Post Syndicated from Kalaiselvi Kamaraj original https://aws.amazon.com/blogs/big-data/accelerate-amazon-redshift-data-lake-queries-with-column-level-statistics/

Amazon Redshift enables you to efficiently query and retrieve structured and semi-structured data from open format files in Amazon S3 data lake without having to load the data into Amazon Redshift tables. Amazon Redshift extends SQL capabilities to your data lake, enabling you to run analytical queries. Amazon Redshift supports a wide variety of tabular data formats like CSV, JSON, Parquet, ORC and open tabular formats like Apache Hudi, Linux foundation Delta Lake and Apache Iceberg.

You create Redshift external tables by defining the structure for your files, S3 location of the files and registering them as tables in an external data catalog. The external data catalog can be AWS Glue Data Catalog, the data catalog that comes with Amazon Athena, or your own Apache Hive metastore.

Over the last year, Amazon Redshift added several performance optimizations for data lake queries across multiple areas of query engine such as rewrite, planning, scan execution and consuming AWS Glue Data Catalog column statistics. To get the best performance on data lake queries with Redshift, you can use AWS Glue Data Catalog’s column statistics feature to collect statistics on Data Lake tables. For Amazon Redshift Serverless instances, you will see improved scan performance through increased parallel processing of S3 files and this happens automatically based on RPUs used.

In this post, we highlight the performance improvements we observed using industry standard TPC-DS benchmarks. Overall execution time of TPC-DS 3 TB benchmark improved by 3x. Some of the queries in our benchmark experienced up to 12x speed up.

Performance Improvements

Several performance optimizations were done over the last year to improve performance of data lake queries including the following.

• Consume AWS Glue Data Catalog column statistics and tuning of Redshift optimizer to improve quality of query plans
• Utilize bloom filters for partition columns
• Improved scan efficiency for Amazon Redshift Serverless instances through increased parallel processing of files
• Novel query rewrite rules to merge similar scans
• Faster retrieval of metadata from AWS Glue Data Catalog

To understand the performance gains, we tested the performance on the industry-standard TPC-DS benchmark using 3 TB data sets and queries which represents different customer use cases. Performance was tested on a Redshift serverless data warehouse with 128 RPU. In our testing, the dataset was stored in Amazon S3 in Parquet format and AWS Glue Data Catalog was used to manage external databases and tables. Fact tables were partitioned on the date column, and each fact table consisted of approximately 2,000 partitions. All of the tables had their row count table property, numRows, set as per the spectrum query performance guidelines.

We did a baseline run on Redshift patch version (patch 172) from last year. Later, we ran all TPC-DS queries on latest patch version (patch 180) that includes all performance optimizations added over last year. Then we used AWS Glue Data Catalog’s column statistics feature to compute statistics for all the tables and measured improvements with the presence of AWS Glue Data Catalog column statistics.

Our analysis revealed that the TPC-DS 3TB Parquet benchmark saw substantial performance gains with these optimizations. Specifically, partitioned Parquet with our latest optimizations achieved 2x faster runtimes compared to the previous implementation. Enabling AWS Glue Data Catalog column statistics further improved performance by 3x versus last year. The following graph illustrates these runtime improvements for the full benchmark (all TPC-DS queries) over the past year, including the additional boost from using AWS Glue Data Catalog column statistics.

Figure 1: Improvement in total runtime of TPC-DS 3T workload

The following graph presents the top queries from the TPC-DS benchmark with the greatest performance improvement over the last year with and without AWS Glue Data Catalog column statistics. You can see that performance improves a lot when statistics exist on AWS Glue Data Catalog (for details on how to get statistics for your Data Lake tables, please refer to optimizing query performance using AWS Glue Data Catalog column statistics). Specifically, multi-join queries will benefit the most from AWS Glue Data Catalog column statistics because the optimizer uses statistics to choose the right join order and distribution strategy.

Figure 2: Speed-up in TPC-DS queries

Let’s discuss some of the optimizations that contributed to improved query performance.

Optimizing with table-level statistics

Amazon Redshift’s design enables it to handle large-scale data challenges with superior speed and cost-efficiency. Its massively parallel processing (MPP) query engine, AI-powered query optimizer, auto-scaling capabilities, and other advanced features allow Redshift to excel at searching, aggregating, and transforming petabytes of data.

