Incremental Batch Strategy

class onetl.strategy.incremental_strategy.IncrementalBatchStrategy(*, hwm: HWM | None = None, step: Any = None, start: Any = None, stop: Any = None, offset: Any = None)

Incremental batch strategy for DB Reader.

Note

Cannot be used with File Downloader

Same as IncrementalStrategy, but reads data from the source in sequential batches (1..N) like:

1: SELECT id, data
   FROM public.mydata
   WHERE id > 1000 AND id <= 1100; -- previous HWM value is 1000, step is 100

2: WHERE id > 1100 AND id <= 1200; -- + step
3: WHERE id > 1200 AND id <= 1200; -- + step
N: WHERE id > 1300 AND id <= 1400; -- until stop

This allows to use less CPU and RAM than reading all the data in the one batch, but takes proportionally more time.

Warning

Unlike SnapshotBatchStrategy, it saves current HWM value after each batch into HWM Store.

So if code inside the context manager raised an exception, like:

with IncrementalBatchStrategy() as batches:
    for _ in batches:
        df = reader.run()  # something went wrong here
        writer.run(df)  # or here
        # or here...

DBReader will NOT update HWM in HWM Store for the failed batch.

All of that allows to resume reading process from the last successful batch.

Warning

Not every DB connection supports batch strategy. For example, Kafka connection doesn’t support it. Make sure the connection you use is compatible with the IncrementalBatchStrategy.

Parameters:

stepAny

Step size used for generating batch SQL queries like:

SELECT id, data
FROM public.mydata
WHERE id > 1000 AND id <= 1100; -- 1000 is previous HWM value, step is 100

Note

Step defines a range of values will be fetched by each batch. This is not a number of rows, it depends on a table content and value distribution across the rows.

Note

step value will be added to the HWM, so it should have a proper type.

For example, for TIMESTAMP column step type should be datetime.timedelta, not int

stopAny, default: None

If passed, the value will be used for generating WHERE clauses with hwm.expression filter, as a stop value for the last batch.

If not set, the value is determined by a separated query:

SELECT MAX(id) as stop
FROM public.mydata
WHERE id > 1000; -- 1000 is previous HWM value (if any)

Note

stop should be the same type as hwm.expression value, e.g. datetime.datetime for TIMESTAMP column, datetime.date for DATE, and so on

offsetAny, default: None

If passed, the offset value will be used to read rows which appeared in the source after the previous read.

For example, previous incremental run returned rows:

898
899
900
1000

Current HWM value is 1000.

But since then few more rows appeared in the source:

898
899
900
901 # new
902 # new
...
999 # new
1000

and you need to read them too.

So you can set offset=100, so the first batch of a next incremental run will look like:

SELECT id, data
FROM public.mydata
WHERE id > 900 AND id <= 1000; -- 900 = 1000 - 100 = HWM - offset

and return rows from 901 (not 900) to 1000 (duplicate).

Warning

This can lead to reading duplicated values from the table. You probably need additional deduplication step to handle them

Note

offset value will be subtracted from the HWM, so it should have a proper type.

For example, for TIMESTAMP column offset type should be datetime.timedelta, not int

Examples

IncrementalBatch run:

from onetl.connection import Postgres, Hive
from onetl.db import DBReader
from onetl.strategy import IncrementalBatchStrategy

from pyspark.sql import SparkSession

maven_packages = Postgres.get_packages()
spark = (
    SparkSession.builder.appName("spark-app-name")
    .config("spark.jars.packages", ",".join(maven_packages))
    .getOrCreate()
)

postgres = Postgres(
    host="postgres.domain.com",
    user="myuser",
    password="*****",
    database="target_database",
    spark=spark,
)

hive = Hive(cluster="rnd-dwh", spark=spark)

reader = DBReader(
    connection=postgres,
    source="public.mydata",
    columns=["id", "data"],
    hwm=DBReader.AutoDetectHWM(name="some_hwm_name", expression="id"),
)

writer = DBWriter(connection=hive, target="newtable")

with IncrementalBatchStrategy(step=100) as batches:
    for _ in batches:
        df = reader.run()
        writer.run(df)

