Add a cmake find module to locate cuFile. If found, add the include directory and link to the shared library.

This shouldn't have any effect if cuFile is not installed locally.

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

[gpuCI] Auto-merge branch-0.17 to branch-0.18 [skip ci]

)

This implements the `non_numeric` argument for `DataFrame.quantile` meaning that it now works on `datetime` and `timedelta` data. However, because of the difference in how `DataFrame.iloc` behaves between Pandas and cuDF, this implementation returns a DataFrame when `non_numeric=False` even when Pandas returns a Series

Passes tests locally

This closes #6799

Authors:
  - Chris Jarrett <cjarrett@dt08.aselab.nvidia.com>
  - ChrisJar <chris.jarrett.0@gmail.com>

Approvers:
  - Keith Kraus

URL: #6902

When using parameter `--rmm_mode=managed` for gtests `Invalid RMM allocation mode: managed` exception is thrown.
The logic in `include/cudf_test/base_fixture.hpp` is just missing a return statement.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - Paul Taylor
  - Mark Harris

URL: #6912

Resolves: #6870 

This PR adds support for `set_names` API in both `Index` & `MultiIndex`.

Authors:
  - galipremsagar <sagarprem75@gmail.com>
  - GALI PREM SAGAR <sagarprem75@gmail.com>

Approvers:
  - Keith Kraus

URL: #6929

Fixes: #6821 

This PR fixes issue where `columns` and `index` are currently not being handled correctly in specific scenarios.

Authors:
  - galipremsagar <sagarprem75@gmail.com>
  - GALI PREM SAGAR <sagarprem75@gmail.com>

Approvers:
  - Richard (Rick) Zamora
  - Ashwin Srinath

URL: #6838

)

Update to libcu++ on Github.

Authors:
  - ptaylor <paul.e.taylor@me.com>
  - Paul Taylor <paul.e.taylor@me.com>

Approvers:
  - Mark Harris
  - Keith Kraus
  - Christopher Harris
  - Mark Harris

URL: #6275

This PR removes `**kwargs` from the string/categorical accessors where unnecessary, and exposes keyword arguments like `inplace` to the user directly.

If we want to maintain parity with Pandas APIs for Dask/others using cuDF internally, we can consider using the approach described in #6135, which will automatically raise `NotimplementedError` when unsupported kwargs are passed.

Authors:
  - Ashwin Srinath <shwina@users.noreply.github.com>

Approvers:
  - GALI PREM SAGAR
  - Keith Kraus
  - Keith Kraus

URL: #6750

Fixes #6682, #6680 

Currently, empty fields are treated as N/A regardless on parsing options. However, the desired behavior is to handle empty fields the same way as fields with special values (apply default_na_values, na_filter logic). 
This PR irons out the behavior so it matches Pandas in this regard.

- Tries now support matching empty strings.
- The list of special NA values is now generated more robustly, so it has correct elements in any parameter combination.
- Empty string is added to the list of special NA values.
- Empty string string ("/"/"") is added to NA value list if empty string ("") is included (mirrors Pandas behavior).
- Added tests for previously failing parameter combinations.
- Reworked some of the tests to check against Pandas results instead of assumed desired behavior.

Authors:
  - vuule <vmilovanovic@nvidia.com>
  - vuule <vukasin.milovanovic.87@gmail.com>
  - Vukasin Milovanovic <vukasin.milovanovic.87@gmail.com>
  - Vukasin Milovanovic <vmilovanovic@nvidia.com>

Approvers:
  - Ram (Ramakrishna Prabhu)
  - Christopher Harris
  - Keith Kraus

URL: #6922

The include directory was renamed from `simt` to `cuda`.

Authors:
  - Rong Ou <rong.ou@gmail.com>

Approvers:
  - Jason Lowe

URL: #6948

The `cudf::merge` API expects the key columns to be sorted. This means that if null rows are included, these null entries should all appear either at beginning or at the end of the column depending on the null_order for the sort. The `MergeDictionaryTest.WithNull` gtest placed null rows in the middle of the column. The expected results should also have included null entries at the beginning or the end.

