Optimize your workloads with Amazon Redshift Serverless AI-driven scaling and optimization

The present scaling method of Amazon Redshift Serverless will increase your compute capability based mostly on the question queue time and scales down when the queuing reduces on the information warehouse. Nevertheless, you may must robotically scale compute sources based mostly on elements like question complexity and information quantity to fulfill price-performance targets, regardless of question queuing. To deal with this requirement, Redshift Serverless launched the synthetic intelligence (AI)-driven scaling and optimization function, which scales the compute not solely based mostly on the queuing, but in addition factoring information quantity and question complexity.

On this publish, we describe how Redshift Serverless makes use of the brand new AI-driven scaling and optimization capabilities to handle frequent use circumstances. This publish additionally consists of instance SQLs, which you’ll run by yourself Redshift Serverless information warehouse to expertise the advantages of this function.

Answer overview

The AI-powered scaling and optimization function in Redshift Serverless offers a user-friendly visible slider to set your required steadiness between value and efficiency. By transferring the slider, you’ll be able to select between optimized for value, balanced efficiency and price, or optimized for efficiency. Primarily based on the place you place the slider, Amazon Redshift will robotically add or take away sources to make sure higher conduct and carry out different AI-driven optimizations like computerized materialized views and computerized desk design optimization to fulfill your chosen price-performance goal.

The slider affords the next choices:

• Optimized for value – Prioritizes value financial savings. Redshift makes an attempt to robotically scale up compute capability when doing so and doesn’t incur extra prices. And it’ll additionally try to scale down compute for decrease value, regardless of longer runtime.
• Balanced – Provides steadiness between efficiency and price. Redshift scales for efficiency with a average value improve.
• Optimized for efficiency – Prioritizes efficiency. Redshift scales aggressively for optimum efficiency, doubtlessly incurring increased prices.

Within the following sections, we illustrate how the AI-driven scaling and optimization function can intelligently predict your workload compute wants and scale proactively for 3 situations:

• Use case 1 – An extended-running complicated question. Compute scales based mostly on question complexity.
• Use case 2 – A sudden spike in ingestion quantity (a three-fold improve, from 720 million to 2.1 billion). Compute scales based mostly on information quantity.
• Use case 3 – A knowledge lake question scanning giant datasets (TBs). Compute scales based mostly on the anticipated information to be scanned from the information lake. The anticipated information scan is predicted by machine studying (ML) fashions based mostly on prior historic run statistics.

Within the current auto scaling mechanism, the use circumstances don’t improve compute capability robotically until queuing is recognized throughout the occasion.

Conditions

To comply with alongside, full the next stipulations:

1. Create a Redshift Serverless workgroup in preview mode. For directions, see Making a preview workgroup.
2. Whereas creating the preview group, select Efficiency and Value Controls and Worth-performance goal, and modify the slider to Optimized for efficiency. For extra data, check with Amazon Redshift provides new AI capabilities, together with Amazon Q, to spice up effectivity and productiveness.
3. Arrange an AWS Id and Entry Administration (IAM) function because the default IAM function. Confer with Managing IAM roles created for a cluster utilizing the console for directions.
4. We use TPC-DS 1TB Cloud Knowledge Warehouse Benchmark information to reveal this function. Run the SQL statements to create tables and cargo the TPC-DS 1TB information.

Use case 1: Scale compute based mostly on question complexity

The next question analyzes product gross sales throughout a number of channels akin to web sites, wholesale, and retail shops. This complicated question usually takes about 25 minutes to run with the default 128 RPUs. Let’s run this workload on the preview workgroup created as a part of stipulations.

When a question is run for the primary time, the AI scaling system might make a suboptimal choice relating to useful resource allocation or scaling because the system remains to be studying the question and information traits. Nevertheless, the system learns from this expertise, and when the identical question is run once more, it may well make a extra optimum scaling choice. Subsequently, if the question didn’t scale through the first run, it is suggested to rerun the question. You possibly can monitor the RPU capability used on the Redshift Serverless console or by querying the SYS_SERVERLSS_USAGE system view.