However, even the most powerful systems can experience performance degradation if they encounter anti-patterns like grossly inaccurate table statistics, such as the row count metadata.

Without this crucial metadata, Redshift’s query optimizer may be limited in the number of possible optimizations, especially those related to data distribution during query execution. This can have a significant impact on overall query performance.

To illustrate this, consider the following simple query involving an inner join between a large table with billions of rows and a small table with only a few hundred thousand rows.

select small_table.sellerid, sum(large_table.qtysold)
from large_table, small_table
where large_table.salesid = small_table.listid
 and small_table.listtime > '2023-12-01'
 and large_table.saletime > '2023-12-01'
group by 1 order by 1

If executed as-is, with the large table on the right-hand side of the join, the query will lead to sub-optimal performance. This is because the large table will need to be distributed (broadcast) to all Redshift compute nodes to perform the inner join with the small table, as shown in the following diagram.

Figure 3: Inaccurate table statistics lead to limited optimizations and large amounts of data broadcast among compute nodes for a simple inner join

Now, consider a scenario where the table statistics, such as the row count, are accurate. This allows the Amazon Redshift query optimizer to make more informed decisions, such as determining the optimal join order. In this case, the optimizer would immediately rewrite the query to have the large table on the left-hand side of the inner join, so that it is the small table that is broadcast across the Redshift compute nodes, as illustrated in the following diagram.

Figure 4: Accurate table statistics lead to high degree of optimizations and very little data broadcast among compute nodes for a simple inner join

Fortunately, Amazon Redshift automatically maintains accurate table statistics for local tables by running the ANALYZE command in the background. For external tables (data lake tables), however, AWS Glue Data Catalog column statistics are recommended for use with Amazon Redshift as we will discuss in the next section. For more general information on optimizing queries in Amazon Redshift, please refer to the documentation on factors affecting query performance, data redistribution, and Amazon Redshift best practices for designing queries.

Improvements with AWS Glue Data Catalog column statistics

AWS Glue Data Catalog has a feature to compute column level statistics for Amazon S3 backed external tables. AWS Glue Data Catalog can compute column level statistics such as NDV, Number of Nulls, Min/Max and Avg. column width for the columns without the need for additional data pipelines. Amazon Redshift cost-based optimizer utilizes these statistics to come up with better quality query plans. In addition to consuming statistics, we also made several improvements in cardinality estimations and cost tuning to get high quality query plans thereby improving query performance.

TPC-DS 3TB dataset showed 40% improvement in total query execution time when these AWS Glue Data Catalog column statistics were provided. Individual TPC-DS queries showed up to 5x improvements in query execution time. Some of the queries that had greater impact in execution time are Q85, Q64, Q75, Q78, Q94, Q16, Q04, Q24 and Q11.

We will go through an example where cost-based optimizer generated a better query plan with statistics and how it improved the execution time.

Let’s consider following simpler version of TPC-DS Q64 to showcase the query plan differences with statistics.

select i_product_name product_name
,i_item_sk item_sk
,ad1.ca_street_number b_street_number
,ad1.ca_street_name b_street_name
,ad1.ca_city b_city
,ad1.ca_zip b_zip
,d1.d_year as syear
,count(*) cnt
,sum(ss_wholesale_cost) s1
,sum(ss_list_price) s2
,sum(ss_coupon_amt) s3
FROM   tpcds_3t_alls3_pp_ext.store_sales
,tpcds_3t_alls3_pp_ext.store_returns
,tpcds_3t_alls3_pp_ext.date_dim d1
,tpcds_3t_alls3_pp_ext.customer
,tpcds_3t_alls3_pp_ext.customer_address ad1
,tpcds_3t_alls3_pp_ext.item
WHERE
ss_sold_date_sk = d1.d_date_sk AND
ss_customer_sk = c_customer_sk AND

ss_addr_sk = ad1.ca_address_sk and
ss_item_sk = i_item_sk and
ss_item_sk = sr_item_sk and
ss_ticket_number = sr_ticket_number and
i_color in ('firebrick','papaya','orange','cream','turquoise','deep') and
i_current_price between 42 and 42 + 10 and
i_current_price between 42 + 1 and 42 + 15
group by i_product_name
,i_item_sk
,ad1.ca_street_number
,ad1.ca_street_name
,ad1.ca_city
,ad1.ca_zip
,d1.d_year

Without Statistics Following figure represents the logical query plan of Q64. You can observe that cardinality estimation of joins is not accurate. With inaccurate cardinalities, optimizer produces a sub-optimal query plan leading to higher execution time.