This PR also includes an extra test for checking merge results are consistent with the sort parameters `cudf::order` and `cudf::null_order`. This test also includes a larger number of rows to ensure `thrust::merge` requires more than one tile/block in its runtime logic.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - Ram (Ramakrishna Prabhu)
  - Vukasin Milovanovic

URL: #6942

Updating the Java bindings package version to match the libcudf version.

Authors:
  - Jason Lowe <jlowe@nvidia.com>

Approvers:
  - Robert (Bobby) Evans

URL: #6949

This exposes `logical_cast` through a JNI API. I also updated some of the test code to take a ColumnView instead of a ColumnVector so I could test it more easily.

Authors:
  - Robert (Bobby) Evans <bobby@apache.org>

Approvers:
  - Jason Lowe

URL: #6954

Updates libcudf to use the new, simplified `rmm::exec_policy` and include the new refactored headers `rmm/exec_policy.hpp` and `rmm/device_vector.hpp`

The new `exec_policy` can be passed directly to Thrust, no longer any need to call `rmm::exec_policy(stream)->on(stream)`.

Depends on rapidsai/rmm#647

As a part of trying to support upper and lower bounds for decimal I found that type checking for this function was broken because it used `==` for equality instead of `.equals`.  Looking further I found a few other places where this was a bug (one in ColumnVector that is mostly a performance issue and one in Scalar)  I decided to update all of the code to use .equals for comparison of types to make it consistent so it is less likely to have bugs like this crop up in the future.

I also took the opportunity to internally move away from using `isTimestamp` (which is deprecated) to `isTimestampType`

Authors:
  - Robert (Bobby) Evans <bobby@apache.org>

Approvers:
  - Jason Lowe

URL: #6970

Adding Java bindings for the `url_decode` and `url_encode` functions.

Authors:
  - Jason Lowe <jlowe@nvidia.com>

Approvers:
  - Robert (Bobby) Evans
  - Kuhu Shukla

URL: #6972

This adds a `libcufilejni.so` that's by default not built nor loaded. The unit tests are controlled similarly.

Tested locally with the corresponding spark-rapids plugin changes.

@jlowe @revans2 @abellina

Authors:
  - Rong Ou <rong.ou@gmail.com>

Approvers:
  - Robert (Bobby) Evans

URL: #6940

Currently we're missing a few kwargs in `Series.groupby` which is causing issues due to the dask change in dask/dask#6854

Adds the missing kwargs and validates that we support the values passed in.

Authors:
  - Keith Kraus <keith.j.kraus@gmail.com>

Approvers:
  - GALI PREM SAGAR
  - Michael Wang

URL: #6964

closes #6778 

Fixes typo missed in PR #6887, the else condition where this is situated would be the last resort for any unaccounted scalar type values.

Authors:
  - Ramakrishna Prabhu <ramakrishnap@nvidia.com>

Approvers:
  - Keith Kraus

URL: #6957

More Pandas-like behaviour for groupby when no keys are passed.

Possibly fixes #6927.

Authors:
  - Ashwin Srinath <shwina@users.noreply.github.com>

Approvers:
  - Keith Kraus

URL: #6945

… no-op(#6975)

@nartal1 found a small bug while working on: NVIDIA/spark-rapids#1244

Problem is that for `fixed_point`, when the column `scale = -decimal_places`, it should be a no-op. Fix is to make it a no-op.

Authors:
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - David
  - Karthikeyan

URL: #6975

array_tests.cu with same content in same directory exists.
(can't convert to .cpp because .cuh is included in array_tests.cu and template source code is tested)
Also, array_tests.cpp is not referred in `cpp/tests/CMakeLists.txt`

Authors:
  - Karthikeyan <6488848+karthikeyann@users.noreply.github.com>

Approvers:
  - David
  - Vukasin Milovanovic

URL: #6953

Addresses groupby part of #2188
- [x] Add cython interfaces for aggregation argmin, argmax as idxmin, idxmax
- [x] unit tests

Authors:
  - Karthikeyan Natarajan <karthikeyann@users.noreply.github.com>
  - Karthikeyan <6488848+karthikeyann@users.noreply.github.com>

Approvers:
  - David
  - Ram (Ramakrishna Prabhu)
  - GALI PREM SAGAR
  - Jake Hemstad

URL: #6856

… column support(#6907)

Part 1 for issue #1361 

- Adds `PRECEDING` and `FOLLOWING` options to `replace_nulls` in `libcudf`. This PR provides support for `fixed_width_type` type columns.
- Adds Cython binding

Authors:
  - Michael Wang <michaelwang0905@gmail.com>
  - Michael Wang <isVoid@users.noreply.github.com>

Approvers:
  - Ashwin Srinath
  - Jake Hemstad
  - Mark Harris
  - Mark Harris

URL: #6907

This is a small cleanup that replaces a `cudf::binary_operation` with a much cleaner `cudf::cast`.

