fields defined in the attributes list:

wine_datum = RowStruct(
    name="wine_datum",
    attributes=attributes,
)

Then, we declare that our data can be found somewhere on the Web, in the remote storage. Note that we also record the data being CSV-encoded, and the location corresponding to a single file. This is where we could note compression algorithms, headers, etc.:

remote = RemoteStorage(
    location=WebLocation(
        address=("https://archive.ics.uci.edu/ml/"
                 "machine-learning-databases/wine/wine.data"),
    ),
    layout=SingleFileLayout(),
    encoding=CSVEncoding(),
)

We need this data to live locally, in a Hive table in ORC format, backed by a MinIO location with the prefix wine:

local = HiveTableStorage(
    location=MinioLocation(name="wine"),
    layout=StaticHiveTableLayout(),
    encoding=ORCEncoding(),
)

Note a few things:

  • we don't specify the table name, as this is automatically-generated from the asset name (we will define that momentarily)
  • we declare, "this thing needs to be stored in MinIO", but do not concern ourselves with endpoints at this moment. Aorist will find the right endpoints for us and fill in secrets, etc. Or if MinIO is unavailable it will fail.
  • this is also where we can indicate whether our table is static (i.e. there is no time dimension, or dynamic).

We are now ready to define our asset, called wine_table:

wine_table = StaticDataTable(
    name="wine_table",
    schema=default_tabular_schema(wine_datum),
    setup=RemoteImportStorageSetup(
        tmp_dir="/tmp/wine",
        remote=remote,
        local=[local],
    ),
    tag="wine",
)

Here's what we do here:

  • we define an asset called wine_table. This is also going to be the name of any Hive table that will be created to back this asset (or file, or directory, etc., depending on the dataset storage).
  • we also define a schema. A schema tells us exactly how we can turn a row into a template. For instance, we need the exact order of columns in a row to know unambiguously how to convert it into a struct.
  • default_tabular_schema is a helper function that allows us to derive a schema where columnns in the table are in exactly the same order as fields in the struct.
  • the setup field introduces the notion of a "replicated" remote storage, via RemoteImportStorageSetup. The idea expressed here is that we should make sure the data available at the remote location is replicated exactly in the local locations (either by copying it over, or, if already availalbe, by checking that the remote and target data has the same checksum, etc.)
  • we also use a tag field to help generate legible task names and IDs (e.g., in Airflow)

Finally, let's define our dataset:

wine_dataset = DataSet(
    name="wine",
    description="A DataSet about wine",
    sourcePath=__file__,
    datumTemplates=[wine_datum],
    assets={"wine_table": wine_table},
)

This dataset can then be imported into the universe discussed previously.

Aside: The asset / template split

An Aorist dataset is meant to be a collection of two things:

  • data assets -- concrete information, stored in one or multiple locations, remotely, on-premise, or in some sort of hybrid arrangement.
  • datum templates -- information about what an instance of our data (i.e., a datum) represents.

For instance, a table is a data asset. It has rows and columns, and those rows and columns are filled with some values that can be read from some location.

What those rows and columns mean depends on the template. Oftentimes rows in tables translate to structs, for instance in a typical dim_customers table. But if we're talking about graph data, then a row in our table represents a tuple (more specifically a pair), and not a struct.

Other examples of data assets would be:

  • directories with image files,
  • concrete machine learning models,
  • aggregations,
  • scatterplots,

Other examples of data templates could be:

  • a tensor data template corresponding to RGB images,
  • an ML model template that takes a certain set of features (e.g. number of rooms and surface of a house, and produces a prediction, e.g. a valuation),
  • a histogram data template, expressing the meaning of margin columns used for aggregations, as well as the aggregation function (a count for a histogram)
  • a scatterplot template, encoding the meaning of the x and y axis, etc.

This conceptual differentiation allows us to use the same template to refer to multiple assets. For instance, we may have multiple tables with exactly the same schema, some being huge tables with real data, and others being downsampled tables used for development. These tables should be refered to using the same template.

This is also very useful in terms of tracking data lineage, on two levels: semantically (how does template Y follow from template X?) and concretely (how does row A in table T1 follow from row B in table T2?).

