Skip to main content

2 posts tagged with "pyspark"

View All Tags

· 7 min read
Jeffrey Aven

Loading Parquet format files into BigQuery is straightforward, you just need to specify the file location (local, Google Cloud Storage, Drive, Amazon S3 or Azure Blob storage) and thats pretty much it, BigQuery works the rest out from there.

bq load \
--location=australia-southeast2 \
--project_id=parquet-demo \
--source_format=PARQUET \
parquet_test.dim_calendar \
.\Calendar.gzip

In Snowflake, however, it is not as simple, I'll share my approach to automating this here.

info

Parquet is a self-describing, column-oriented storage format commonly used in distributed systems for input and output. Data in Parquet files is serialised for optimised consumption from Parquet client libraries and packages such as pandas, pyarrow, fastparquet, dask, and pyspark.

Background

Data in a Parquet file is stored in a single column for a self-contained dataset. If you were to ingest this into Snowflake without knowing the schema you could do something like this...

CREATE OR REPLACE TABLE PARQUET_TEST.PUBLIC.DIM_CALENDAR (
Data variant
);

COPY INTO PARQUET_TEST.PUBLIC.DIM_CALENDAR
(
Data
) FROM (
SELECT
*
FROM
@PARQUET_TEST.PUBLIC.DIM_CALENDAR_STAGE)
file_format = (TYPE = parquet);

You would end up with something like...

RowData
1{"CalMonthOfYearNo": 6, "CalYear": 2020, ... }
2{"CalMonthOfYearNo": 6, "CalYear": 2020, ... }
......

You could then have a second stage of processing to convert this into a normal relational structure.

Or you could do this in one step, with a little prep work ahead of time. In my scenario I was given several parquet files from a client for a one-off load into Snowflake, several files for a fact table and multiple single files representing different dimension tables.

Streamlined Ingestion for Parquet Files into Snowflake

To collapse the formatting and uploading of Parquet files into a materialized table into one step, we need to do a couple of things:

  1. Create the target table with the correct schema (column names and data types); and
  2. perform a projection in our COPY command from the single column containing all of the data (represented by $1 in Snowflake) into columns defined in step 1

Since this is technically a transformation and only named stages are supported for COPY transformations, we need to create a stage for the copy. In my case there is a pre-existing Storage Integration in place that can be used by the stage.

Generate Table DDL

To automate the generation of the DDL to create the table and stage and the COPY command, I used Python and Spark (which has first class support for Parquet files). Parquet datatypes are largely the same as Snowflake, but if we needed to, we could create a map and modify the target types during the DDL generation.

First copy specimen Parquet formatted files to a local directory, the script we are creating can then iterate through the parquet files and generate all of the commands we will need saved to a .sql file.

With some setup information provided (not shown for brevity), we will first go through each file in the directory, capture metadata along with the schema (column name and data type) as shown here:

for file in files:
tableMap = {}
table = file.stem
spark = launch_spark_session()
parquetFile = spark.read.parquet("%s/%s" %(BASE_DIR, file))
data_types = parquetFile.dtypes
stop_spark_session(spark)
tableMap['name'] = table
tableMap['file'] = file
tableMap['data_types'] = data_types
allTables.append(tableMap)

The allTables list looks something like this...

[{'name': 'Calendar', 'file': PosixPath('data/dim/Calendar.gzip'), 'data_types': [('Time_ID', 'bigint'), ('CalYear', 'bigint'), ('CalMonthOfYearNo', 'bigint'), ('FinYear', 'bigint'), ('FinWeekOfYearNo', 'bigint')]}, ... ]

Next we generate the CREATE TABLE statement using the allTables list:

# create output file for all sql
with open('all_tables.sql', 'w') as f:
for table in allTables:
print("processing %s..." % table['name'])
f.write("/*** Create %s Table***/" % table['name'].upper())
sql = """
CREATE OR REPLACE TABLE %s.%s.%s (
""" % (database, schema, table['name'].upper())
for column in table['data_types']:
sql += " %s %s,\n" % (column[0], column[1])
sql = sql[:-2] + "\n);"
f.write(sql)
f.write("\n\n")

Generate Named Stage DDL

Then we generate the stage in S3 from which the files will be loaded:

        f.write("/*** Create %s Stage***/" % table['name'].upper())
sql = """
CREATE OR REPLACE STAGE %s.%s.%s_STAGE
url='%s/%s'
storage_integration = %s
encryption=(type='AWS_SSE_KMS' kms_key_id = '%s');
""" % (database, schema, table['name'].upper(), s3_prefix, table['file'], storage_int, kms_key_id)
f.write(sql)
f.write("\n\n")