Authors:
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - Vukasin Milovanovic
  - Mark Harris

URL: #6976

Currently has changes for #6950 included.

The full set of null mask `fixed_point_column_wrapper` constructors aren't supported. This PR adds them all and also adds unit tests for each of them across difference `fixed_point` API tests.

**To Do List:**
* [x] Add constructors
* [x] Add basic unit test
* [x] Add all unit tests
* [x] Update docs

Authors:
  - Mark Harris <mharris@nvidia.com>
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - null
  - Vukasin Milovanovic

URL: #6951

Fixes #6671, #6851

- Set the `rows_per_chunk` in `csv_writer_options` to the size of the input table.
- Change `rows_per_chunk` type to `size_type` (used for number of rows).
- Set the default compression in `to_parquet`/`write_parquet` to "snappy".

Authors:
  - vuule <vmilovanovic@nvidia.com>

Approvers:
  - Keith Kraus
  - Conor Hoekstra
  - Ram (Ramakrishna Prabhu)
  - Mark Harris

URL: #6967

Just disables the tests when cufile is not installed.

Authors:
  - Robert (Bobby) Evans <bobby@apache.org>

Approvers:
  - Kuhu Shukla

URL: #6987

…ated as unequal(#6943)

This change mirrors what is done in `groupby` to eliminate null-containing columns from the join hash table if nulls not equal is set. This prevents absolute runaway of the process. I added benchmarks for joins with nulls and I can't even get it to finish without these changes. The 195ms test without nulls takes 2,000,000ms to complete and the larger tests I haven't had the patience to even see complete. With this change, the timings are faster than without nulls proportional to the % of nulls. Meaning half the table is nulls means the query is twice as fast as the non-null version, which makes sense.

closes #6052

Authors:
  - Mike Wilson <knobby@burntsheep.com>
  - Mike Wilson <hyperbolic2346@users.noreply.github.com>

Approvers:
  - Jake Hemstad
  - Jake Hemstad
  - null
  - Mark Harris

URL: #6943

…6959)

Fixes #6947 

When TZif file has no transitions (e.g. GMT), `build_timezone_transition_table` has an out-of-bounds read that leads to undefined behavior and intermittent issues.

This PR makes two changes to behavior:
1. When there are no transitions, the ancient rule is initialized from the first time offset (instead of the first transition rule, which does not exist in this case).
2. When there are no transitions and the time offset is zero, an empty table is returned (avoid using a no-op table in CUDA).

Authors:
  - vuule <vmilovanovic@nvidia.com>
  - Vukasin Milovanovic <vukasin.milovanovic.87@gmail.com>

Approvers:
  - GALI PREM SAGAR
  - null
  - Ram (Ramakrishna Prabhu)
  - David

URL: #6959

Fixes #6719

Authors:
  - Kumar Aatish <kaatish@nvidia.com>

Approvers:
  - David
  - GALI PREM SAGAR
  - Mark Harris

URL: #6991

Closes #6955 

This will improve the compile time and size for any function using `thrust::sort` and `thrust::stable_sort`. 
The PR includes a .patch file to be applied during cmake when downloading the thrust library.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - Mark Harris

URL: #6982

Fixes #6642

Authors:
  - Kumar Aatish <kaatish@nvidia.com>
  - skirui-source <71867292+skirui-source@users.noreply.github.com>

Approvers:
  - Vukasin Milovanovic
  - Devavret Makkar
  - Ram (Ramakrishna Prabhu)

URL: #6889

Allows for scalars of the same dtype as a column to be passed along a fast codepath to libcudf, instead of being inspected to reduce their dtype beforehand.

