Amazon Redshift out-of-the-box efficiency improvements for knowledge lake queries

Databases and question engines, together with Amazon Redshift, usually depend on totally different statistics in regards to the underlying knowledge to find out the simplest method to execute a question, such because the variety of distinct values and which values have low selectivity. When Amazon Redshift receives a question, reminiscent of

SELECT insert_date, sum(gross sales)
FROM receipts
WHERE insert_date BETWEEN '2024-12-01' AND '2024-12-31'
GROUP BY insert_date

the question planner makes use of statistics to make an informed guess on the best methodology to load and course of knowledge from storage. Extra statistics in regards to the underlying knowledge can usually assist a question planner choose a plan that results in the most effective question efficiency, however this could require a tradeoff among the many price of computing, storing, and sustaining statistics, and would possibly require extra question planning time.

Information lakes are a strong structure to prepare knowledge for analytical processing, as a result of they let builders use environment friendly analytical columnar codecs like Apache Parquet, whereas letting them proceed to switch the form of their knowledge as their purposes evolve with open desk codecs like Apache Iceberg. One problem with knowledge lakes is that they don’t at all times have statistics about their underlying knowledge, making it tough for question engines to find out the optimum execution path. This could result in points, together with sluggish queries and surprising modifications in question efficiency.

In 2024, Amazon Redshift clients queried over 77 EB (exabytes) of information residing in knowledge lakes. Given this utilization, the Amazon Redshift staff works to innovate on knowledge lake question efficiency to assist clients effectively entry their open knowledge to get close to real-time insights to make vital enterprise selections. In 2024, Amazon Redshift launched a number of options that enhance question efficiency for knowledge lakes, together with sooner question occasions when an information lake doesn’t have statistics. With Amazon Redshift patch 190, the TPC-DS 3TB benchmark confirmed an total 2x question efficiency enchancment on Apache Iceberg tables with out statistics, together with TPC-DS Question #72, which improved by 125 occasions from 690 seconds to five.5 seconds.

On this put up, we first briefly overview how planner statistics are collected and what impression they’ve on queries. Then, we focus on Amazon Redshift options that ship optimum plans on Iceberg tables and Parquet knowledge even with the shortage of statistics. Lastly, we overview some instance queries that now execute sooner due to these newest Amazon Redshift improvements.

Conditions

The benchmarks on this put up have been run utilizing the next atmosphere:

• Amazon Redshift Serverless with a base capability of 88 RPU (Amazon Redshift processing unit)
• The Cloud Information Warehouse Benchmark derived from the TPC-DS 3TB dataset. The next tables have been partitioned on this dataset (the remainder have been unpartitioned):
  • catalog_returns on cr_returned_date_sk
  • catalog_sales on cs_sold_date_sk
  • store_returns on sr_returned_date_sk
  • store_sales on ss_sold_date_sk
  • web_returns on wr_returned_date_sk
  • web_saleson ws_sold_date_sk
  • stock on inv_date_sk

For extra info on loading the Cloud Information Warehouse Benchmark into your Amazon Redshift Serverless workgroup, see the Cloud Information Warehouse Benchmark documentation.

Now, let’s overview how database statistics work and the way they impression question efficiency.

Overview of the impression of planner statistics on question efficiency

To know why database statistics are necessary, first let’s overview what a question planner does. A question planner is the mind of a database: whenever you ship a question to a database, the question planner should decide probably the most environment friendly method to load and compute all the knowledge required to reply the question. Having details about the underlying dataset, reminiscent of statistics in regards to the variety of rows in a dataset, or the distribution of information, might help the question planner generate an optimum plan for retrieving the information. Amazon Redshift makes use of statistics in regards to the underlying knowledge in tables and columns statistics to find out the best way to construct an optimum question execution path.

Let’s see how this works in an instance. Take into account the next question to find out the highest 5 gross sales dates in December 2024 for shops in North America:

SELECT insert_date, sum(gross sales) AS total_sales
FROM receipts
JOIN shops ON shops.id = receipts.store_id
WHERE
  shops.area = 'NAMER' AND
  receipts.insert_date BETWEEN '2024-12-01' AND '2024-12-31'
GROUP BY receipts.insert_date
ORDER BY total_sales DESC
LIMIT 5;

On this question, the question planner has to think about a number of elements, together with:

• Which desk is bigger, shops or receipts? Am I capable of question the smaller desk first to scale back the quantity of looking on the bigger desk?
• Which returns extra rows, receipts.insert_date BETWEEN '2024-12-01' AND '2024-12-31' or shops.area = 'NAMER'?
• Is there any partitioning on the tables? Can I search over a smaller set of information to hurry up the question?

Having details about the underlying knowledge might help to generate an optimum question plan. For instance, shops.area = 'NAMER' would possibly solely return just a few rows (that’s, it’s extremely selective), which means it’s extra environment friendly to execute that step of the question first earlier than filtering by means of the receipts desk. What helps a question planner make this choice is the statistics accessible on columns and tables.