Generate COPY commands

Then we generate the COPY commands...

        f.write("/*** Copying Data into %s ***/" % table['name'].upper())
sql = """
COPY INTO %s.%s.%s
(\n""" % (database, schema, table['name'].upper())
for column in table['data_types']:
sql += " %s,\n" % column[0]
sql = sql[:-2] + "\n)"
sql += " FROM (\nSELECT\n"
for column in table['data_types']:
sql += " $1:%s::%s,\n" % (column[0], column[1])
sql = sql[:-2] + "\nFROM\n"
sql += "@%s.%s.%s_STAGE)\n" % (database, schema, table['name'].upper())
sql += " file_format = (TYPE = parquet);"
f.write(sql)
f.write("\n\n")

Since this is a one off load, we will go ahead and drop the stage we created as it is no longer needed (this step is optional)..

        f.write("/*** Dropping stage for %s ***/" % table['name'].upper())
sql = """
DROP STAGE %s.%s.%s_STAGE;
""" % (database, schema, table['name'].upper())
f.write(sql)
f.write("\n\n")

The resultant file created looks like this..

/*** Create CALENDAR Table***/
CREATE OR REPLACE TABLE PARQUET_TEST.PUBLIC.DIM_CALENDAR (
Time_ID bigint,
CalYear bigint,
CalMonthOfYearNo bigint,
FinYear bigint,
FinWeekOfYearNo bigint
);

/*** Create DIM_CALENDAR Stage***/
CREATE OR REPLACE STAGE PARQUET_TEST.PUBLIC.DIM_CALENDAR_STAGE
url='s3://my-bucket/data/dim/Calendar.gzip'
storage_integration = my_storage_int
encryption=(type='AWS_SSE_KMS' kms_key_id = '4f715ec9-ee8e-44ab-b35d-8daf36c05f19');

/*** Copying Data into DIM_CALENDAR ***/
COPY INTO PARQUET_TEST.PUBLIC.DIM_CALENDAR
(
Time_ID,
CalYear,
CalMonthOfYearNo,
FinYear,
FinWeekOfYearNo
) FROM (
SELECT
$1:Time_ID::bigint,
$1:CalYear::bigint,
$1:CalMonthOfYearNo::bigint,
$1:FinYear::bigint,
$1:FinWeekOfYearNo::bigint
FROM
@PARQUET_TEST.PUBLIC.DIM_CALENDAR_STAGE)
file_format = (TYPE = parquet);

/*** Dropping stage for DIM_CALENDAR ***/
DROP STAGE PARQUET_TEST.PUBLIC.DIM_CALENDAR_STAGE;

Load your data

You can then run this along with all of the other dimension and fact table DDL and COPY commands generated to perform the one-off load from parquet files. You can find the complete code below, enjoy!

Complete Code
from pathlib import Path
from pyspark.sql import SparkSession
def launch_spark_session():
return SparkSession \
.builder \
.appName("Parquet DDL Generation") \
.getOrCreate()

def stop_spark_session(spark):
spark.stop()

allTables = []
database = "PARQUET_TEST"
schema = "PUBLIC"
s3_prefix = 's3://my-bucket'
storage_int = 'my_storage_int'
kms_key_id = '4f715ec9-ee8e-44ab-b35d-8daf36c05f19'

BASE_DIR = Path(__file__).resolve().parent
directory = 'data/dim'
files = Path(directory).glob('*.gzip')
for file in files:
tableMap = {}
table = file.stem
spark = launch_spark_session()
parquetFile = spark.read.parquet("%s/%s" %(BASE_DIR, file))
data_types = parquetFile.dtypes
stop_spark_session(spark)
tableMap['name'] = table
tableMap['file'] = file
tableMap['data_types'] = data_types
allTables.append(tableMap)