Authors:
  - brandon-b-miller <brmiller@nvidia.com>
  - GALI PREM SAGAR <sagarprem75@gmail.com>

Approvers:
  - GALI PREM SAGAR

URL: #6938

PR #6982 added a `PATCH_COMMAND` when fetching Thrust to remove unrolling in `thrust::sort`, thereby improving compile time and performance in some cases. But the command failed on local builds from source (At least on my machine under rapids-compose). 

This PR simplifies the command.

Authors:
  - Mark Harris <mharris@nvidia.com>

Approvers:
  - Keith Kraus

URL: #7002

Closes #6801 

This PR adds an extra reduce call in the libcudf gather specialization logic for strings column. This will check to make sure the output size of the gather does not exceed the size limit for the child characters column. The offsets column is first created with the individual output string sizes. Then the reduce call will add these sizes to check for overflow.

Also added a gtest to check for the overflow condition.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - Devavret Makkar
  - Karthikeyan

URL: #6997

I think this is how I started and verified it was working, but then I was trying to exclude the source as well, which didn't work for tests. Then I realized we need the source built for the plugin so remove it. Anyway, no presubmit CI check is really painful. :(

@revans2 

```console
$ mvn test
...
[WARNING] Tests run: 775, Failures: 0, Errors: 0, Skipped: 4

$ mvn test -DUSE-GDS=ON
...
[WARNING] Tests run: 777, Failures: 0, Errors: 0, Skipped: 4
```

Authors:
  - Rong Ou <rong.ou@gmail.com>

Approvers:
  - Robert (Bobby) Evans

URL: #6988

Fix #6926 .

Hi!

When invoking from_dlpack() and to_dlpack, the following warnings are displayed:

from_dlpack()
```
/opt/conda/envs/rapids/lib/python3.7/site-packages/cudf/io/dlpack.py:33: UserWarning: WARNING: cuDF from_dlpack() assumes column-major (Fortran order) input. If the input tensor is row-major, transpose it before passing it to this function.
res = libdlpack.from_dlpack(pycapsule_obj)
```

to_dlpack()
```
/opt/conda/envs/rapids/lib/python3.7/site-packages/cudf/io/dlpack.py:74: UserWarning: WARNING: cuDF to_dlpack() produces column-major (Fortran order) output. If the output tensor needs to be row major, transpose the output of this function.
return libdlpack.to_dlpack(gdf_cols)
```

I think those warnings should be removed, because it contains information that should be available in the API documentation, and not necessarily displayed each time the methods are invoked.

Some users, like me, love to have their notebooks/code without warnings. Even if it is possible to disable those warnings, I think the user should not go that way, because the warning is just repeating what the API documentation should cover.

Hope it helps!
Miguel

Authors:
  - Miguel Martínez <26169771+miguelusque@users.noreply.github.com>
  - GALI PREM SAGAR <sagarprem75@gmail.com>

Approvers:
  - GALI PREM SAGAR
  - Ram (Ramakrishna Prabhu)

URL: #7001

This PR adds a test to exercise the issue described in #6733. This issue was only reproduced on a laptop Pascal GPU, but I think it's a good test to have.  In summary, `copy_if`, used by `apply_boolean_mask` computes the output null count during as part of its custom scatter kernel, rather than using `cudf::count_unset_bits`.  #6733 describes an issue where the former is different from the latter. So it's good to have a test that verifies they get the same null count.

And since it's difficult to get a repro on a similar machine, this is a first step.

Authors:
  - Mark Harris <mharris@nvidia.com>
  - Keith Kraus <kkraus@nvidia.com>

Approvers:
  - Karthikeyan
  - Devavret Makkar

URL: #6903

#7002 attempted to fix the temporary Thrust sort patch introduced in #6982 which didn't work with CMake 3.19+.  

This PR updates the thirdparty CMakeLists.txt file to continue if the Thrust sort patch has already been applied.
Today, the first time cmake is run, the Thrust sort.h is patched. But if cmake is run again without cleaning the build directory, the build will fail, because the file has already been patched.  @trxcllnt showed us the correct `patch` incantation to ignore the patch if already applied.

CC @davidwendt

Authors:
  - Mark Harris <mharris@nvidia.com>

Approvers:
  - Paul Taylor
  - Keith Kraus

URL: #7009

This pull request is to address #6909.