With Statistics Following figure represents the logical query plan after consuming AWS Glue Data Catalog column statistics. Based on the highlighted changes, you can observe that the cardinality estimations of JOIN improved by many magnitudes helping the optimizer to choose a better join order and join strategy (broadcast DS_BCAST_INNER vs. distribute DS_DIST_BOTH). Switching the customer_address and customer table from inner to outer table and making join strategies as distribute has major impact because this reduces the data movement between the nodes and avoids spilling from hash table.

Figure 5: Logical query plan of Q64 without statistics

Figure 6: Logical query plan of Q64 after consuming AWS Glue Data Catalog column statistics

This change in query plan improved the query execution time of Q64 from 383s to 81s.

Given the greater benefits with AWS Glue Data Catalog column statistics for the optimizer, you should consider collecting stats for your data lake using AWS Glue. If your workload is a JOIN heavy workload, then collecting stats will show greater improvement on your workload. Refer to generating AWS Glue Data Catalog column statistics for instructions on how to collect statistics in AWS Glue Data Catalog.

Query rewrite optimization

We introduced a new query rewrite rule which combines scalar aggregates over the same common expression using slightly different predicates. This rewrite resulted in performance improvements on TPC-DS queries Q09, Q28, and Q88. Let’s focus on Q09 as a representative of these queries, given by the following fragment:

SELECT CASE
WHEN (SELECT COUNT(*)
FROM store_sales
WHERE ss_quantity BETWEEN 1 AND 20) > 48409437
THEN (SELECT AVG(ss_ext_discount_amt)
FROM store_sales
WHERE ss_quantity BETWEEN 1 AND 20)
ELSE (SELECT AVG(ss_net_profit)
FROM store_sales
WHERE ss_quantity BETWEEN 1 AND 20) END
AS bucket1,
<<4 more variations of the CASE expression above>>
FROM reason
WHERE r_reason_sk = 1

In total, there are 15 scans of the fact table store_sales, each one returning various aggregates over different subsets of data. The engine first performs subquery removal and transforms the various expressions in the CASE statements into relational subtrees connected via cross products, and then they are fused into one subquery handling all scalar aggregates. The resulting plan for Q09, described below using SQL for clarity, is given by:

SELECT CASE WHEN v1 > 48409437 THEN t1 ELSE e1 END,
<4 more variations>
FROM (SELECT COUNT(CASE WHEN b1 THEN 1 END) AS v1,
AVG(CASE WHEN b1 THEN ss_ext_discount_amt END) AS t1,
AVG(CASE WHEN b1 THEN ss_net_profit END) AS e1,
<4 more variations>
FROM reason,
(SELECT *,
ss_quantity BETWEEN 1 AND 20 AS b1,
<4 more variations>
FROM store_sales
WHERE ss_quantity BETWEEN 1 AND 20 OR
<4 more variations>))
WHERE r_reason_sk = 1)

In general, this rewrite rule results in the largest improvements both in latency (from 3x to 8x improvements) and bytes read from Amazon S3 (from 6x to 8x reduction in scanned bytes and, consequently, cost).

Bloom filter for partition columns

Amazon Redshift already uses Bloom filters on data columns of external tables in Amazon S3 to enable early and effective data filtering. Last year, we extended this support for partition columns as well. A Bloom filter is a probabilistic, memory-efficient data structure that accelerates join queries at scale by filtering rows that do not match the join relation, significantly reducing the amount of data transferred over the network. Amazon Redshift automatically determines what queries are suitable for leveraging Bloom filters at query runtime.

This optimization resulted in performance improvements on TPC-DS queries Q05, Q17 and Q54. This optimization resulted in large improvements in both latency (from 2x to 3x improvement) and bytes read from S3 (from 9x to 15x reduction in scanned bytes and, consequently cost).