Back to the SNAP dataset

The SNAP dataset we discussed initially is a bit different from the simple Wine dataset. For one, it contains many assets -- this is a collection of different graphs used for Machine Learning applications -- each graph is its own asset. But the meaning of a row remains the same: it's a 2-tuple made up of identifiers. We record this by defining the template:

edge_tuple = IdentifierTuple(
    name="edge",
    attributes=attr_list([
        attr.NumericIdentifier("from_id"),
        attr.NumericIdentifier("to_id"),
    ]),
)

Then we define an asset for each of 12 datasets. Note that the names come from the URL patterns corresponding to each dataset. We need to replace dashes with underscores when creating asset names however (Hive tables don't like dashes in their names):

names = [
    "ca-AstroPh", "ca-CondMat", "ca-GrQc", "ca-HepPh",
    "ca-HepTh", "web-BerkStan", "web-Google", "web-NotreDame",
    "web-Stanford", "amazon0302", "amazon0312", "amazon0505",
]
tables = {}
for name in names:

    name_underscore = name.replace("-", "_").lower()
    remote = RemoteStorage(
        location=WebLocation(
            address="https://snap.stanford.edu/data/%s.txt.gz" % name,
        ),
        layout=SingleFileLayout(),
        encoding=TSVEncoding(
            compression=GzipCompression(),
            header=UpperSnakeCaseCSVHeader(num_lines=4),
        ),
    )
    local = HiveTableStorage(
        location=MinioLocation(name=name_underscore),
        layout=StaticHiveTableLayout(),
        encoding=ORCEncoding(),
    )
    table = StaticDataTable(
        name=name_underscore,
        schema=default_tabular_schema(edge_tuple),
        setup=RemoteImportStorageSetup(
            tmp_dir="/tmp/%s" % name_underscore,
            remote=remote,
            local=[local],
        ),
        tag=name_underscore,
    )
    tables[name] = table

snap_dataset = DataSet(
    name="snap",
    description="The Snap DataSet",
    sourcePath=__file__,
    datumTemplates=[edge_tuple],
    assets=tables,
    tag="snap",
)

What if we want to do Machine Learning?

As a proof-of-concept, ML models are not substantively different from tabular-based data assets. Here's an example for how we can declare the existence of an SVM regression model trained on the wine table:

# We will train a classifier and store it in a local file.
classifier_storage = LocalFileStorage(
    location=MinioLocation(name="wine"),
    layout=SingleFileLayout(),
    encoding=ONNXEncoding(),
)
# We will use these as the features in our classifier.
features = attributes[2:10]
# This is the "recipe" for our classifier.
classifier_template = TrainedFloatMeasure(
    name="predicted_alcohol",
    comment="""
    Predicted alcohol content, based on the following inputs:
    %s
    """ % [x.name for x in features],
    features=features,
    objective=attributes[1],
    source_asset_name="wine_table",
)
# We now augment the dataset with this recipe.
wine_dataset.add_template(classifier_template)
# The classifier is computed from local data
# (note the source_asset_names dictionary)
classifier_setup = ComputedFromLocalData(
    source_asset_names={"training_dataset": "wine_table"},
    target=classifier_storage,
    tmp_dir="/tmp/wine_classifier",
)
# We finally define our regression_model as a concrete
# data asset, following a recipe defined by the template,
# while also connected to concrete storage, as defined
# by classifier_setup
regression_model = SupervisedModel(
    name="wine_alcohol_predictor",
    tag="predictor",
    setup=classifier_setup,
    schema=classifier_template.get_model_storage_tabular_schema(),
    algorithm=SVMRegressionAlgorithm(),
)
wine_dataset.add_asset(regression_model)

Note the use of imperative directives such as wine_dataset.add_asset. This is a small compromise on our mostly-declarative syntax, but it maps well on the following thought pattern common to ML models:

  • we have some "primary sources", datasets external to the project,
  • we then derive other data assets by building, iteratively on the primary sources.

The common development cycle, therefore, is one where, after the original data sources are imported, we add new templates and assets to our dataset, fine-tuning Python code by first running it in Jupyter, then in Native python, then as an Airflow task, etc.

Also note that while currently Aorist only supports generating single files as DAGs, in the future we expect it will support multiple file generation for complex projects.

SQL and derived assets

Especially when datasets are in tabular form, it makes sense to think of data transformations in terms of standard SQL operations -- selections, projections, groups, explodes, and joins. These transformations can be supported via a derive_asset directive used in the process of Universe creation. For instance, if we are interested in training a model for high-ABV wines only, we can write:

universe.derive_asset(
    """
    SELECT *
    FROM wine.wine_table
    WHERE wine.wine_table.alcohol > 14.0
    """,
    name="high_abv_wines",
    storage=HiveTableStorage(
        location=MinioLocation(name="high_abv_wines"),
        layout=StaticHiveTableLayout(),
        encoding=ORCEncoding(),
    ),
    tmp_dir="/tmp/high_abv_wines",
)

Behind the scenes, this directive does two things:

  • if necessary, creates a new template expressing the operation of filtering a table on the alcohol attribute.
  • it creates a new StaticDataTable asset living in the indicated storage. This table will only be computed after its source tables (the ones in the FROM clause) are ready.

How to build

To build cargo library (need Rust installed):

cargo build

To try out Python code against .so library:

./run.sh

To rebuild pip wheel (requires maturin):

maturin build

Internals

Aorist's "secret sauce" is a Rust core. Even though we usually interact with Aorist via Python, deep down we rely on Rust's efficiency and type-safety to help us deal with the impressive complexity of data tasks. The following notes deal with some of the core concepts in Aorist. They are still Work-in-Progress.