The outcomes cache is turned off within the following queries to keep away from fetching outcomes from the cache.

SET enable_result_cache_for_session TO off;
with /* TPC-DS demo question */
    ws as
    (choose d_year AS ws_sold_year, ws_item_sk,    ws_bill_customer_sk
     ws_customer_sk,    sum(ws_quantity) ws_qty,    sum(ws_wholesale_cost) ws_wc,
        sum(ws_sales_price) ws_sp   from web_sales   left be part of web_returns on
     wr_order_number=ws_order_number and ws_item_sk=wr_item_sk   be part of date_dim
     on ws_sold_date_sk = d_date_sk   the place wr_order_number is null   group by
     d_year, ws_item_sk, ws_bill_customer_sk   ),
    cs as  
    (choose d_year AS cs_sold_year,
     cs_item_sk,    cs_bill_customer_sk cs_customer_sk,    sum(cs_quantity) cs_qty,
        sum(cs_wholesale_cost) cs_wc,    sum(cs_sales_price) cs_sp   from catalog_sales
       left be part of catalog_returns on cr_order_number=cs_order_number and cs_item_sk=cr_item_sk
       be part of date_dim on cs_sold_date_sk = d_date_sk   the place cr_order_number is
     null   group by d_year, cs_item_sk, cs_bill_customer_sk   ),
    ss as  
    (choose
     d_year AS ss_sold_year, ss_item_sk,    ss_customer_sk,    sum(ss_quantity)
     ss_qty,    sum(ss_wholesale_cost) ss_wc,    sum(ss_sales_price) ss_sp
       from store_sales left be part of store_returns on sr_ticket_number=ss_ticket_number
     and ss_item_sk=sr_item_sk   be part of date_dim on ss_sold_date_sk = d_date_sk
       the place sr_ticket_number is null   group by d_year, ss_item_sk, ss_customer_sk
       ) 
       
       choose 
       ss_customer_sk,spherical(ss_qty/(coalesce(ws_qty+cs_qty,1)),2)
     ratio,ss_qty store_qty, ss_wc store_wholesale_cost, ss_sp store_sales_price,
    coalesce(ws_qty,0)+coalesce(cs_qty,0) other_chan_qty,coalesce(ws_wc,0)+coalesce(cs_wc,0)
     other_chan_wholesale_cost,coalesce(ws_sp,0)+coalesce(cs_sp,0) other_chan_sales_price
    from ss left be part of ws on (ws_sold_year=ss_sold_year and ws_item_sk=ss_item_sk
     and ws_customer_sk=ss_customer_sk)left be part of cs on (cs_sold_year=ss_sold_year
     and cs_item_sk=cs_item_sk and cs_customer_sk=ss_customer_sk)the place coalesce(ws_qty,0)>0
     and coalesce(cs_qty, 0)>0 order by   ss_customer_sk,  ss_qty desc, ss_wc
     desc, ss_sp desc,  other_chan_qty,  other_chan_wholesale_cost,  other_chan_sales_price,
      spherical(ss_qty/(coalesce(ws_qty+cs_qty,1)),2);

When the question is full, run the next SQL to seize the beginning and finish instances of the question, which will probably be used within the subsequent question:

choose query_id,query_text,start_time,end_time, elapsed_time/1000000.0 duration_in_seconds
from sys_query_history
the place query_text like '%TPC-DS demo question%'
and query_text not like '%sys_query_history%'
order by start_time desc

Let’s assess the compute scaled through the previous start_time and end_time interval. Change start_time and end_time within the following question with the output of the previous question:

choose * from sys_serverless_usage
the place end_time >= 'start_time'
and end_time <= DATEADD(minute,1,'end_time')
order by end_time asc

-- Instance
--select * from sys_serverless_usage
--where end_time >= '2024-06-03 00:17:12.322353'
--and end_time <= DATEADD(minute,1,'2024-06-03 00:19:11.553218')
--order by end_time asc

The next screenshot reveals an instance output.