Following is the subquery of Q05 which showcased improvements with runtime filter.

select s_store_id,
sum(sales_price) as sales,
sum(profit) as profit,
sum(return_amt) as returns,
sum(net_loss) as profit_loss
from
( select  ss_store_sk as store_sk,
ss_sold_date_sk  as date_sk,
ss_ext_sales_price as sales_price,
ss_net_profit as profit,
cast(0 as decimal(7,2)) as return_amt,
cast(0 as decimal(7,2)) as net_loss
from tpcds_3t_alls3_pp_ext.store_sales
union all
select sr_store_sk as store_sk,
sr_returned_date_sk as date_sk,
cast(0 as decimal(7,2)) as sales_price,
cast(0 as decimal(7,2)) as profit,
sr_return_amt as return_amt,
sr_net_loss as net_loss
from tpcds_3t_alls3_pp_ext.store_returns
) salesreturnss,
tpcds_3t_alls3_pp_ext.date_dim,
tpcds_3t_alls3_pp_ext.store
where date_sk = d_date_sk
and d_date between cast('1998-08-13' as date)
and (cast('1998-08-13' as date) +  14)
and store_sk = s_store_sk
group by s_store_id

Without bloom filter support on partition columns Following figure is the logical query plan for sub-query of Q05. This appends two large fact tables store_sales (8B rows) and store_returns (863M rows) and then joins with very selective dimension tables date_dim and then with dimension table store. You can observe that join with date_dim table reduces the number of rows from 9B to 93M rows.

With bloom filter support on partition columns With support of bloom filter on partition columns, we now create bloom filter for d_date_sk column of date_dim table and push down the bloom filters to store_sales and store_returns table. These bloom filters help to filter out the partitions in both store_sales and store_returns table because join happens on partition column (number of partitions processed reduces by 10x).

Figure 7: Logical query plan for sub-query of Q05 without bloom filter support on partition columns

Figure 8: Logical query plan for sub-query of Q05 with bloom filter support on partition columns

Overall, bloom filter on partition column will reduce the number of partitions processed resulting in reduced S3 listing calls and lesser number of data files to be read (reduction in scanned bytes). You can see that we only scan 89M rows from store_sales and 4M rows from store_returns because of the bloom filter. This reduced number of rows to process at JOIN level and helped in improving the overall query performance by 2x and scanned bytes by 9x.

Conclusion

In this post, we covered new performance optimizations in Amazon Redshift data lake query processing and how AWS Glue Data Catalog statistics helps to enhance quality of query plans for data lake queries in Amazon Redshift. These optimizations together improved TPC-DS 3 TB benchmark by 3x. Some of the queries in our benchmark benefited up to 12x speed up.

In summary, Amazon Redshift now offers enhanced query performance with optimizations such as AWS Glue Data Catalog column statistics, bloom filters on partition columns, new query rewrite rules and faster retrieval of metadata. These optimizations are enabled by default and Amazon Redshift users will benefit with better query response times for their workloads. For more information, please reach out to your AWS technical account manager or AWS account solutions architect. They will be happy to provide additional guidance and support.

About the authors

Kalaiselvi Kamaraj is a Sr. Software Development Engineer with Amazon. She has worked on several projects within Redshift Query processing team and currently focusing on performance related projects for Redshift Data Lake.

Mark Lyons is a Principal Product Manager on the Amazon Redshift team. He works on the intersection of data lakes and data warehouses. Prior to joining AWS, Mark held product leadership roles with Dremio and Vertica. He is passionate about data analytics and empowering customers to change the world with their data.

Asser Moustafa is a Principal Worldwide Specialist Solutions Architect at AWS, based in Dallas, Texas, USA. He partners with customers worldwide, advising them on all aspects of their data architectures, migrations, and strategic data visions to help organizations adopt cloud-based solutions, maximize the value of their data assets, modernize legacy infrastructures, and implement cutting-edge capabilities like machine learning and advanced analytics. Prior to joining AWS, Asser held various data and analytics leadership roles, completing an MBA from New York University and an MS in Computer Science from Columbia University in New York. He is passionate about empowering organizations to become truly data-driven and unlock the transformative potential of their data.