Concepts

Aorist uses two types of concepts:

  • abstract concepts (e.g. a "location")
  • concrete ones (e.g. "a Google Cloud Storage location", or GCSLocation).

The relationship between abstract concepts represents the core semantic model offered by Aorist. This is not expected to change on a regular basis. Ideally this would not change at all.

Concrete concepts "instantiate" abstract ones, much like classes can instantiate traits or interfaces in OOP languages (in fact, this is how concrete concepts are implemented in Aorist).

Abstract concepts have the following hierarchy:

  • Universe: the core Aorist abstraction, one per project
    • DataSet: a grouping of instantiated objects which have inter-related schemas
    • User: someone accessing the objects
    • Group: a group of users
    • Roles: a way in which a user may access data.
    • RoleBindings: a connection between users and roles

Here is the current hierarchy of Aorist concepts:

Hierarchy of Aorist Concepts

Constraints

A constraint is a fact that can be verified about a concept. A constraint may have dependent constraints. For instance, we may have the constraint "is consistent" on Universe, that further breaks down into:

  • "datasets are replicated",
  • "users are instantiated",
  • "role bindings are created",
  • "data access policies are enforced".

Dependent constraints simply tell us what needs to hold, in order for a constraint to be true. Dependent constraints may be defined on the same concept, on dependent concepts, or on higher-order concepts, but they may not create a cycle. So we cannot say that constraint A depends on B, B depends on C, and C depends on A.

This is quite dry stuff. Here is a diagram of an example set of constraints to help better visualize what's going on:

Hierarchy of Aorist Concepts and Constraints

When dependent constraints are defined on lower-order concepts, we will consider the dependency to be satisfied when ALL constraints of the dependent kind associated with the lower-order concepts directly inheriting from the constrained concept have been satisfied.

For instance we may say that a constraint placed on the Universe (our abstraction for a Data Warehouse or Data Lake), of the kind: "no columns contain PII" is to be satisfied when all columns in ALL the tables are confirmed to not contain any PII.

When dependent constraints are defined on higher-order concepts, we will consider the dependency to be satisfied when the dependent constraint placed on the exact higher-order ancestor has been satisfied.

So for instance, a model trained on data from the data warehouse may be publishable on the web if we can confirm that no data in the warehouse whatsoever contains any PII. This is a very strict guarantee, but it is logically correct -- if there is no PII in the warehouse, there can be no PII in the model. This is why we could have a constraint at the model-level that depends on the Universe-level "no PII" constraint.

Constraint DAG generation

Both constraints and concept operate at a very abstract level. They are basic semantic building blocks of how we understand the things we care about in our data streams. But our YAML file will define ``instances'' of concepts, i.e., Aorist objects. StaticDataTable is a concept, but we may have 200 static data tables, on which we would like to impose the same constraints. For instance, we would like all these tables to be audited, etc.[1]

Looking back at the concept hierarchy mentioned above, we turn the constraint DAG into the prototype of an ETL pipeline by "walking" both the concept (black) and constraint (red) part of the DAG.

Here's what the Constraint DAG looks like

Constraint DAG

[1] (NOTE: in the future we will support filters on constraints, but for now assume that all constraints must hold for all instances).

Some things to note about this DAG:

  • it includes some superfluous dependencies, such as the one between DownloadDataFromRemote and ReplicatedData
  • some constraints are purely "cosmetic" -- DataFromRemoteDownloaded is really just a wrapper around DownloadDataFromRemote that "elevates" it to the root level, so that UploadDataToLocal can depend on it.

Programs

Constraints are satisfiable in two stages:

  1. First, any dependent constraints must be satisfied.
  2. Then, a program associated with the constraint must run successfully.

The program is where the actual data manipulation happens. Examples of programs are: "move this data from A to B", or "train this model", or "anonymize this data," etc. The programs are written as templates, with full access to the instantiated object hierarchy.

A program is written in a "dialect" that encompases what is considered to be valid code. For instance, "Python3 with numpy and PyTorch" would be a dialect. For Python dialects, we may attach a conda requirements.txt file, or a Docker image to the dialect, etc. For R dialects we may attach a list of libraries and an R version, or a Docker image.

Drivers

Note that multiple programs may exist that could technically satisfy a constraint. A driver decides which program to apply (given a preference ordering) and is responsible for instantiating it into valid code that will run in the specific deployment. A driver could, for instance, be responsible for translating the constraint graph into valid Airflow code that will run in a particular data deployment, etc.

Testing

cargo test --no-default-features

Note: make sure that libpython is in your LD_LIBRARY_PATH. E.g.:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/bogdan/anaconda3/lib/

Dependencies

~15–21MB
~339K SLoC