Constructing schedules from partitioned jobs#

In this guide, we'll walk you through how to construct schedules from partitioned assets and jobs. By the end, you'll be able to:

• Construct a schedule for a time-partitioned job
• Customize a partitioned job's starting time
• Customize the most recent partition in a set
• Construct a schedule for a statically-partitioned job

Prerequisites#

To follow this guide, you need to be familiar with:

• Schedules
• Partitions
• Asset definitions
• Asset jobs and op jobs

Working with time-based partitions#

For jobs partitioned by time, you can use the build_schedule_from_partitioned_job to construct a schedule for the job. The schedule's interval will match the spacing of the partitions in the job. For example, if you have a daily partitioned job that fills in a date partition of a table each time it runs, you likely want to run that job every day.

Refer to the following tabs for examples of asset and op-based jobs using build_schedule_from_partitioned_job to construct schedules:

from dagster import (
    asset,
    build_schedule_from_partitioned_job,
    define_asset_job,
    DailyPartitionsDefinition,
)

daily_partition = DailyPartitionsDefinition(start_date="2024-05-20")


@asset(partitions_def=daily_partition)
def daily_asset(): ...


partitioned_asset_job = define_asset_job("partitioned_job", selection=[daily_asset])


asset_partitioned_schedule = build_schedule_from_partitioned_job(
    partitioned_asset_job,
)

from dagster import build_schedule_from_partitioned_job, job


@job(config=partitioned_config)
def partitioned_op_job(): ...


partitioned_op_schedule = build_schedule_from_partitioned_job(
    partitioned_op_job,
)

Customizing schedule timing#

The minute_of_hour, hour_of_day, day_of_week, and day_of_month parameters of build_schedule_from_partitioned_job can be used to control the timing of the schedule.

Consider the following job:

from dagster import build_schedule_from_partitioned_job

asset_partitioned_schedule = build_schedule_from_partitioned_job(
    partitioned_asset_job, hour_of_day=1, minute_of_hour=30
)

On May 20, 2024, the schedule will evaluate at 1:30 AM UTC and then start a run for the partition key of the previous day, 2024-05-19.

Customizing the ending partition in a set#

Each schedule tick of a partitioned job targets the latest partition in the partition set that exists as of the tick time. For example, consider a schedule that runs a daily-partitioned job. When the schedule runs on 2024-05-20, it will target the most recent partition, which will correspond to the previous day: 2024-05-19.

This occurs because each partition is a time window. A time window is a set period of time with a start and an end time. The partition's key is the start of the time window, but the partition isn't included in the partition set until its time window has completed. Kicking off a run after the time window completes allows the run to process data for the entire time window.

Continuing with the daily partition example, the 2024-05-20 partition would have the following start and end times:

• Start time - 2024-05-20 00:00:00
• End time - 2024-05-20 23:59:59

After 2024-05-20 23:59:59 passes, the time window is complete and Dagster will add a new 2024-05-20 partition to the partition set. At this point, the process will repeat with the next time window of 2024-05-21.

If you need to customize the ending, or most recent partition in a set, use the end_offset parameter in the partition's config:

from dagster import DailyPartitionsDefinition

daily_partition_with_offset = DailyPartitionsDefinition(
    start_date="2024-05-20", end_offset=-1
)

Setting this parameter changes the partition that will be filled in at each schedule tick. Positive and negative integers are accepted, which will have the following effects:

• Positive numbers, like 1, cause the schedule to fill in the partition of the current hour/day/week/month
• Negative numbers, like -1, cause the schedule to fill in the partition of an earlier hour/day/week/month

Generally, the calculation for end_offset can be expressed as:

current_date - 1 type_of_partition + end_offset

Let's look at an example schedule that's partitioned by day and how different end_offset values would affect the most recent partition in the set. In this example, we're using a start date of 2024-05-20:

Working with static partitions#

Next, we'll demonstrate how to create a schedule for a job with a static partition. To do this, we'll construct the schedule from scratch using the @schedule decorator, rather than using a helper function like build_schedule_from_partitioned_job. This will allow more flexibility in determining which partitions should be run by the schedule.

In this example, the job is partitioned by continent:

from dagster import (
    AssetExecutionContext,
    Config,
    asset,
    define_asset_job,
    static_partitioned_config,
)

CONTINENTS = [
    "Africa",
    "Antarctica",
    "Asia",
    "Europe",
    "North America",
    "Oceania",
    "South America",
]


@static_partitioned_config(partition_keys=CONTINENTS)
def continent_config(partition_key: str):
    return {"ops": {"continents": {"config": {"continent_name": partition_key}}}}


class ContinentOpConfig(Config):
    continent_name: str


@asset
def continents(context: AssetExecutionContext, config: ContinentOpConfig):
    context.log.info(config.continent_name)


continent_job = define_asset_job(
    name="continent_job", selection=[continents], config=continent_config
)

Using the @schedule decorator, we'll write a schedule that targets each partition, or continent:

from dagster import RunRequest, schedule


@schedule(cron_schedule="0 0 * * *", job=continent_job)
def continent_schedule():
    for c in CONTINENTS:
        yield RunRequest(run_key=c, partition_key=c)

If we only want to target the Antarctica partition, we can create a schedule like the following:

from dagster import RunRequest, schedule


@schedule(cron_schedule="0 0 * * *", job=continent_job)
def antarctica_schedule():
    return RunRequest(partition_key="Antarctica")

APIs in this guide#

Related#