Desk statistics (often known as planner statistics) present a snapshot of the information accessible in a desk to assist the question planner make an knowledgeable choice on execution methods. Databases acquire desk statistics by means of sampling, which entails reviewing a subset of rows to find out the general distribution of information. The standard of statistics, together with the freshness of information, can considerably impression a question plan, which is why databases will reanalyze and regenerate statistics after a sure threshold of the underlying knowledge modifications.

Amazon Redshift helps a number of desk and column degree statistics to help in constructing question plans. These embody:

Open codecs reminiscent of Apache Parquet don’t have any of the previous statistics by default and desk codecs like Apache Iceberg have a subset of the previous statistics reminiscent of variety of rows, NULL depend and min/max values. This could make it difficult for question engines to plan environment friendly queries. Amazon Redshift has added improvements that enhance total question efficiency on knowledge lake knowledge saved in Apache Iceberg and Apache Parquet codecs even when all or partial desk or column-level statistics are unavailable. The following part evaluations options in Amazon Redshift that assist enhance question efficiency on knowledge lakes even when desk statistics aren’t current or are restricted.

Amazon Redshift options when knowledge lakes don’t have statistics for Iceberg tables and Parquet

As talked about beforehand, there are lots of circumstances the place tables saved in knowledge lakes lack statistics, which creates challenges for question engines to make knowledgeable selections on selecting the right question plan. Nonetheless, Amazon Redshift has launched a collection of improvements that enhance efficiency for queries on knowledge lakes even when there aren’t desk statistics accessible. On this part, we overview a few of these enhancements and the way they impression your question efficiency.

Dynamic partition elimination by means of distributed joins

Dynamic partition elimination is a question optimization approach that permits Amazon Redshift to skip studying knowledge unnecessarily throughout question execution on a partitioned desk. It does this by figuring out which partitions of a desk are related to a question and solely scanning these partitions, considerably lowering the quantity of information that must be processed.

For instance, think about a schema that has two tables:

• gross sales (truth desk) with columns:
  • sale_id
  • product_id
  • sale_amount
  • sale_date
• merchandise (dimension desk) with columns:
  • product_id
  • product_name
  • class

The gross sales desk is partitioned by product_id. Within the following instance, you need to discover the whole gross sales quantity for merchandise within the Electronics class in December 2024.

SQL question:

SELECT SUM(s.sale_amount) 
FROM gross sales s
JOIN merchandise p ON s.product_id = p.product_id
WHERE p.class = 'Electronics';

How Amazon Redshift improves this question:

1. Filter on dimension desk:
   • The question filters the merchandise desk to solely embody merchandise within the Electronics class.
2. Determine related partitions:
   • With the brand new enhancements, Amazon Redshift analyzes this filter and determines which partitions of the gross sales desk must be scanned.
   • It appears to be like on the product_id values within the merchandise desk that match the Electronics class and solely scans these particular partitions within the gross sales desk.
   • As an alternative of scanning your complete gross sales desk, Amazon Redshift solely scans the partitions that include gross sales knowledge for electronics merchandise.
   • This considerably reduces the quantity of information Amazon Redshift must course of, making the question sooner.

Beforehand, this optimization was solely utilized on broadcast joins when all baby joins under the be part of have been additionally broadcast joins. The Amazon Redshift staff prolonged this functionality to work on all broadcast joins, regardless if the kid joins under them are broadcast. This permits extra queries to profit from dynamic partition elimination, reminiscent of TPC-DS Q64 and Q75 for Iceberg tables, and TPC-DS Q25 in Parquet.

Metadata caching for Iceberg tables

The Iceberg open desk format employs a two-layer construction: a metadata layer and an information layer. The metadata layer has three ranges of information (metadata.json, manifest lists, and manifests), which permits for efficiency options reminiscent of sooner scan planning and superior knowledge filtering. Amazon Redshift makes use of the Iceberg metadata construction to effectively establish the related knowledge information to scan, utilizing partition worth ranges and column-level statistics and eliminating pointless knowledge processing.

The Amazon Redshift staff noticed that Iceberg metadata is incessantly fetched a number of occasions each inside and throughout queries, resulting in potential efficiency bottlenecks. We applied an in-memory LRU (least not too long ago used) cache for parsed metadata, manifest record information, and manifest information. This cache retains probably the most not too long ago used metadata in order that we keep away from fetching them repeatedly from Amazon Easy Storage Service (Amazon S3) throughout queries. This caching has helped with total efficiency enhancements of as much as 2% in a TPC-DS 3TB workload. We observe greater than 90% cache hits for these metadata buildings, lowering the iceberg metadata processing occasions significantly.

Stats inference for Iceberg tables

As talked about beforehand, the Apache Iceberg file format comes with some statistics reminiscent of variety of rows, variety of nulls, column min/max values and column storage dimension within the metadata information known as manifest information. Nonetheless, they don’t at all times present all of the statistics that we’d like particularly common width which is necessary for the cost-based optimizer utilized by Amazon Redshift.

