hsfs.constructor.prediction_times #

PredictionTimes #

The timestamps a batch-inference read is anchored on.

Each timestamp becomes the prediction time of one row per entity: the feature values returned for it are the newest at or before it in every feature group of the view. The timestamps may be in the future, which is the point of the class.

Build one with cron, every, or of.

Example

entities = pd.DataFrame([{"country": "SE", "city": "Stockholm", "street": "Sveavagen"}])
schedule = PredictionTimes.every("daily", offset="08:00", count=7)
fv.get_batch_data(spine_df=schedule.cross(entities, event_time="date"))

timestamps `property` #

timestamps: list[datetime]

The resolved prediction times, timezone-aware in UTC, ascending and unique.

cron `classmethod` #

cron(
    expression: str,
    start: date | datetime | str | int | None = None,
    end: date | datetime | str | int | None = None,
    count: int | None = None,
    timezone: str = "UTC",
    max_horizon_days: int = DEFAULT_MAX_HORIZON_DAYS,
) -> PredictionTimes

Prediction times from a five-field cron expression.

The fields are minute hour day-of-month month day-of-week, in the Vixie cron dialect: *, an integer, a-b, a-b/n, */n, comma-separated lists, and the three-letter month and weekday names. Both 0 and 7 mean Sunday. When both day-of-month and day-of-week are restricted, a day matches if either matches. This is not the Quartz dialect used by Hopsworks job schedules, which has a seconds field and numbers Sunday as 1.

start is inclusive and defaults to now truncated to the minute; end is exclusive. Exactly one of end and count is required. A naive datetime and a date are read in timezone; an aware one is converted from its own. Expansion happens in timezone, so a daily schedule keeps its local time across a DST change.

Example

PredictionTimes.cron("0 8 * * MON-FRI", count=5)

PARAMETER	DESCRIPTION
`expression`	The cron expression. TYPE: `str`
`start`	First instant the schedule may fire at, inclusive. TYPE: `date \| datetime \| str \| int \| None` DEFAULT: `None`
`end`	Instant the schedule stops before, exclusive. TYPE: `date \| datetime \| str \| int \| None` DEFAULT: `None`
`count`	How many timestamps to produce. TYPE: `int \| None` DEFAULT: `None`
`timezone`	IANA time zone name the schedule is expressed in. TYPE: `str` DEFAULT: `'UTC'`
`max_horizon_days`	How far past `start` to search before giving up on `count`. TYPE: `int` DEFAULT: `DEFAULT_MAX_HORIZON_DAYS`

RETURNS	DESCRIPTION
`PredictionTimes`	The resolved prediction times.

RAISES	DESCRIPTION
`ValueError`	If the expression is malformed, neither or both of `end` and `count` are given, or the schedule yields fewer than `count` timestamps within the horizon.

cross #

cross(spine_df: Any, event_time: str) -> Any

Cross a frame of entities with these times, one row per entity per time.

The ordering is the contract: entities in the order given, and within an entity ascending in time. A batch read returns its rows in that same order, so predictions zip back onto the frame positionally.

PARAMETER	DESCRIPTION
`spine_df`	The entities, one row each. TYPE: `Any`
`event_time`	Name of the feature view's event time column, which the result carries the times under. TYPE: `str`

RETURNS	DESCRIPTION
`Any`	A pandas DataFrame ready to pass as `spine_df`.

every `classmethod` #

every(
    interval: str,
    offset: str | None = None,
    start: date | datetime | str | int | None = None,
    end: date | datetime | str | int | None = None,
    count: int | None = None,
    timezone: str = "UTC",
    max_horizon_days: int = DEFAULT_MAX_HORIZON_DAYS,
) -> PredictionTimes

Prediction times from a named interval, which compiles to a cron expression.

interval is one of "hourly", "daily", "weekly" or "monthly". offset places the firing within the period and defaults to its start: a minute ("15") for hourly, "HH:MM" for daily, "DDD:HH:MM" with a three-letter weekday for weekly, and "D:HH:MM" with a day of the month for monthly. Every other parameter behaves as in cron.

Example

PredictionTimes.every("daily", offset="08:00", start=tomorrow, count=7)

PARAMETER	DESCRIPTION
`interval`	The named interval. TYPE: `str`
`offset`	Where in the period the schedule fires. TYPE: `str \| None` DEFAULT: `None`
`start`	First instant the schedule may fire at, inclusive. TYPE: `date \| datetime \| str \| int \| None` DEFAULT: `None`
`end`	Instant the schedule stops before, exclusive. TYPE: `date \| datetime \| str \| int \| None` DEFAULT: `None`
`count`	How many timestamps to produce. TYPE: `int \| None` DEFAULT: `None`
`timezone`	IANA time zone name the schedule is expressed in. TYPE: `str` DEFAULT: `'UTC'`
`max_horizon_days`	How far past `start` to search before giving up on `count`. TYPE: `int` DEFAULT: `DEFAULT_MAX_HORIZON_DAYS`

RETURNS	DESCRIPTION
`PredictionTimes`	The resolved prediction times.

RAISES	DESCRIPTION
`ValueError`	If the interval or the offset is not recognized.

of `classmethod` #

of(
    timestamps: list[date | datetime | str | int],
    timezone: str = "UTC",
) -> PredictionTimes

Prediction times from an explicit list.

Accepts datetime, date, ISO-8601 strings and epoch milliseconds, the same unit an integer column of spine_df is read in. Naive values are read in timezone, at the precision given: an instant keeps its seconds. Use this for anything a schedule cannot express.

Example

PredictionTimes.of([datetime(2026, 9, 14, 8), datetime(2026, 9, 21, 8)])

PARAMETER	DESCRIPTION
`timestamps`	The prediction times. TYPE: `list[date \| datetime \| str \| int]`
`timezone`	IANA time zone name naive values are read in. TYPE: `str` DEFAULT: `'UTC'`

RETURNS	DESCRIPTION
`PredictionTimes`	The resolved prediction times, sorted ascending and deduplicated.

RAISES	DESCRIPTION
`ValueError`	If the list is empty or contains a null.

hsfs.constructor.prediction_times #

PredictionTimes #

timestamps property #

cron classmethod #

cross #

every classmethod #

of classmethod #

timestamps `property` #

cron `classmethod` #

every `classmethod` #

of `classmethod` #