You possibly can discover the rise in compute over the length of this question. This demonstrates how Redshift Serverless scales based mostly on question complexity.

Use case 2: Scale compute based mostly on information quantity

Let’s contemplate the web_sales ingestion job. For this instance, your day by day ingestion job processes 720 million data and completes in a mean of two minutes. That is what you ingested within the prerequisite steps.

As a consequence of some occasion (akin to month finish processing), your volumes elevated by thrice and now your ingestion job must course of 2.1 billion data. In an current scaling method, this is able to improve your ingestion job runtime until the queue time is sufficient to invoke extra compute sources. However with AI-driven scaling, in efficiency optimized mode, Amazon Redshift robotically scales compute to finish your ingestion job inside typical runtimes. This helps shield your ingestion SLAs.

Run the next job to ingest 2.1 billion data into the web_sales desk:

copy web_sales from 's3://redshift-downloads/TPC-DS/2.13/3TB/web_sales/' iam_role default gzip delimiter '|' EMPTYASNULL area 'us-east-1';

Run the next question to check the length of ingesting 2.1 billion data and 720 million data. Each ingestion jobs accomplished in roughly an identical time, regardless of the three-fold improve in quantity.

choose query_id,table_name,data_source,loaded_rows,length/1000000.0 duration_in_seconds , start_time,end_time
from sys_load_history
the place
table_name="web_sales"
order by start_time desc

Run the next question with the beginning instances and finish instances from the earlier output:

choose * from sys_serverless_usage
the place end_time >= 'start_time'
and end_time <= DATEADD(minute,1,'end_time')
order by end_time asc

The next is an instance output. You possibly can discover the rise in compute capability for the ingestion job that processes 2.1 billion data. This illustrates how Redshift Serverless scaled based mostly on information quantity.

Use case 3: Scale information lake queries

On this use case, you create exterior tables pointing to TPC-DS 3TB information in an Amazon Easy Storage Service (Amazon S3) location. Then you definately run a question that scans a big quantity of knowledge to reveal how Redshift Serverless can robotically scale compute capability as wanted.

Within the following SQL, present the ARN of the default IAM function you connected within the stipulations:

-- Create exterior schema
create exterior schema ext_tpcds_3t
from information catalog
database ext_tpcds_db
iam_role '<ARN of the default IAM function connected>'
create exterior database if not exists;

Create exterior tables by operating DDL statements within the following SQL file. You must see seven exterior tables within the question editor underneath the ext_tpcds_3t schema, as proven within the following screenshot.

Run the next question utilizing exterior tables. As talked about within the first use case, if the question didn’t scale through the first run, it is suggested to rerun the question, as a result of the system may have realized from the earlier expertise and may doubtlessly present higher scaling and efficiency for the next run.

The outcomes cache is turned off within the following queries to keep away from fetching outcomes from the cache.

SET enable_result_cache_for_session TO off;

with /* TPC-DS demo information lake question */

ws as
(choose d_year AS ws_sold_year, ws_item_sk, ws_bill_customer_sk
ws_customer_sk,    sum(ws_quantity) ws_qty,    sum(ws_wholesale_cost) ws_wc,
sum(ws_sales_price) ws_sp   from ext_tpcds_3t.web_sales   left be part of ext_tpcds_3t.web_returns on
wr_order_number=ws_order_number and ws_item_sk=wr_item_sk   be part of ext_tpcds_3t.date_dim
on ws_sold_date_sk = d_date_sk   the place wr_order_number is null   group by
d_year, ws_item_sk, ws_bill_customer_sk   ),

