Reading from Postgres using `DBReader`¶

DBReader supports strategy for incremental data reading, but does not support custom queries, like JOIN.

Warning

Please take into account Postgres types

Supported DBReader features¶

✅︎ columns
✅︎ where
✅︎ hwm, supported strategies:
- ✅︎ Snapshot strategy
- ✅︎ Incremental strategy
- ✅︎ Snapshot batch strategy
- ✅︎ Incremental batch strategy
❌ hint (is not supported by Postgres)
❌ df_schema
✅︎ options (see Postgres.ReadOptions)

Examples¶

Snapshot strategy:

from onetl.connection import Postgres
from onetl.db import DBReader

postgres = Postgres(...)

reader = DBReader(
    connection=postgres,
    source="schema.table",
    columns=["id", "key", "CAST(value AS text) value", "updated_dt"],
    where="key = 'something'",
    options=Postgres.ReadOptions(partitionColumn="id", numPartitions=10),
)
df = reader.run()

Incremental strategy:

from onetl.connection import Postgres
from onetl.db import DBReader
from onetl.strategy import IncrementalStrategy

postgres = Postgres(...)

reader = DBReader(
    connection=postgres,
    source="schema.table",
    columns=["id", "key", "CAST(value AS text) value", "updated_dt"],
    where="key = 'something'",
    hwm=DBReader.AutoDetectHWM(name="postgres_hwm", expression="updated_dt"),
    options=Postgres.ReadOptions(partitionColumn="id", numPartitions=10),
)

with IncrementalStrategy():
    df = reader.run()

Recommendations¶

Select only required columns¶

Instead of passing "*" in DBReader(columns=[...]) prefer passing exact column names. This reduces the amount of data passed from Postgres to Spark.

Pay attention to `where` value¶

Instead of filtering data on Spark side using df.filter(df.column == 'value') pass proper DBReader(where="column = 'value'") clause. This both reduces the amount of data send from Postgres to Spark, and may also improve performance of the query. Especially if there are indexes or partitions for columns used in where clause.

Options¶

`PostgresReadOptions` ¶

Spark JDBC reading options.

Added in 0.5.0

Replace Postgres.Options → Postgres.ReadOptions

Examples¶

Note

You can pass any value supported by Spark, even if it is not mentioned in this documentation. Option names should be in camelCase!

The set of supported options depends on Spark version.

from onetl.connection import Postgres

options = Postgres.ReadOptions(
    partitioning_mode="range",
    partitionColumn="reg_id",
    numPartitions=10,
    customSparkOption="value",
)

Parameters:

query_timeout ¶
(int | None, default: None ) –

The number of seconds the driver will wait for a statement to execute. Zero means there is no limit.

This option depends on driver implementation, some drivers can check the timeout of each query instead of an entire JDBC batch.
fetchsize ¶
(int, default: 100000 ) –

Fetch N rows from an opened cursor per one read round.

Tuning this option can influence performance of reading.

Warning

Default value is different from Spark.

Spark uses driver's own value, and it may be different in different drivers, and even versions of the same driver. For example, Oracle has default fetchsize=10, which is absolutely not usable.

Thus we've overridden default value with 100_000, which should increase reading performance.

Changed in 0.2.0

Set explicit default value to 100_000
partition_column ¶
(str | None, default: None ) –
Column used to parallelize reading from a table.

Warning

It is highly recommended to use primary key, or column with an index to avoid performance issues.
Note

Column type depends on partitioning_mode.
- partitioning_mode="range" requires column to be an integer, date or timestamp (can be NULL, but not recommended).
- partitioning_mode="hash" accepts any column type (NOT NULL).
- partitioning_mode="mod" requires column to be an integer (NOT NULL).
See documentation for partitioning_mode for more details
num_partitions ¶
(int, default: 1 ) –

Number of jobs created by Spark to read the table content in parallel. See documentation for partitioning_mode for more details
lower_bound ¶
(int | None, default: None ) –

See documentation for partitioning_mode for more details
upper_bound ¶
(int | None, default: None ) –

See documentation for partitioning_mode for more details
session_init_statement ¶
(str | None, default: None ) –
After each database session is opened to the remote DB and before starting to read data, this option executes a custom SQL statement (or a PL/SQL block).

Use this to implement session initialization code.

Example:
```
sessionInitStatement = """
    BEGIN
        execute immediate
        'alter session set "_serial_direct_read"=true';
    END;
"""
```

`partitioning_mode` ¶

(JDBCPartitioningMode, default: <JDBCPartitioningMode.RANGE: 'range'> ) –

Defines how Spark will parallelize reading from table.

