Monthly Archives: January 2023

Amazon SageMaker is a fully managed machine learning (ML) service. With SageMaker, data scientists and developers can quickly and easily build and train ML models, and then directly deploy them into a production-ready hosted environment. It provides an integrated Jupyter authoring notebook instance for easy access to your data sources for exploration and analysis, so you don’t have to manage servers. It also provides common ML algorithms that are optimized to run efficiently against extremely large data in a distributed environment.

SageMaker real-time inference is ideal for workloads that have real-time, interactive, low-latency requirements. With SageMaker real-time inference, you can deploy REST endpoints that are backed by a specific instance type with a certain amount of compute and memory. Deploying a SageMaker real-time endpoint is only the first step in the path to production for many customers. We want to be able to maximize the performance of the endpoint to achieve a target transactions per second (TPS) while adhering to latency requirements. A large part of performance optimization for inference is making sure you select the proper instance type and count to back an endpoint.

This post describes the best practices for load testing a SageMaker endpoint to find the right configuration for the number of instances and size. This can help us understand the minimum provisioned instance requirements to meet our latency and TPS requirements. From there, we dive into how you can track and understand the metrics and performance of the SageMaker endpoint utilizing Amazon CloudWatch metrics.

We first benchmark the performance of our model on a single instance to identify the TPS it can handle per our acceptable latency requirements. Then we extrapolate the findings to decide on the number of instances we need in order to handle our production traffic. Finally, we simulate production-level traffic and set up load tests for a real-time SageMaker endpoint to confirm our endpoint can handle the production-level load. The entire set of code for the example is available in the following GitHub repository.

Overview of solution

For this post, we deploy a pre-trained Hugging Face DistilBERT model from the Hugging Face Hub. This model can perform a number of tasks, but we send a payload specifically for sentiment analysis and text classification. With this sample payload, we strive to achieve 1000 TPS.

Deploy a real-time endpoint

This post assumes you are familiar with how to deploy a model. Refer to Create your endpoint and deploy your model to understand the internals behind hosting an endpoint. For now, we can quickly point to this model in the Hugging Face Hub and deploy a real-time endpoint with the following code snippet:

# Hub Model configuration. https://huggingface.co/models
hub = {
'HF_MODEL_ID':'distilbert-base-uncased',
'HF_TASK':'text-classification'
}

# create Hugging Face Model Class
huggingface_model = HuggingFaceModel(
transformers_version='4.17.0',
pytorch_version='1.10.2',
py_version='py38',
env=hub,
role=role,
)

# deploy model to SageMaker Inference
predictor = huggingface_model.deploy(
initial_instance_count=1, # number of instances
instance_type='ml.m5.12xlarge' # ec2 instance type
)

Let’s test our endpoint quickly with the sample payload that we want to use for load testing:

import boto3
import json
client = boto3.client('sagemaker-runtime')
content_type = "application/json"
request_body = {'inputs': "I am super happy right now."}
data = json.loads(json.dumps(request_body))
payload = json.dumps(data)
response = client.invoke_endpoint(
EndpointName=predictor.endpoint_name,
ContentType=content_type,
Body=payload)
result = response['Body'].read()
result

Note that we’re backing the endpoint using a single Amazon Elastic Compute Cloud (Amazon EC2) instance of type ml.m5.12xlarge, which contains 48 vCPU and 192 GiB of memory. The number of vCPUs is a good indication of the concurrency the instance can handle. In general, it’s recommended to test different instance types to make sure we have an instance that has resources that are properly utilized. To see a full list of SageMaker instances and their corresponding compute power for real-time Inference, refer to Amazon SageMaker Pricing.

Metrics to track

Before we can get into load testing, it’s essential to understand what metrics to track to understand the performance breakdown of your SageMaker endpoint. CloudWatch is the primary logging tool that SageMaker uses to help you understand the different metrics that describe your endpoint’s performance. You can utilize CloudWatch logs to debug your endpoint invocations; all logging and print statements you have in your inference code are captured here. For more information, refer to How Amazon CloudWatch works.