Authors:
  - sperlingxx <lovedreamf@gmail.com>
  - Alfred Xu <lovedreamf@gmail.com>

Approvers:
  - Robert (Bobby) Evans
  - Mike Wilson
  - Devavret Makkar

URL: #6969

This resolves #6996

Authors:
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - Mike Wilson
  - Devavret Makkar

URL: #7006

Implements `cudf.DateOffset` - an object used for calendrical arithmetic, similar to pandas.DateOffset - for month units only. 

Closes #6754

Authors:
  - brandon-b-miller <brmiller@nvidia.com>
  - brandon-b-miller <53796099+brandon-b-miller@users.noreply.github.com>
  - Keith Kraus <kkraus@nvidia.com>

Approvers:
  - GALI PREM SAGAR
  - Keith Kraus
  - Keith Kraus

URL: #6775

Resolves: #6370 

This PR enables the parallel execution of pytests of `cudf`, `dask_cudf` & `custreamz` in CI. The changes also include adding `pytest-xdist` to dev environments.

With these changes, here is the change in pytest execution times in CI:

| module      | without pytest-xdist | with pytest-xdist(n=6)  |
| ----------- | ----------- | -----------|
| cudf      | 1 hr       |  14 min |
| dask_cudf   |    4 min     | 1 min |
| custreamz  |  6 min |  2 min |


Related Integration changes: rapidsai/integration#188

Authors:
  - galipremsagar <sagarprem75@gmail.com>

Approvers:
  - AJ Schmidt
  - Keith Kraus

URL: #6958

`librdkakfa` 1.5.3 required gcc 9.3 (https://anaconda.org/conda-forge/librdkafka/files?version=1.5.3) and `libcudf_kakfa` (https://anaconda.org/rapidsai-nightly/libcudf_kafka/files?version=0.18.0a201215) is being built requiring 1.5.3 which is not compatible with the rest of RAPIDS.

Authors:
  - Ray Douglass

Approvers:
  - Mike Wendt

* Fixes #6823
* Raise a `KeyError` similar to Pandas rather than an `IndexError` when loc fails
* Improve tests to compare more directly with Pandas behaviour

Authors:
  - Ashwin Srinath <shwina@users.noreply.github.com>

Approvers:
  - Ram (Ramakrishna Prabhu)
  - Michael Wang
  - GALI PREM SAGAR

URL: #6993

Found a small bug while working on NVIDIA/spark-rapids#1244. 
For negative integers, it was not rounding to nearest even number.

Authors:
  - Niranjan Artal <nartal@nvidia.com>
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - Conor Hoekstra
  - Mark Harris

URL: #7014

This PR resolves a part of #3556.

Supporting `cudf::reduce`:
1. Part 1 (`MIN`, `MAX`, `SUM` & `PRODUCT` & `NUNIQUE`) #6814
2. Part 2 (the rest) ◀️ 

**Reduction Ops:**

**Done in Previous PR**
✔️  `SUM,             ///< sum reduction`
✔️ `PRODUCT,         ///< product reduction`
✔️ `MIN,             ///< min reduction`
✔️ `MAX,             ///< max reduction`
✔️ `NUNIQUE,         ///< count number of unique elements`

**Not supported by `cudf::reduce`:**
* [x] `COUNT_VALID,     ///< count number of valid elements`
* [x] `COUNT_ALL,       ///< count number of elements`
* [x] `COLLECT,         ///< collect values into a list`
* [x] `LEAD,            ///< window function, accesses row at specified offset following current row`
* [x] `LAG,             ///< window function, accesses row at specified offset preceding current row`
* [x] `PTX,             ///< PTX UDF based reduction`
* [x] `CUDA             ///< CUDA UDf based reduction`
* [x] `ARGMAX,          ///< Index of max element`
* [x] `ARGMIN,          ///< Index of min element`
* [x] `ROW_NUMBER,      ///< get row-number of element`

**Won't be supported:**
* [x] `ANY,             ///< any reduction`
* [x] `ALL,             ///< all reduction`

**To Do / Investigate:**
* [x] `SUM_OF_SQUARES,  ///< sum of squares reduction`
* [x] `MEDIAN,          ///< median reduction`
* [x] `QUANTILE,        ///< compute specified quantile(s)`
* [x] `NTH_ELEMENT,     ///< get the nth element`

**Deferred until requested**
* [x] `MEAN,            ///< arithmetic mean reduction`
* [x] `VARIANCE,        ///< groupwise variance`
* [x] `STD,             ///< groupwise standard deviation`

Authors:
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - null
  - Karthikeyan
  - David

URL: #6980

Follow up for PR #6907 

- `replace_null` policy function now supports `string` and `dictionary` dtype column.