Possible values:

range (default) Allocate each executor a range of values from column passed into partition_column.
Spark generates for each executor an SQL query

Executor 1:
```
SELECT ... FROM table
WHERE (partition_column >= lowerBound
        OR partition_column IS NULL)
AND partition_column < (lowerBound + stride)
```
Executor 2:
```
SELECT ... FROM table
WHERE partition_column >= (lowerBound + stride)
AND partition_column < (lowerBound + 2 * stride)
```
...

Executor N:
```
SELECT ... FROM table
WHERE partition_column >= (lowerBound + (N-1) * stride)
AND partition_column <= upperBound
```
Where stride=(upperBound - lowerBound) / numPartitions.
Column type must be integer, date or timestamp.

Note

lower_bound, upper_bound and num_partitions are used just to calculate the partition stride, NOT for filtering the rows in table. So all rows in the table will be returned (unlike Incremental strategy).

Note

All queries are executed in parallel. To execute them sequentially, use Batch strategy.
hash Allocate each executor a set of values based on hash of the partition_column column.
Spark generates for each executor an SQL query

Executor 1:
```
SELECT ... FROM table
WHERE (some_hash(partition_column) mod num_partitions) = 0 -- lower_bound
```
Executor 2:
```
SELECT ... FROM table
WHERE (some_hash(partition_column) mod num_partitions) = 1 -- lower_bound + 1
```
...

Executor N:
```
SELECT ... FROM table
WHERE (some_hash(partition_column) mod num_partitions) = num_partitions-1 -- upper_bound
```
Note

The hash function implementation depends on RDBMS. It can be MD5 or any other fast hash function, or expression based on this function call. Usually such functions accepts any column type as an input.

mod Allocate each executor a set of values based on modulus of the partition_column column.

Spark generates for each executor an SQL query

Executor 1:

SELECT ... FROM table
WHERE (partition_column mod num_partitions) = 0 -- lower_bound

Executor 2:

SELECT ... FROM table
WHERE (partition_column mod num_partitions) = 1 -- lower_bound + 1

Executor N:

SELECT ... FROM table
WHERE (partition_column mod num_partitions) = num_partitions-1 -- upper_bound

Note

Can be used only with columns of integer type.

Added in 0.5.0

Examples¶

Read data in 10 parallel jobs by range of values in id_column column:

ReadOptions(
    partitioning_mode="range",  # default mode, can be omitted
    partitionColumn="id_column",
    numPartitions=10,
    # Options below can be discarded because they are
    # calculated automatically as MIN and MAX values of `partitionColumn`
    lowerBound=0,
    upperBound=100_000,
)

Read data in 10 parallel jobs by hash of values in some_column column:

ReadOptions(
    partitioning_mode="hash",
    partitionColumn="some_column",
    numPartitions=10,
    # lowerBound and upperBound are automatically set to `0` and `9`
)

Read data in 10 parallel jobs by modulus of values in id_column column:

ReadOptions(
    partitioning_mode="mod",
    partitionColumn="id_column",
    numPartitions=10,
    # lowerBound and upperBound are automatically set to `0` and `9`
)

`parse(options)` ¶

If a parameter inherited from the ReadOptions class was passed, then it will be returned unchanged. If a Dict object was passed it will be converted to ReadOptions.

Otherwise, an exception will be raised

Reading from Postgres using `DBReader`¶

Supported DBReader features¶

Examples¶

Recommendations¶

Select only required columns¶

Pay attention to `where` value¶

Options¶

`PostgresReadOptions` ¶

Examples¶

`query_timeout` ¶

`fetchsize` ¶

`partition_column` ¶

`num_partitions` ¶

`lower_bound` ¶

`upper_bound` ¶

`session_init_statement` ¶

`partitioning_mode` ¶

Examples¶

`parse(options)` ¶

Reading from Postgres using DBReader¶

Supported DBReader features¶

Examples¶

Recommendations¶

Select only required columns¶

Pay attention to where value¶

Options¶

PostgresReadOptions ¶

Examples¶

query_timeout ¶

fetchsize ¶

partition_column ¶

num_partitions ¶

lower_bound ¶

upper_bound ¶

session_init_statement ¶

partitioning_mode ¶

Examples¶

parse(options) ¶

Reading from Postgres using `DBReader`¶

Pay attention to `where` value¶

`PostgresReadOptions` ¶

`query_timeout` ¶

`fetchsize` ¶

`partition_column` ¶

`num_partitions` ¶

`lower_bound` ¶

`upper_bound` ¶

`session_init_statement` ¶

`partitioning_mode` ¶

`parse(options)` ¶