There are two different types of metrics CloudWatch covers for SageMaker: instance-level and invocation metrics.

Instance-level metrics

The first set of parameters to consider is the instance-level metrics: CPUUtilization and MemoryUtilization (for GPU-based instances, GPUUtilization). For CPUUtilization, you may see percentages above 100% at first in CloudWatch. It’s important to realize for CPUUtilization, the sum of all the CPU cores is being displayed. For example, if the instance behind your endpoint contains 4 vCPUs, this means the range of utilization is up to 400%. MemoryUtilization, on the other hand, is in the range of 0–100%.

Specifically, you can use CPUUtilization to get a deeper understanding of if you have sufficient or even an excess amount of hardware. If you have an under-utilized instance (less than 30%), you could potentially scale down your instance type. Conversely, if you are around 80–90% utilization, it would benefit to pick an instance with greater compute/memory. From our tests, we suggest around 60–70% utilization of your hardware.

Invocation metrics

As suggested by the name, invocation metrics is where we can track the end-to-end latency of any invokes to your endpoint. You can utilize the invocation metrics to capture error counts and what type of errors (5xx, 4xx, and so on) that your endpoint may be experiencing. More importantly, you can understand the latency breakdown of your endpoint calls. A lot of this can be captured with ModelLatency and OverheadLatency metrics, as illustrated in the following diagram.

The ModelLatency metric captures the time that inference takes within the model container behind a SageMaker endpoint. Note that the model container also includes any custom inference code or scripts that you have passed for inference. This unit is captured in microseconds as an invocation metric, and generally you can graph a percentile across CloudWatch (p99, p90, and so on) to see if you’re meeting your target latency. Note that several factors can impact model and container latency, such as the following:

• Custom inference script – Whether you have implemented your own container or used a SageMaker-based container with custom inference handlers, it’s best practice to profile your script to catch any operations that are specifically adding a lot of time to your latency.
• Communication protocol – Consider REST vs. gRPC connections to the model server within the model container.
• Model framework optimizations – This is framework specific, for example with TensorFlow, there are a number of environment variables you can tune that are TF Serving specific. Make sure to check what container you’re using and if there are any framework-specific optimizations you can add within the script or as environment variables to inject in the container.

OverheadLatency is measured from the time that SageMaker receives the request until it returns a response to the client, minus the model latency. This part is largely outside of your control and falls under the time taken by SageMaker overheads.

End-to-end latency as a whole depends on a variety of factors and isn’t necessarily the sum of ModelLatency plus OverheadLatency. For example, if you client is making the InvokeEndpoint API call over the internet, from the client’s perspective, the end-to-end latency would be internet + ModelLatency + OverheadLatency. As such, when load testing your endpoint in order to accurately benchmark the endpoint itself, it’s recommended to focus on the endpoint metrics (ModelLatency, OverheadLatency, and InvocationsPerInstance) to accurately benchmark the SageMaker endpoint. Any issues related to end-to-end latency can then be isolated separately.

A few questions to consider for end-to-end latency:

• Where is the client that is invoking your endpoint?
• Are there any intermediary layers between your client and the SageMaker runtime?

Auto scaling

We don’t cover auto scaling in this post specifically, but it’s an important consideration in order to provision the correct number of instances based on the workload. Depending on your traffic patterns, you can attach an auto scaling policy to your SageMaker endpoint. There are different scaling options, such as TargetTrackingScaling, SimpleScaling, and StepScaling. This allows your endpoint to scale in and out automatically based on your traffic pattern.

A common option is target tracking, where you can specify a CloudWatch metric or custom metric that you have defined and scale out based on that. A frequent utilization of auto scaling is tracking the InvocationsPerInstance metric. After you have identified a bottleneck at a certain TPS, you can often use that as a metric to scale out to a greater number of instances to be able to handle peak loads of traffic. To get a deeper breakdown of auto scaling SageMaker endpoints, refer to Configuring autoscaling inference endpoints in Amazon SageMaker.