# create output file for all sql
with open('all_tables.sql', 'w') as f:
for table in allTables:
print("processing %s..." % table['name'])
f.write("/*** Create %s Table***/" % table['name'].upper())
sql = """
CREATE OR REPLACE TABLE %s.%s.%s (
""" % (database, schema, table['name'].upper())
for column in table['data_types']:
sql += " %s %s,\n" % (column[0], column[1])
sql = sql[:-2] + "\n);"
f.write(sql)
f.write("\n\n")

f.write("/*** Create %s Stage***/" % table['name'].upper())
sql = """
CREATE OR REPLACE STAGE %s.%s.%s_STAGE
url='%s/%s'
storage_integration = %s
encryption=(type='AWS_SSE_KMS' kms_key_id = '%s');
""" % (database, schema, table['name'].upper(), s3_prefix, table['file'], storage_int, kms_key_id)
f.write(sql)
f.write("\n\n")

f.write("/*** Copying Data into %s ***/" % table['name'].upper())
sql = """
COPY INTO %s.%s.%s
(\n""" % (database, schema, table['name'].upper())
for column in table['data_types']:
sql += " %s,\n" % column[0]
sql = sql[:-2] + "\n)"
sql += " FROM (\nSELECT\n"
for column in table['data_types']:
sql += " $1:%s::%s,\n" % (column[0], column[1])
sql = sql[:-2] + "\nFROM\n"
sql += "@%s.%s.%s_STAGE)\n" % (database, schema, table['name'].upper())
sql += " file_format = (TYPE = parquet);"
f.write(sql)
f.write("\n\n")

f.write("/*** Dropping stage for %s ***/" % table['name'].upper())
sql = """
DROP STAGE %s.%s.%s_STAGE;
""" % (database, schema, table['name'].upper())
f.write(sql)
f.write("\n\n")

· 9 min read
Jeffrey Aven

CDC using Spark

Change Data Capture (CDC) is one of the most challenging processing patterns to implement at scale. I personally have had several cracks at this using various different frameworks and approaches, the most recent of which was implemented using Spark – and I think I have finally found the best approach. Even though the code examples referenced use Spark, the pattern is language agnostic – the focus is on the approach not the specific implementation (as this could be applied to any framework or runtime).

The first challenge you are faced with, is to compare a very large dataset (representing the current state of an object) with another potentially very large dataset (representing new or incoming data). Ideally, you would like the process to be configuration driven and accommodate such things as composite primary keys, or operational columns which you would like to restrict from change detection. You may also want to implement a pattern to segregate sensitive attributes from non-sensitive attributes.

Overview

This pattern (and all my other recent attempts) is fundamentally based upon calculating a deterministic hash of the key and non-key attribute(s), and then using this hash as the basis for comparison. The difference between this pattern and my other attempts is in the distillation and reconstitution of data during the process, as well as breaking the pattern into discrete stages (designed to minimize the impact to other applications). This pattern can be used to process delta or full datasets.

A high-level flowchart representing the basic pattern is shown here:

CDC Flowchart

The Example

The example provided uses the Synthetic CDC Data Generator application, configuring an incoming set with 5 uuid columns acting as a composite key, and 10 random number columns acting as non key values. The initial days payload consists of 10,000 records, the subsequent days payload consists of another 10,000 records. From the initial dataset, a DELETE operation was performed at the source system for 20% of records, an UPDATE was performed on 40% of the records and the remaining 40% of records were unchanged. In this case the 20% of records that were deleted at the source, were replaced by new INSERT operations creating new keys.

After creating the synthesized day 1 and day 2 datasets, the files are processed as follows:

$ spark-submit cdc.py config.yaml data/day1 2019-06-18
$ spark-submit cdc.py config.yaml data/day2 2019-06-19

Where config.yaml is the configuration for the dataset, data/day1 and data/day2 represent the different data files, and 2019-06-18 and 2019-06-19 represent a business effective date.

The Results

You should see the following output from running the preceding commands for day 1 and day 2 respectively:

Day 1:

Day 2:

A summary analysis of the resultant dataset should show:

Pattern Details

Details about the pattern and its implementation follow.

Current and Historical Datasets

The output of each operation will yield a current dataset (that is the current stateful representation of a give object) and a historical dataset partition (capturing the net changes from the previous state in an appended partition).

This is useful, because often consumers will primarily query the latest state of an object. The change sets (or historical dataset partitions) can be used for more advanced analysis by sophisticated users.

Type 2 SCDs (sort of)

Two operational columns are added to each current and historical object:

  • OPERATION : Represents the last known operation to the record, valid values include :
    • I (INSERT)
    • U (UPDATE)
    • D (DELETE – hard DELETEs, applies to full datasets only)
    • X (Not supplied, applies to delta processing only)
    • N (No change)
  • EFF_START_DATE

Since data structures on most big data or cloud storage platforms are immutable, we only store the effective start date for each record, this is changed as needed with each coarse-grained operation on the current object. The effective end date is inferred by the presence of a new effective start date (or change in the EFF_START_DATE value for a given record).

