[DOC] RAPIDS 22.12 Release Blog Outline

[caryr35]

Please provide bullet point list of major user facing changes to Sophie Watson by the end of burn down

[shwina]

cuDF Python

New Features

• Searching for multiple occurrences of a substring: [REVIEW] Add .str.find_multiple API #11928
• ngroup groupby method: Add ngroup #11871
• Support string operations in UDFs that return strings (note that this is currently expected to be quite slow):
  • Support + in strings_udf #12117
  • Support strip, lstrip, and rstrip in strings_udf #12091
  • Support upper and lower in strings_udf #12099
• Add DataFrame.pivot_table. #12015
• Constructing DataFrames from dictionaries: [REVIEW] Add support for DataFrame.from_dict\to_dict and Series.to_dict #12048
• (highly experimental, limited functionality, benchmarks TBD) cuDF can spill to host memory automatically: Spilling to host memory #12106

Breaking changes

• Some type promotion rules have changed (to match Pandas behaviour): Fix type promotion edge cases in numerical binops #12074

[randerzander]

** Dask-SQL **

New/Restored Features:

• Support for create experiment syntax
• Automatic handling of nullable types for ML train & predict features
• Accept split_out to be used by SQL operations that result in a Dask DataFrame group by
• Accept more SQL types
• More query optimization rules from DataFusion
• Fix Hive partition type and quote handling
• Optimize "select ... from table limit N" type queries, Sort + limit topk optimization (initial) dask-contrib/dask-sql#893
• Expanded timestamp support:
  Support to_timestamp dask-contrib/dask-sql#838
  Add parser support for timestampadd dask-contrib/dask-sql#857
  Check for np.timedelta64 in as_timelike dask-contrib/dask-sql#860
  Add TIMESTAMPDIFF support dask-contrib/dask-sql#876
• Add pip packages to release artifacts, including for Apple silicon
• Upgrade to DataFusion 14
• Replace Dask-ML w/ builtin functionality
• Add support for replace and coalesce
• Expanded support for predicate pushdown filtering

[quasiben]

Dask

New Features

CPU/GPU dispatching is now available for Dask-Dataframe creation allowing users to easily switch between CPU/GPU for Dask workflows: #9475 and #11920

Groupby aggregation performance increase with shuffle-based algorithms is now on by default: #9406

[cjnolet]

RAFT

Here's the Github board w/ all the work completed in release 22.12: https://github.com/orgs/rapidsai/projects/55/views/1?filterQuery=release%3A%2222.12%22+status%3ADone+.

Summary: In preparation for becoming a top-level RAPIDS library with its own userbase, lots of effort was spent on cleaning up the APIs, documentation, tests and general user experience during this release. In addition, lots of improvements were made to existing primitives and pylibraft.

Highlights

Primitives

• Public APIs: Several primitives were converted to new, cleaner, and more documenting APIs based on mdspan in raft::neighbors, raft::stats, raft::random, raft::linalg and raft::matrix

• New Primitives:

  • Primitives for kernel gramm matrix construction moved from cuML to raft::distance::kernel
  • New method added to raft::neighbors for refining and reducing a set of candidate nearest neighbors into a final set of size k.
  • New argmin and argmax primitives added to raft::stats

• Performance Enhancements:

  • Distance primitives now utilizing CUTLASS for L2/Cosine distances. This also makes use of tensor cores for Ampere architectures and above.
  • Several important performance enhancements made to k-means clustering and approximate nearest neighbors algorithms

Python

• New pylibraft.common.device_ndarray to enable pylibraft functions to return device arrays which interop with other __cuda_array_interface__ libraries (like cupy and pytorch)
• Wrappers for several primitives added to pylibraft:
  • k-means clustering algorithm
  • ivf-pq approximate nearest neighbors graph construction
  • approximate nearest neighbors refinement
  • efficient fused computation of 1-nearest neighbors

User Experience

• Moved Getting Started, Build & Install Guide, and Developer Guide to the main RAFT docs.
• Many enhancements to API documentation overall- improved coverage of API and usage examples.