cs as
(choose d_year AS cs_sold_year,
cs_item_sk,    cs_bill_customer_sk cs_customer_sk,    sum(cs_quantity) cs_qty,
sum(cs_wholesale_cost) cs_wc,    sum(cs_sales_price) cs_sp   from ext_tpcds_3t.catalog_sales
left be part of ext_tpcds_3t.catalog_returns on cr_order_number=cs_order_number and cs_item_sk=cr_item_sk
be part of ext_tpcds_3t.date_dim on cs_sold_date_sk = d_date_sk   the place cr_order_number is
null   group by d_year, cs_item_sk, cs_bill_customer_sk   ),

ss as
(choose
d_year AS ss_sold_year, ss_item_sk,    ss_customer_sk,    sum(ss_quantity)
ss_qty,    sum(ss_wholesale_cost) ss_wc,    sum(ss_sales_price) ss_sp
from ext_tpcds_3t.store_sales left be part of ext_tpcds_3t.store_returns on sr_ticket_number=ss_ticket_number
and ss_item_sk=sr_item_sk   be part of ext_tpcds_3t.date_dim on ss_sold_date_sk = d_date_sk
the place sr_ticket_number is null   group by d_year, ss_item_sk, ss_customer_sk)

SELECT           ss_customer_sk,spherical(ss_qty/(coalesce(ws_qty+cs_qty,1)),2)
ratio,ss_qty store_qty, ss_wc store_wholesale_cost, ss_sp store_sales_price,
coalesce(ws_qty,0)+coalesce(cs_qty,0) other_chan_qty,coalesce(ws_wc,0)+coalesce(cs_wc,0)    other_chan_wholesale_cost,coalesce(ws_sp,0)+coalesce(cs_sp,0) other_chan_sales_price
FROM ss left be part of ws on (ws_sold_year=ss_sold_year and ws_item_sk=ss_item_sk and ws_customer_sk=ss_customer_sk)left be part of cs on (cs_sold_year=ss_sold_year and cs_item_sk=cs_item_sk and cs_customer_sk=ss_customer_sk)
the place coalesce(ws_qty,0)>0
and coalesce(cs_qty, 0)>0
order by   ss_customer_sk,  ss_qty desc, ss_wc desc, ss_sp desc,  other_chan_qty,  other_chan_wholesale_cost,  other_chan_sales_price,     spherical(ss_qty/(coalesce(ws_qty+cs_qty,1)),2);

Evaluation the entire elapsed time of the question. You want the start_time and end_time from the outcomes to feed into the subsequent question.

choose query_id,query_text,start_time,end_time, elapsed_time/1000000.0 duration_in_seconds
from sys_query_history
the place query_text like '%TPC-DS demo information lake question%'
and query_text not like '%sys_query_history%'
order by start_time desc

Run the next question to see how compute scaled through the previous start_time and end_time interval. Change start_time and end_time within the following question from the output of the previous question:

choose * from sys_serverless_usage
the place end_time >= 'start_time'
and end_time <= DATEADD(minute,1,'end_time')
order by end_time asc

The next screenshot reveals an instance output.

The elevated compute capability for this information lake question reveals that Redshift Serverless can scale to match the information being scanned. This demonstrates how Redshift Serverless can dynamically allocate sources based mostly on question wants.

Issues when selecting your price-performance goal

You need to use the price-performance slider to decide on your required price-performance goal to your workload. The AI-driven scaling and optimizations present holistic optimizations utilizing the next fashions:

• Question prediction fashions – These decide the precise useful resource wants (reminiscence, CPU consumption, and so forth) for every particular person question
• Scaling prediction fashions – These predict how the question would behave on completely different capability sizes

Let’s contemplate a question that takes 7 minutes and prices $7. The next determine reveals the question runtimes and price with no scaling.

A given question may scale in just a few other ways, as proven beneath. Primarily based on the price-performance goal you selected on the slider, AI-driven scaling predicts how the question trades off efficiency and price, and scales it accordingly.