Since original implementation depends only on column validity and index, this extension trivially removes SFINAE on `replace_null` functor and removes `type_dispatcher`.

Authors:
  - Michael Wang <isVoid@users.noreply.github.com>

Approvers:
  - Mark Harris
  - Karthikeyan

URL: #7004

`pd.DateOffset` uses a metaclass that overrides the usual instance/subclass checking behaviour. Any subclass of `pd._libs.tslibs.offsets.BaseOffset` will be reported as a subclass of `pd.DateOffset` (itself a `pd.DateOffset`). This can lead to some surprising behaviour:

```python
In [3]: isinstance(pd.DateOffset(), cudf.DateOffset)
Out[3]: True
``` 

Note that `cudf.DateOffset` inherits from `pd.DateOffset`. But, a `pd.DateOffset` is reported as an instance of `cudf.DateOffset` -- [Child Is Father of the Man](https://en.wikipedia.org/wiki/Child_Is_Father_of_the_Man)!

Authors:
  - Ashwin Srinath <shwina@users.noreply.github.com>

Approvers:
  - GALI PREM SAGAR

URL: #7029

Closes #6850 

dask_cudf version of the `dask.dataframe` changes proposed in [dask#6960](dask/dask#6960).  Uses `fsspec` to infer the default `compression` argument from the suffix of the first file-path argument.

Authors:
  - rjzamora <rzamora217@gmail.com>

Approvers:
  - Keith Kraus

URL: #7013

Closes #6472.

`rolling.cu` is taking inordinately long to compile, slowing down the `libcudf` build. The following changes were made to mitigate this:

1. Moved `grouped_rolling_window()` and `grouped_time_based_rolling_window()` to `grouped_rolling.cu`. Common functions were moved to `rolling_detail.cuh`.
2. Normalized timestamp columns to use int64_t representations. This reduces the number of template instantiations for `time_based_grouped_rolling_window()`.
3. `grouped_*_rolling_window()` functions used to pass around fancy iterators, causing massive template instantiations. This has been changed to materialize the window offsets as separate columns, and use those with existing `rolling_window()` functions to produce the final result.

These changes have been tested by running a window function test from SparkSQL, over a 2.4GB ORC file with 155M records (1.5M groups of about 97 records each on average):
1. There has been no discernible change in the end-to-end runtime. (The `nsys` profile seems to indicate that the total time spent in the `gpu_rolling` kernel has reduced. This is still being examined, to confirm.)
2. Compiling `rolling.cu` and `grouped_rolling.cu` in parallel now takes 60s as opposed to about 300s before.
3. The object file size seems to have reduced by a factor of 3.

Authors:
  - Mithun RK <mythrocks@gmail.com>

Approvers:
  - Vukasin Milovanovic
  - Karthikeyan

URL: #6512

We compared the wrong thing on a cast optimization.  This fixes that.

Authors:
  - Robert (Bobby) Evans <bobby@apache.org>

Approvers:
  - Jason Lowe
  - Alessandro Bellina

URL: #7032

Closes #1361 

- Provides "`ffill`" and "`bfill`" `fillna` methods for `Numerical`, `Datetime`, `Timedelta` and `Categorical` type column.
- Supports `method` parameter for `Series.fillna` and `DataFrame.fillna`

Authors:
  - Michael Wang <isVoid@users.noreply.github.com>

Approvers:
  - Ashwin Srinath
  - GALI PREM SAGAR

URL: #6998

Fixes #6923

Included other minor cuIO improvements that are too small for individual PRs:
- Remove unnecessary NaN-related conditions in JSON, CSV.
- Expand a comment in `createSerializedTrie` to make initialization clearer.