-- previous HWM value was 1000
-- each batch (1..N) will perform a query which return some part of input data

1:  SELECT id, data
    FROM public.mydata
    WHERE id > 1100 AND id <= 1200; --- from HWM to HWM+step (EXCLUDING first row)

2:  WHERE id > 1200 AND id <= 1300; -- + step
N:  WHERE id > 1300 AND id <= 1400; -- until max value of HWM column

IncrementalBatch run with stop value:

with IncrementalBatchStrategy(step=100, stop=2000) as batches:
    for _ in batches:
        df = reader.run()
        writer.run(df)

-- previous HWM value was 1000
-- each batch (1..N) will perform a query which return some part of input data

1:  SELECT id, data
    FROM public.mydata
    WHERE id > 1000 AND id <= 1100; --- from HWM to HWM+step (EXCLUDING first row)

2:  WHERE id > 1100 AND id <= 1200; -- + step
...
N:  WHERE id > 1900 AND id <= 2000; -- until stop

IncrementalBatch run with offset value:

with IncrementalBatchStrategy(step=100, offset=100) as batches:
    for _ in batches:
        df = reader.run()
        writer.run(df)

-- previous HWM value was 1000
-- each batch (1..N) will perform a query which return some part of input data

1:  SELECT id, data
    FROM public.mydata
    WHERE id >  900 AND id <= 1000; --- from HWM-offset to HWM-offset+step (EXCLUDING first row)

2:  WHERE id > 1000 AND id <= 1100; -- + step
3:  WHERE id > 1100 AND id <= 1200; -- + step
...
N:  WHERE id > 1300 AND id <= 1400; -- until max value of HWM column

IncrementalBatch run with all possible options:

with IncrementalBatchStrategy(
    step=100,
    stop=2000,
    offset=100,
) as batches:
    for _ in batches:
        df = reader.run()
        writer.run(df)

-- previous HWM value was 1000
-- each batch (1..N) will perform a query which return some part of input data

1:  SELECT id, data
    FROM public.mydata
    WHERE id > 900 AND id <= 1000; --- from HWM-offset to HWM-offset+step (EXCLUDING first row)

2:  WHERE id > 1000 AND id <= 1100; -- + step
3:  WHERE id > 1100 AND id <= 1200; -- + step
...
N:  WHERE id > 1900 AND id <= 2000; -- until stop

hwm.expression can be a date or datetime, not only integer:

from datetime import date, timedelta

reader = DBReader(
    connection=postgres,
    source="public.mydata",
    columns=["business_dt", "data"],
    hwm=DBReader.AutoDetectHWM(name="some_hwm_name", expression="business_dt"),
)

with IncrementalBatchStrategy(
    step=timedelta(days=5),
    stop=date("2021-01-31"),
    offset=timedelta(days=1),
) as batches:
    for _ in batches:
        df = reader.run()
        writer.run(df)

-- previous HWM value was '2021-01-10'
-- each batch (1..N) will perform a query which return some part of input data

1:  SELECT business_dt, data
    FROM public.mydata
    WHERE business_dt  > CAST('2021-01-09' AS DATE)  -- from HWM-offset (EXCLUDING first row)
    AND   business_dt <= CAST('2021-01-14' AS DATE); -- to HWM-offset+step

2:  WHERE business_dt  > CAST('2021-01-14' AS DATE) -- + step
    AND   business_dt <= CAST('2021-01-19' AS DATE);

3:  WHERE business_dt  > CAST('2021-01-19' AS DATE) -- + step
    AND   business_dt <= CAST('2021-01-24' AS DATE);

...

N:  WHERE business_dt  > CAST('2021-01-29' AS DATE)
    AND   business_dt <= CAST('2021-01-31' AS DATE); -- until stop

__init__(**kwargs)

Create a new model by parsing and validating input data from keyword arguments.

Raises ValidationError if the input data cannot be parsed to form a valid model.