### What changes were proposed in this pull request?
This PR targets to document the Pandas UDF redesign with type hints introduced at SPARK-28264.
Mostly self-describing; however, there are few things to note for reviewers.
1. This PR replace the existing documentation of pandas UDFs to the newer redesign to promote the Python type hints. I added some words that Spark 3.0 still keeps the compatibility though.
2. This PR proposes to name non-pandas UDFs as "Pandas Function API"
3. SCALAR_ITER become two separate sections to reduce confusion:
- `Iterator[pd.Series]` -> `Iterator[pd.Series]`
- `Iterator[Tuple[pd.Series, ...]]` -> `Iterator[pd.Series]`
4. I removed some examples that look overkill to me.
5. I also removed some information in the doc, that seems duplicating or too much.
### Why are the changes needed?
To document new redesign in pandas UDF.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existing tests should cover.
Closes#27466 from HyukjinKwon/SPARK-30722.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR proposes to redesign pandas UDFs as described in [the proposal](https://docs.google.com/document/d/1-kV0FS_LF2zvaRh_GhkV32Uqksm_Sq8SvnBBmRyxm30/edit?usp=sharing).
```python
from pyspark.sql.functions import pandas_udf
import pandas as pd
pandas_udf("long")
def plug_one(s: pd.Series) -> pd.Series:
return s + 1
spark.range(10).select(plug_one("id")).show()
```
```
+------------+
|plug_one(id)|
+------------+
| 1|
| 2|
| 3|
| 4|
| 5|
| 6|
| 7|
| 8|
| 9|
| 10|
+------------+
```
Note that, this PR address one of the future improvements described [here](https://docs.google.com/document/d/1-kV0FS_LF2zvaRh_GhkV32Uqksm_Sq8SvnBBmRyxm30/edit#heading=h.h3ncjpk6ujqu), "A couple of less-intuitive pandas UDF types" (by zero323) together.
In short,
- Adds new way with type hints as an alternative and experimental way.
```python
pandas_udf(schema='...')
def func(c1: Series, c2: Series) -> DataFrame:
pass
```
- Replace and/or add an alias for three types below from UDF, and make them as separate standalone APIs. So, `pandas_udf` is now consistent with regular `udf`s and other expressions.
`df.mapInPandas(udf)` -replace-> `df.mapInPandas(f, schema)`
`df.groupby.apply(udf)` -alias-> `df.groupby.applyInPandas(f, schema)`
`df.groupby.cogroup.apply(udf)` -replace-> `df.groupby.cogroup.applyInPandas(f, schema)`
*`df.groupby.apply` was added from 2.3 while the other were added in the master only.
- No deprecation for the existing ways for now.
```python
pandas_udf(schema='...', functionType=PandasUDFType.SCALAR)
def func(c1, c2):
pass
```
If users are happy with this, I plan to deprecate the existing way and declare using type hints is not experimental anymore.
One design goal in this PR was that, avoid touching the internal (since we didn't deprecate the old ways for now), but supports type hints with a minimised changes only at the interface.
- Once we deprecate or remove the old ways, I think it requires another refactoring for the internal in the future. At the very least, we should rename internal pandas evaluation types.
- If users find this experimental type hints isn't quite helpful, we should simply revert the changes at the interface level.
### Why are the changes needed?
In order to address old design issues. Please see [the proposal](https://docs.google.com/document/d/1-kV0FS_LF2zvaRh_GhkV32Uqksm_Sq8SvnBBmRyxm30/edit?usp=sharing).
### Does this PR introduce any user-facing change?
For behaviour changes, No.
It adds new ways to use pandas UDFs by using type hints. See below.
**SCALAR**:
```python
pandas_udf(schema='...')
def func(c1: Series, c2: DataFrame) -> Series:
pass # DataFrame represents a struct column
```
**SCALAR_ITER**:
```python
pandas_udf(schema='...')
def func(iter: Iterator[Tuple[Series, DataFrame, ...]]) -> Iterator[Series]:
pass # Same as SCALAR but wrapped by Iterator
```
**GROUPED_AGG**:
```python
pandas_udf(schema='...')
def func(c1: Series, c2: DataFrame) -> int:
pass # DataFrame represents a struct column
```
**GROUPED_MAP**:
This was added in Spark 2.3 as of SPARK-20396. As described above, it keeps the existing behaviour. Additionally, we now have a new alias `groupby.applyInPandas` for `groupby.apply`. See the example below:
```python
def func(pdf):
return pdf
df.groupby("...").applyInPandas(func, schema=df.schema)
```
**MAP_ITER**: this is not a pandas UDF anymore
This was added in Spark 3.0 as of SPARK-28198; and this PR replaces the usages. See the example below:
```python
def func(iter):
for df in iter:
yield df
df.mapInPandas(func, df.schema)
```
**COGROUPED_MAP**: this is not a pandas UDF anymore
This was added in Spark 3.0 as of SPARK-27463; and this PR replaces the usages. See the example below:
```python
def asof_join(left, right):
return pd.merge_asof(left, right, on="...", by="...")
df1.groupby("...").cogroup(df2.groupby("...")).applyInPandas(asof_join, schema="...")
```
### How was this patch tested?
Unittests added and tested against Python 2.7, 3.6 and 3.7.
Closes#27165 from HyukjinKwon/revisit-pandas.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR proposes to move pandas related functionalities into pandas package. Namely:
```bash
pyspark/sql/pandas
├── __init__.py
├── conversion.py # Conversion between pandas <> PySpark DataFrames
├── functions.py # pandas_udf
├── group_ops.py # Grouped UDF / Cogrouped UDF + groupby.apply, groupby.cogroup.apply
├── map_ops.py # Map Iter UDF + mapInPandas
├── serializers.py # pandas <> PyArrow serializers
├── types.py # Type utils between pandas <> PyArrow
└── utils.py # Version requirement checks
```
In order to separately locate `groupby.apply`, `groupby.cogroup.apply`, `mapInPandas`, `toPandas`, and `createDataFrame(pdf)` under `pandas` sub-package, I had to use a mix-in approach which Scala side uses often by `trait`, and also pandas itself uses this approach (see `IndexOpsMixin` as an example) to group related functionalities. Currently, you can think it's like Scala's self typed trait. See the structure below:
```python
class PandasMapOpsMixin(object):
def mapInPandas(self, ...):
...
return ...
# other Pandas <> PySpark APIs
```
```python
class DataFrame(PandasMapOpsMixin):
# other DataFrame APIs equivalent to Scala side.
```
Yes, This is a big PR but they are mostly just moving around except one case `createDataFrame` which I had to split the methods.
### Why are the changes needed?
There are pandas functionalities here and there and I myself gets lost where it was. Also, when you have to make a change commonly for all of pandas related features, it's almost impossible now.
Also, after this change, `DataFrame` and `SparkSession` become more consistent with Scala side since pandas is specific to Python, and this change separates pandas-specific APIs away from `DataFrame` or `SparkSession`.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existing tests should cover. Also, I manually built the PySpark API documentation and checked.
Closes#27109 from HyukjinKwon/pandas-refactoring.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
HyukjinKwon