Authors:
  - vuule <vmilovanovic@nvidia.com>
  - Vukasin Milovanovic <vukasin.milovanovic.87@gmail.com>

Approvers:
  - GALI PREM SAGAR
  - Karthikeyan
  - Christopher Harris

URL: #7012

…umn data(#7020)

I tried experimenting with changing the `cudf::size_type` to `int64_t` and found many, many places that assume `size_type` and `int32_t` (and `int`) are interchangeable. This PR attempts to fix some of the places where offsets column is created as INT32 but the column data is incorrectly referenced as `data<size_type>()` for example.
Also, this PR fixes some places that accepts/returns only int32_t (regex internal functions) or size_type (factories) which should be casted or accounted for.
This is not a full set of possible violations found but may help minimize future errors. No function has changed/added.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - Conor Hoekstra
  - Devavret Makkar

URL: #7020

This corrects an issue with the sampling range used when replacement=True. Before, it sampled the range 0 through `num_rows` meaning it could sample `num_rows` even though it's one position out of bounds. This caused sample to return values not present in the original DataFrame.

I also created exceptions for sampling on empty DataFrames that match pandas, as well as an exception for sampling when `axis=1` and `replace=True` as cudf does not support DataFrames with duplicate columns.

This closes #6532

Authors:
  - Chris Jarrett <cjarrett@dt08.aselab.nvidia.com>
  - Mark Harris <mharris@nvidia.com>
  - ChrisJar <chris.jarrett.0@gmail.com>

Approvers:
  - Keith Kraus
  - Mark Harris

URL: #6884

Follow up of PR #7004 

Adds `method` field to `fillna` method in string type column to support `ffill` and `bfill`.
Also involves a small change to a `datetime64` `ffill`, `bfill` test case to improve test robustness.

Authors:
  - Michael Wang <isVoid@users.noreply.github.com>

Approvers:
  - GALI PREM SAGAR

URL: #7036

The `run_pos`  which was being used was from data rather from secondary stream which was for scale, but resulted value was being used for secondary stream `scale`. The code change fixes that issue and also adds test case to cover the issue.

closes #7016

Authors:
  - Ramakrishna Prabhu <ramakrishnap@nvidia.com>

Approvers:
  - Vukasin Milovanovic
  - GALI PREM SAGAR
  - Devavret Makkar

URL: #7034

This PR closes #6921 by dispatching to appropriate cudf alias for numpy functions from the UFUNC_ALIASES dictionary :  
```python
 _UFUNC_ALIASES = {
    "power": "pow",
    "equal": "eq",
    "not_equal": "ne",
    "less": "lt",
    "less_equal": "le",
    "greater": "gt",
    "greater_equal": "ge",
    "absolute": "abs",
}
```

Authors:
  - Vibhu Jawa <vibhujawa@gmail.com>

Approvers:
  - Keith Kraus
  - null

URL: #6973

Fixes: #7025 

This PR:

1. Handles loading of pickle files which have been created with rangeIndex prior to introduction of `step` parameter support.
2. Introduces special-case handling of stringcolumn size where we were previously storing it as a pickled object.

Authors:
  - galipremsagar <sagarprem75@gmail.com>

Approvers:
  - Ram (Ramakrishna Prabhu)

URL: #7033

Share the implementation of `cudf.Series.factorize` with the `Index` class and the `cudf` module namespace. 

Closes #6871

Authors:
  - brandon-b-miller <brmiller@nvidia.com>
  - Keith Kraus <kkraus@nvidia.com>
  - brandon-b-miller <53796099+brandon-b-miller@users.noreply.github.com>

Approvers:
  - Ashwin Srinath
  - Keith Kraus

URL: #6885

Fixes: #6936 

This PR introduces changes to `MultiIndex.__repr__`, where the output is now more readable and easy to understand similar to that of pandas MultiIndex. Changes also include handling of `<NA>`, `nan` values and spacing issues around them.

Authors:
  - galipremsagar <sagarprem75@gmail.com>

Approvers:
  - null
  - Keith Kraus

URL: #6992

Making this PR since wrong formatting keeps getting propagated in new PRs and (sometimes) corrected in code review.

Changes:

- Ironed out the formatting of Doxygen comments to match the guidelines. 
- Removed the outdated file with formatting examples.

Authors:
  - vuule <vmilovanovic@nvidia.com>
  - Vukasin Milovanovic <vukasin.milovanovic.87@gmail.com>
  - vukasin <vmilovanovic@nvidia.com>

Approvers:
  - David
  - Karthikeyan

URL: #7041

…7050)

Fixes: #7046

This PR:

- [x] Updates all doc examples with new NA_REP(`<NA>`)
- [x] Fixes reference warnings during doc build.