The Configuration

I am using a YAML document to store the configuration for the pattern. Important attributes to include in your configuration are a list of keys and non keys and their datatype (this implementation does type casting as well). Other important attributes include the table names and file paths for the current and historical data structures.

The configuration is read at the beginning of a routine as an input along with the path of an incoming data file (a CSV file in this case) and a business effective date (which will be used as the EFF_START_DATE for new or updated records).

Processing is performed using the specified key and non key attributes and the output datasets (current and historical) are written to columnar storage files (parquet in this case). This is designed to make subsequent access and processing more efficient.

The Algorithm

I have broken the process into stages as follows:

Stage 1 – Type Cast and Hash Incoming Data

The first step is to create deterministic hashes of the configured key and non key values for incoming data. The hashes are calculated based upon a list of elements representing the key and non key values using the MD5 algorithm. The hashes for each record are then stored with the respective record. Furthermore, the fields are casted their target datatype as specified in the configuration. Both of these operations can be performed in a single pass of each row using a map() operation.

Importantly we only calculate hashes once upon arrival of new data, as the hashes are persisted for the life of the data – and the data structures are immutable – the hashes should never change or be invalidated.

Stage 2 – Determine INSERTs

We now compare Incoming Hashes with previously calculated hash values for the (previous day’s) current object. If no current object exists for the dataset, then it can be assumed this is a first run. In this case every record is considered as an INSERT with an EFF_START_DATE of the business effective date supplied.

If there is a current object, then the key and non key hash values (only the hash values) are read from the current object. These are then compared to the respective hashes of the incoming data (which should still be in memory).

Given the full outer join:

incoming_data(keyhash, nonkeyhash) FULL OUTER JOIN
current_data(keyhash, nonkeyhash) ON keyhash

Keys which exist in the left entity which do not exist in the right entity must be the results of an INSERT operation.

Tag these records with an operation of I with an EFF_START_DATE of the business effective date, then rejoin only these records with their full attribute payload from the incoming dataset. Finally, write out these records to the current and historical partition in overwrite mode.

Stage 3 - Determine DELETEs or Missing Records

Referring the previous full outer join operation, keys which exist in the right entity (current object) which do not appear in the left entity (incoming data) will be the result of a (hard) DELETE operation if you are processing full snapshots, otherwise if you are processing deltas these would be missing records (possibly because there were no changes at the source).

Tag these records as D or X respectively with an EFF_START_DATE of the business effective date, rejoin these records with their full attribute payload from the current dataset, then write out these records to the current and historical partition in append mode.

Stage 4 - Determine UPDATEs or Unchanged Records

Again, referring to the previous full outer join, keys which exist in both the incoming and current datasets must be either the result of an UPDATE or they could be unchanged. To determine which case they fall under, compare the non key hashes. If the non key hashes differ, it must have been a result of an UPDATE operation at the source, otherwise the record would be unchanged.

Tag these records as U or N respectively with an EFF_START_DATE of the business effective date (in the case of an update - otherwise maintain the current EFF_START_DATE), rejoin these records with their full attribute payload from the incoming dataset, then write out these records to the current and historical partition in append mode.

Key Callouts

A summary of the key callouts from this pattern are:

  • Use the RDD API for iterative record operations (such as type casting and hashing)
  • Persist hashes with the records
  • Use Dataframes for JOIN operations
  • Only perform JOINs with the keyhash and nonkeyhash columns – this minimizes the amount of data shuffled across the network
  • Write output data in columnar (Parquet) format
  • Break the routine into stages, covering each operation, culminating with a saveAsParquet() action – this may seem expensive but for large datsets it is more efficient to break down DAGs for each operation
  • Use caching for objects which will be reused between actions

Metastore Integration

Although I did not include this in my example, you could easily integrate this pattern with a metastore (such as a Hive metastore or AWS Glue Catalog), by using table objects and ALTER TABLE statements to add historical partitions.

Further optimisations

If the incoming data is known to be relatively small (in the case of delta processing for instance), you could consider a broadcast join where the smaller incoming data is distributed to all of the different Executors hosting partitions from the current dataset.

Also you could add a key to the column config to configure a column to be nullable or not.

Happy CDCing!

Full source code for this article can be found at: https://github.com/avensolutions/cdc-at-scale-using-spark