The slider choices yield the next outcomes:

• Optimized for value – If you select Optimized for value, the warehouse scales up if there isn’t a extra value or lesser prices to the person. Within the previous instance, the superlinear scaling method demonstrates this conduct. Scaling will solely happen if it may be finished in an economical method in keeping with the scaling mannequin predictions. If the scaling fashions predict that cost-optimized scaling isn’t doable for the given workload, then the warehouse gained’t scale.
• Balanced – With the Balanced possibility, the system will scale in favor of efficiency and there will probably be a price improve, however it will likely be a restricted improve in value. Within the previous instance, the linear scaling method demonstrates this conduct.
• Optimized for efficiency – With the Optimized for efficiency possibility, the system will scale in favor of efficiency regardless that the prices are increased and non-linear. Within the previous instance, the sublinear scaling method demonstrates this conduct. The nearer the slider place is to the Optimized for efficiency place, the extra sublinear scaling is permitted.

The next are extra factors to notice:

• The worth-performance slider choices are dynamic and they are often modified anytime. Nevertheless, the affect of those adjustments won’t be realized instantly. The affect of that is efficient because the system learns methods to scale the present workload and any extra workloads higher.
• The worth-performance slider choices, Max capability and Max RPU-hours are designed to work collectively. Max capability and Max RPU-hours are the controls to restrict most RPUs the information warehouse allowed to scale and most RPU hours allowed to eat respectively. These controls are all the time honored and enforced whatever the settings on the price-performance goal slider.
• The AI-driven scaling and optimization function dynamically adjusts compute sources to optimize question runtime pace whereas adhering to your price-performance necessities. It considers elements akin to question queueing, concurrency, quantity, and complexity. The system can both run queries on a compute useful resource with decrease concurrent queries or spin up extra compute sources to keep away from queueing. The objective is to supply the perfect price-performance steadiness based mostly in your selections.

Monitoring

You possibly can monitor the RPU scaling within the following methods:

• Evaluation the RPU capability used graph on the Amazon Redshift console.
• Monitor the ComputeCapacity metric underneath AWS/Redshift-Serverless and Workgroup in Amazon CloudWatch.
• Question the SYS_QUERY_HISTORY view, offering the precise question ID or question textual content to determine the time interval. Use this time interval to question the SYS_SERVERLSS_USAGE system view to search out the compute_capacity The compute_capacity discipline will present the RPUs scaled through the question runtime.

Confer with Configure monitoring, limits, and alarms in Amazon Redshift Serverless to maintain prices predictable for the step-by-step directions on utilizing these approaches.

Clear up

Full the next steps to delete the sources you created to keep away from sudden prices:

1. Delete the Redshift Serverless workgroup.
2. Delete the Redshift Serverless related namespace.

Conclusion

On this publish, we mentioned methods to optimize your workloads to scale based mostly on the adjustments in information quantity and question complexity. We demonstrated an method to implement extra responsive, proactive scaling with the AI-driven scaling function in Redshift Serverless. Do this function in your atmosphere, conduct a proof of idea in your particular workloads, and share your suggestions with us.

Concerning the Authors

Satesh Sonti is a Sr. Analytics Specialist Options Architect based mostly out of Atlanta, specialised in constructing enterprise information platforms, information warehousing, and analytics options. He has over 19 years of expertise in constructing information property and main complicated information platform packages for banking and insurance coverage shoppers throughout the globe.

Ashish Agrawal is a Principal Product Supervisor with Amazon Redshift, constructing cloud-based information warehouses and analytics cloud providers. Ashish has over 25 years of expertise in IT. Ashish has experience in information warehouses, information lakes, and platform as a service. Ashish has been a speaker at worldwide technical conferences.

Davide Pagano is a Software program Growth Supervisor with Amazon Redshift based mostly out of Palo Alto, specialised in constructing cloud-based information warehouses and analytics cloud providers options. He has over 10 years of expertise with databases, out of which 6 years of expertise tailor-made to Amazon Redshift.