Authors:
  - galipremsagar <sagarprem75@gmail.com>

Approvers:
  - Keith Kraus

URL: #7050

Improves performance of parquet reader on certain multi-GPU systems, which take a long time to allocate pinned memory, by reducing the number of `hostdevice_vector` allocations.

Closes #7049

Authors:
  - Devavret Makkar <dmakkar@nvidia.com>

Approvers:
  - null
  - Ram (Ramakrishna Prabhu)
  - Karthikeyan

URL: #7005

This PR resolves a part of #3556.

Aggregation ops supported:
* `MIN`
* `MAX`
* `COUNT` (both `null_policy` - `EX/INCLUDE`)
* `LEAD`
* `LAG`

**To Do List:**
* [x] Basic unit tests
* [x] Comprehensive unit tests
* [x] Implementation
* [x] Figure out which rolling ops to suppport

Authors:
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - Vukasin Milovanovic
  - Ram (Ramakrishna Prabhu)

URL: #7037

Reference #7027 and #5698

This adds a strings column to the current gbenchmark for sort. This will help measure improvements or changes over time to the column and strings comparator functions.

No code logic changed or added.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - Vukasin Milovanovic
  - Devavret Makkar
  - Keith Kraus

URL: #7040

Closes #6699 
The timestamp format(s) used by the CSV writer have the form `%Y-%m-%dT%H:%M:%SZ`. This means if the column delimiter `','` or the line delimiter `\n` is either `':'` or `'-'` then the timestamp string output could conflict with these delimiters. The current logic simply removed these delimiters from the format if they detected a conflicting column or line delimiter. For example, specifying a dash `'-'` as column delimiter caused the timestamp format to change to `%Y%m%d...` (the dash is removed). I admit this was kind of hacky and also made the output inconsistent with Pandas `to_csv()`.

It is easy enough to simply add double-quotes around the timestamp format to prevent these conflicts as well as make the output consistent. This PR fixes that logic.

Exception logic to check for a dash as column separator was also found in [csv.py](https://github.com/rapidsai/cudf/blob/8c1f01e1fd713d873cf3d943ab409f3e9efc48f8/python/cudf/cudf/io/csv.py#L139-L149), specifically citing issue 6699 in the exception message. Also, there was a pytest specifically created to check for this exception. The exception is removed and the pytest function updated in this PR as well.

Authors:
  - davidwendt <dwendt@nvidia.com>

Approvers:
  - GALI PREM SAGAR
  - Karthikeyan
  - null

URL: #7023

Adding support for `cudf::scan` for `decimal32` and `decimal64`. `cudf::scan` only supports 4 operations (sum, product, min and max) but the decimal types will only support `SUM`, `MAX` and `MIN`.

This PR resolves a part of #3556.

Authors:
  - Conor Hoekstra <codereport@outlook.com>

Approvers:
  - Jake Hemstad
  - Mark Harris

URL: #7063

Resolves #6863

Expands existing murmur3 hashing functionality to match Spark's murmur3 hashing algorithm by modifying tail processing for unaligned bytes and processing booleans as 32bit integers rather than singular bytes.

Authors:
  - Ryan Lee <ryanlee@nvidia.com>
  - rwlee <rwlee@users.noreply.github.com>

Approvers:
  - Jake Hemstad
  - null
  - Robert (Bobby) Evans
  - GALI PREM SAGAR

URL: #7024

This version is more friendly to ccache:
```console
ccache -C  # clear the cache
time mvn clean package -DskipTests

real	4m43.015s
user	11m18.426s
sys	0m21.891s

time mvn clean package -DskipTests  # everything is now cached

real	0m20.265s
user	0m45.810s
sys	0m3.670s
```

Not sure about the ABI flag, but leaving it in causes the .so to not load:
```console
/usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java: symbol lookup error: /tmp/nvcomp5478764208255606671.so: undefined symbol: _ZN6nvcomp5Check8not_nullEPKvRKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEESA_i
```

@jlowe

Authors:
  - Rong Ou <rong.ou@gmail.com>

Approvers:
  - Jason Lowe

URL: #7069