We delivered a characteristic to estimate common width for variable size columns reminiscent of string and binary from Iceberg metadata. We do that through the use of the column storage dimension and the variety of rows, and we alter for column compression when essential. By inferring these extra statistics, our optimizer could make extra correct price estimates for various question plans. This stats inference characteristic, launched in Amazon Redshift patch 186, gives as much as a 7% enchancment within the TPC-DS benchmarks. We now have additionally enhanced Amazon Redshift optimizer’s price mannequin. The enhancements embody planner optimizations that enhance the estimations of the totally different be part of distribution methods to bear in mind the networking price of distributing the information between the nodes of an Amazon Redshift cluster. The enhancements additionally embody enhancements to Amazon Redshift question optimizer. These enhancements, that are a end result of a number of years of analysis, testing, and implementation demonstrated as much as a forty five% enchancment in a group of TPC-DS benchmarks.

Instance: TPC-DS benchmark highlights on Amazon Redshift no stats queries on knowledge lakes

One method to measure knowledge lake question efficiency for Amazon Redshift is utilizing the TPC-DS benchmark. The TPC-DS benchmark is a standardized benchmark designed to check choice assist techniques, particularly concurrently accessed techniques the place queries can vary from shorter analytical queries (for instance, reporting, dashboards) to longer working ETL-style queries for transferring and remodeling knowledge into a unique system. For these checks, we used the Cloud Information Warehouse Benchmark derived from the TPC-DS 3TB to align our testing with many frequent analytical workloads, and supply a normal set of comparisons to measure enhancements to Amazon Redshift knowledge lake question efficiency.

We ran these checks throughout knowledge saved each within the Apache Parquet knowledge format, along with Apache Iceberg tables with knowledge in Apache Parquet information. As a result of we targeted these checks on out-of-the-box efficiency, none of those knowledge units had any desk statistics accessible. We carried out these checks utilizing the required Amazon Redshift patch variations within the following desk, and used Amazon Redshift Serverless with 88 RPU with none extra tuning. The next outcomes characterize a energy run, which is the sum of how lengthy it took to run all of the checks, from a heat run, that are the outcomes of the facility run after at the very least one execution of the workload:

We noticed notable enhancements in a number of question run occasions. For this put up, we deal with the enhancements we noticed in question 82:

SELECT
    i_brand_id brand_id, i_brand model,
    sum(ss_ext_sales_price) ext_price
FROM date_dim, store_sales, merchandise
WHERE d_date_sk = ss_sold_date_sk AND
    ss_item_sk = i_item_sk AND
    i_manager_id = 83 AND
    d_moy = 12 AND
    d_year = 2002
GROUP BY i_brand, i_brand_id
ORDER BY ext_price desc, i_brand_id
LIMIT 100;

On this question, we’re trying to find the highest 100 promoting manufacturers from a particular supervisor in December 2002, which represents a sometimes dashboard-style analytical question. In our energy run, we noticed a discount in question time from 512 seconds to 18.1 seconds for Apache Parquet knowledge, or a 28.2x enchancment in efficiency. The accelerated question efficiency for this question in a heat run is because of the enhancements to the cost-based optimizer and dynamic partition elimination.

We noticed question efficiency enhancements throughout most of the queries discovered within the Cloud Information Warehouse Benchmark derived from the TPC-DS take a look at suite. We encourage you to strive your individual efficiency checks utilizing Amazon Redshift Serverless in your knowledge lake knowledge to see what efficiency features you may observe.

Cleanup

Should you ran these checks by yourself and don’t want the sources anymore, you’ll have to delete your Amazon Redshift Serverless workgroup. See Shutting down and deleting a cluster. Should you don’t have to retailer the Cloud Information Warehouse Benchmark knowledge in your S3 bucket anymore, see Deleting Amazon S3 objects.

Conclusion

On this put up, you realized how cost-based optimizers for databases work, and the way statistical details about your knowledge might help Amazon Redshift execute queries extra effectively. You’ll be able to optimize question efficiency for Iceberg tables by mechanically amassing Puffin statistics, which lets Amazon Redshift use these current improvements to extra effectively question your knowledge. Giving extra information to your question planner—the mind of Amazon Redshift—helps to supply extra predictable efficiency and lets you additional scale the way you work together along with your knowledge in your knowledge lakes and knowledge lakehouses.

Concerning the authors

Martin Milenkoski is a Software program Growth Engineer on the Amazon Redshift staff, presently specializing in knowledge lake efficiency and question optimization. Martin holds an MSc in Laptop Science from the École Polytechnique Fédérale de Lausanne.

Kalaiselvi Kamaraj is a Sr. Software program Growth Engineer on the Amazon Redshift staff. She has labored on a number of initiatives inside the Amazon Redshift Question processing staff and presently specializing in efficiency associated initiatives for Amazon Redshift DataLake and question optimizer.

Jonathan Katz is a Principal Product Supervisor – Technical on the AWS Analytics staff and is predicated in New York. He’s a Core Workforce member of the open-source PostgreSQL mission and an lively open-source contributor, together with to the pgvector mission.