pyrecdp 0.1.2

A data processing bundle for spark based recommender system operations

What is RecDP

• RecDP is a Data Process python module, specifically designed for Recommender System.

Objective

• Easy-to-use – simple APIs for data scientists, easy to migrate from NVTabular
• Collaborative pipeline with spark and modin - provide stableness and scalability of handling huge dataset with spark and modin as underlying distributed data process engine.
• Optimized Performance - 1) Adaptive dataframe plan decision making; 2) Intel-OAP accelerator extensions (SIMD, Cache, Native).
• Extensible – provides plugin interface to extend intel RECO & Friesian with optimized adaptive DataProcess pipeline plan with spark and modin.
• Feature Engineer oriented – advanced feature engineering functions (target encoding)

Currently RecDP is proven by four use case:

• Recsys2021: successfully support intel Recsys2021 challenge feature engineering work
• Recsys2020: successfully processing over 600 millions dataset and aligned with Recsys2021 winner feature engineering work.
• DRLM: successfully processing Criteo dataset of 24 days w/wo frequence limit, previously wo/ frequence limit went failed using NVIDIA provided spark script.
• DIEN: w/ RecDP, process time is 6x speeding up comparing with original Ali-Matrix python script.

Design Overview

How to start

compile scala extension

• noted: support spark 3.1 by default, using -pspark3.0 for running with Spark3.0

cd ScalaProcessUtils/
mvn package -Pspark-3.1
or
mvn package -Pspark-3.0

test with provided spark docker img

modify run_docker script

${your_local_codes_data_dir} is the path for current recdp folder

# docker run --network host -v ${your_local_codes_data_dir}:/home/vmagent/app/recdp -w /home/vmagent/app/ -it xuechendi/recdp_spark3.1 /bin/bash
docker run --network host -v /mnt/nvme2/chendi/BlueWhale/recdp:/home/vmagent/app/recdp -w /home/vmagent/app/ -it xuechendi/recdp_spark3.1 /bin/bash

# test with provided python script
python test/test_spark_local.py

# test with DIEN data process
cd recdp/examples/python_tests/dien/
# modify num_cores and memory_capacity
vim j2c_spark.py
# fix at line 87 to 93
spark = SparkSession.builder.master('local[${num_core}]')\
        .appName("dien_data_process")\
        .config("spark.driver.memory", "${memory_size}G")\
        .config("spark.executor.cores", "${num_core}")\
        .config("spark.driver.extraClassPath", f"{scala_udf_jars}")\
        .getOrCreate()

vim dien_data_process.py
#fix line 155 to 161
spark = SparkSession.builder.master(...

# call run
./download_dataset
./run

Expected output as below

test with provided jupyter notebook example

• Recsys2021 example url
• Recsys2020 example url
• Recsys2020 multiitem-categorify example(support for Analytics Zoo Friesian) url
• DLRM example url
• DIEN example url

write your own

• some spark configuration is required

import init

import findspark
findspark.init()

import os
from pyspark.sql import *
from pyspark import *
import pyspark.sql.functions as f
import pyspark.sql.types as t
from pyrecdp.data_processor import *
from pyrecdp.encoder import *
from pyrecdp.utils import *

scala_udf_jars = "${path_to_project}/recdp/ScalaProcessUtils/target/recdp-scala-extensions-0.1.0-jar-with-dependencies.jar"

##### 1. Start spark and initialize data processor #####
spark = SparkSession\
    .builder\
    .master('yarn')\  # switch to local[*] for local mode
    .appName("RecDP_test")\
    .config("spark.sql.broadcastTimeout", "7200")\  # tune up broadcast timeout
    .config("spark.cleaner.periodicGC.interval", "10min")\  # config GC interval according to your shuffle disk capacity, \
                                                            # if capacity is below 2T, smaller interval will trigue \
                                                            # spark shuffle blocks GC more often to release space.
    .config("spark.driver.extraClassPath", f"{scala_udf_jars}")\    # add recdp-scala-extension to spark
    .config("spark.executor.extraClassPath", f"{scala_udf_jars}")\
    .getOrCreate()
    
##### 2. init RecDP processor #####
path_prefix = "hdfs://"
current_path = "/recsys2021_0608_example/"  # workdir for recdp
shuffle_disk_capacity="1200GB"  # spark.local.dir / shuffle capacity, this will help recdp to do better plan.
                                # Please make sure this size is less than(about 80%) of your actual shuffle_disk_capacity.

proc = DataProcessor(spark, path_prefix,
                     current_path=current_path, shuffle_disk_capacity=shuffle_disk_capacity)

df = spark.read.parquet("/recsys2021_0608")

op_feature_from_original = FeatureAdd(
        cols={"has_photo": "f.col('present_media').contains('Photo').cast(t.IntegerType())",              
              "a_ff_rate": "f.col('engaged_with_user_following_count')/f.col('engaged_with_user_follower_count')",
              "dt_dow": "f.dayofweek(f.from_unixtime(f.col('tweet_timestamp'))).cast(t.IntegerType())",        
              "mention": "f.regexp_extract(f.col('tweet'), r'[^RT]\s@(\S+)', 1)"
        }, op='inline')

# execute
proc.reset_ops([op_feature_from_original])
df = proc.transform(df, name=output_name)   # data will be transformed when this line called

Test with OAP Gazelle Project

docker run -it --privileged --network host -w /home/vmagent/app/ xuechendi/recdp_spark3.1:gazelle /bin/bash
./run_jupyter
tail jupyter_error.log
    Or copy and paste one of these URLs:
        http://sr130:8888/?token=c631ab6db797517e3603e7450c00e85cfc3b52653f9da31e
     or http://127.0.0.1:8888/?token=c631ab6db797517e3603e7450c00e85cfc3b52653f9da31e
[I 08:24:19.503 NotebookApp] 302 GET / (10.0.0.101) 0.950000ms
[I 08:24:19.515 NotebookApp] 302 GET /tree? (10.0.0.101) 1.090000ms

run jupyter in browser You'll see sql plan as below

LICENSE

• Apache 2.0

Dependency

• Spark 3.x
• python 3.*