Load testing

Although we utilize Locust to display how we can load test at scale, if you’re trying to right size the instance behind your endpoint, SageMaker Inference Recommender is a more efficient option. With third-party load testing tools, you have to manually deploy endpoints across different instances. With Inference Recommender, you can simply pass an array of the instance types you want to load test against, and SageMaker will spin up jobs for each of these instances.

Locust

For this example, we use Locust, an open-source load testing tool that you can implement using Python. Locust is similar to many other open-source load testing tools, but has a few specific benefits:

• Easy to set up – As we demonstrate in this post, we’ll pass a simple Python script that can easily be refactored for your specific endpoint and payload.
• Distributed and scalable – Locust is event-based and utilizes gevent under the hood. This is very useful for testing highly concurrent workloads and simulating thousands of concurrent users. You can achieve high TPS with a single process running Locust, but it also has a distributed load generation feature that enables you to scale out to multiple processes and client machines, as we will explore in this post.
• Locust metrics and UI – Locust also captures end-to-end latency as a metric. This can help supplement your CloudWatch metrics to paint a full picture of your tests. This is all captured in the Locust UI, where you can track concurrent users, workers, and more.

To further understand Locust, check out their documentation.

Amazon EC2 setup

You can set up Locust in whatever environment is compatible for you. For this post, we set up an EC2 instance and install Locust there to conduct our tests. We use a c5.18xlarge EC2 instance. The client-side compute power is also something to consider. At times when you run out of compute power on the client side, this is often not captured, and is mistaken as a SageMaker endpoint error. It’s important to place your client in a location of sufficient compute power that can handle the load that you are testing at. For our EC2 instance, we use an Ubuntu Deep Learning AMI, but you can utilize any AMI as long as you can properly set up Locust on the machine. To understand how to launch and connect to your EC2 instance, refer to the tutorial Get started with Amazon EC2 Linux instances.

The Locust UI is accessible via port 8089. We can open this by adjusting our inbound security group rules for the EC2 Instance. We also open up port 22 so we can SSH into the EC2 instance. Consider scoping the source down to the specific IP address you are accessing the EC2 instance from.

After you’re connected to your EC2 instance, we set up a Python virtual environment and install the open-source Locust API via the CLI:

virtualenv venv #venv is the virtual environment name, you can change as you desire
source venv/bin/activate #activate virtual environment
pip install locust

We’re now ready to work with Locust for load testing our endpoint.

Locust testing

All Locust load tests are conducted based off of a Locust file that you provide. This Locust file defines a task for the load test; this is where we define our Boto3 invoke_endpoint API call. See the following code:

config = Config(
retries = {
'max_attempts': 0,
'mode': 'standard'
}
)

self.sagemaker_client = boto3.client('sagemaker-runtime',config=config)
self.endpoint_name = host.split('/')[-1]
self.region = region
self.content_type = content_type
self.payload = payload

In the preceding code, adjust your invoke endpoint call parameters to suit your specific model invocation. We use the InvokeEndpoint API using the following piece of code in the Locust file; this is our load test run point. The Locust file we’re using is locust_script.py.

def send(self):

request_meta = {
"request_type": "InvokeEndpoint",
"name": "SageMaker",
"start_time": time.time(),
"response_length": 0,
"response": None,
"context": {},
"exception": None,
}
start_perf_counter = time.perf_counter()

try:
response = self.sagemaker_client.invoke_endpoint(
EndpointName=self.endpoint_name,
Body=self.payload,
ContentType=self.content_type
)
response_body = response["Body"].read()

Now that we have our Locust script ready, we want to run distributed Locust tests to stress test our single instance to find out how much traffic our instance can handle.

