Ocean Runner is a package that eases algorithm creation in the scope of OceanProtocol.
pip install ocean-runner
# or
uv add ocean-runnerimport random
from ocean_runner import Algorithm, EmptyAlgorithm
algorithm: EmptyAlgorithm[int] = Algorithm.create(None)
@algorithm.run
def run(_) -> int:
return random.randint(0, 100)This code snippet will:
- Read the OceanProtocol JobDetails from the environment variables and use default configuration file paths.
- Execute the default input validation function, assessing if there are input dids and ddos.
- Execute the run function.
- Execute the default saving function, storing the result in a "result.txt" file within the default outputs path.
If you do not care about static analysis tools, this snippet will run just fine.
import random
from ocean_runner import Algorithm
algorithm = Algorithm.create(None)
@algorithm.run
def run(_):
return random.randint(0, 100)Having defined an algorithm as in the previous steps, you will now run it. To do so, you need the directory structure as in the data directory. This structure replicates the production environment, reducing the room for errors when running the deployed algorithm.
To run locally, or in the Dockerfile, you only have to run:
ocean-executeThis command will:
- Load the
base_dirthat you pass it via arguments (ocean-execute src.algorithm --base-dir ../_data), which represents the root of the data directory structure in OceanProtocol. - Load the algorithm instance from the module you pass
src.algorithmand run it.
You can see more information running it with --help.
The executable defaults are prepared to run from within the algorithm directory in the ocean-algo templates.
This executable makes the developer not need to make use of the docker-compose that emulates this behaviour, making the development cycle much faster and friendly, being able to debug the code more easily. Even not needing the docker-compose it's still a good tool to test your final Dockerfile image, since it will be the one that will run in the production environment.
As in the previous step, we also added a script to execute the tests without having to resort to the docker-compose. Right now it only supports pytest. To run it:
ocean-testAs in the ocean-execute, it accepts the module and base_dir arguments, and also arguments to pass to pytest after -- as in:
ocean-test -- -vvThe application configuration can be tweaked by passing a Config instance to its constructor.
from ocean_runner import Algorithm, Config
algorithm = Algorithm.create(
Config(
custom_input: ... # dataclass
# Custom algorithm parameters dataclass.
logger: ... # type: logging.Logger
# Custom logger to use.
source_paths: ... # type: Iterable[Path]
# Source paths to include in the PATH
environment: ...
# type: ocean_runner.Environment. Mock of environment variables.
)
)import logging
from pydantic import BaseModel
from ocean_runner import Algorithm, Config
class CustomInput(BaseModel):
foobar: string
logger = logging.getLogger(__name__)
algorithm = Algorithm.create(
Config(
custom_input=CustomInput,
"""
Load the Algorithm's Custom Input into a CustomInput instance.
"""
source_paths=[Path("/algorithm/src")],
"""
Source paths to include in the PATH. '/algorithm/src' is the default since our templates place the algorithm source files there.
"""
logger=logger,
"""
Custom logger to use in the Algorithm.
"""
environment=Environment(
base_dir: "./_data",
"""
Custom data path to use test data.
"""
dids: '["17feb697190d9f5912e064307006c06019c766d35e4e3f239ebb69fb71096e42"]',
"""
Dataset DID.
"""
transformation_did: "1234",
"""
Random transformation DID to use while testing.
"""
secret: "1234",
"""
Random secret to use while testing.
"""
)
"""
Should not be needed in production algorithms, used to mock environment variables, defaults to using env.
"""
)
)To fully configure the behaviour of the algorithm as in the Minimal Example, you can do it decorating your defined function as in the following fully annotated example (Pyton >=3.12), which features all the possible algorithm customization.
from pathlib import Path
from typing import Sequence, Tuple
import pandas as pd
from oceanprotocol_job_details.domain import DID
from ocean_runner import Algorithm
from ocean_runner.runner import EmptyAlgorithm
type ResultT = Tuple[DID, pd.DataFrame]
type ResultsT = Sequence[ResultT]
algorithm: EmptyAlgorithm[ResultsT] = Algorithm.create(None)
@algorithm.on_error
def error_callback(_, ex: Exception):
algorithm.logger.exception(ex)
raise algorithm.Error() from ex
@algorithm.validate
def val(_) -> None:
assert algorithm.job_details.files, "Empty input dir"
@algorithm.run
def run(_) -> ResultsT:
def describe(df: pd.DataFrame) -> pd.DataFrame:
return df.describe(include="all")
return [
(did, describe(pd.read_csv(file_path)))
for did, file_path in algorithm.job_details.inputs()
]
@algorithm.save_results
def save(_, result: ResultsT, base: Path):
for did, analysis in result:
algorithm.logger.info(f"Descriptive statistics {did}: {result}")
analysis.to_csv(base / f"{did}.csv")As seen in the minimal example, all methods implemented in Algorithm have a default implementation which will be commented here.
.validate()
"""
Will validate the algorithm's job detail instance, checking for the existence of:
- `job_details.ddos`
- `job_details.files`
"""
.run()
"""
Has NO default implementation, must pass a callback that returns a result of any type.
"""
.save_results()
"""
Stores the result of running the algorithm in "outputs/results.txt"
"""To load the OceanProtocol JobDetails instance, the program will read some environment variables, they can be mocked passing an instance of Environment through the configuration of the algorithm.
Environment variables:
DIDS(optional) Input dataset(s) DID's, must have format:["abc..90"]. Defaults to reading them automatically from theDDOdata directory.TRANSFORMATION_DID(optional, default="DEFAULT"): Algorithm DID, must have format:abc..90.SECRET(optional, default="DEFAULT"): Algorithm secret.BASE_DIR(optional, default="/data"): Base path to the OceanProtocol data directories.