Locust distributed mode is a little more nuanced than a single-process Locust test. In distributed mode, we have one primary and multiple workers. The primary worker instructs the workers on how to spawn and control the concurrent users that are sending a request. In our distributed.sh script, we see by default that 240 users will be distributed across the 60 workers. Note that the --headless flag in the Locust CLI removes the UI feature of Locust.

#replace with your endpoint name in format https://<<endpoint-name>>
export ENDPOINT_NAME=https://$1

export REGION=us-east-1
export CONTENT_TYPE=application/json
export PAYLOAD='{"inputs": "I am super happy right now."}'
export USERS=240
export WORKERS=60
export RUN_TIME=1m
export LOCUST_UI=false # Use Locust UI

.
.
.

locust -f $SCRIPT -H $ENDPOINT_NAME --master --expect-workers $WORKERS -u $USERS -t $RUN_TIME --csv results &
.
.
.

for (( c=1; c<=$WORKERS; c++ ))
do
locust -f $SCRIPT -H $ENDPOINT_NAME --worker --master-host=localhost &
done

./distributed.sh huggingface-pytorch-inference-2022-10-04-02-46-44-677 #to execute Distributed Locust test

We first run the distributed test on a single instance backing the endpoint. The idea here is we want to fully maximize a single instance to understand the instance count we need to achieve our target TPS while staying within our latency requirements. Note that if you want to access the UI, change the Locust_UI environment variable to True and take the public IP of your EC2 instance and map port 8089 to the URL.

The following screenshot shows our CloudWatch metrics.

Eventually, we notice that although we initially achieve a TPS of 200, we start noticing 5xx errors in our EC2 client-side logs, as shown in the following screenshot.

We can also verify this by looking at our instance-level metrics, specifically CPUUtilization.

Here we notice CPUUtilization at nearly 4,800%. Our ml.m5.12x.large instance has 48 vCPUs (48 * 100 = 4800~). This is saturating the entire instance, which also helps explain our 5xx errors. We also see an increase in ModelLatency.

It seems as if our single instance is getting toppled and doesn’t have the compute to sustain a load past the 200 TPS that we are observing. Our target TPS is 1000, so let’s try to increase our instance count to 5. This might have to be even more in a production setting, because we were observing errors at 200 TPS after a certain point.

We see in both the Locust UI and CloudWatch logs that we have a TPS of nearly 1000 with five instances backing the endpoint.

If you start experiencing errors even with this hardware setup, make sure to monitor CPUUtilization to understand the full picture behind your endpoint hosting. It’s crucial to understand your hardware utilization to see if you need to scale up or even down. Sometimes container-level problems lead to 5xx errors, but if CPUUtilization is low, it indicates that it’s not your hardware but something at the container or model level that might be leading to these issues (proper environment variable for number of workers not set, for example). On the other hand, if you notice your instance is getting fully saturated, it’s a sign that you need to either increase the current instance fleet or try out a larger instance with a smaller fleet.

Although we increased the instance count to 5 to handle 100 TPS, we can see that the ModelLatency metric is still high. This is due to the instances being saturated. In general, we suggest aiming to utilize the instance’s resources between 60–70%.

Clean up

After load testing, make sure to clean up any resources you won’t utilize via the SageMaker console or through the delete_endpoint Boto3 API call. In addition, make sure to stop your EC2 instance or whatever client setup you have to not incur any further charges there as well.

Summary

In this post, we described how you can load test your SageMaker real-time endpoint. We also discussed what metrics you should be evaluating when load testing your endpoint to understand your performance breakdown. Make sure to check out SageMaker Inference Recommender to further understand instance right-sizing and more performance optimization techniques.

About the Authors

Marc Karp is a ML Architect with the SageMaker Service team. He focuses on helping customers design, deploy and manage ML workloads at scale. In his spare time, he enjoys traveling and exploring new places.

Ram Vegiraju is a ML Architect with the SageMaker Service team. He focuses on helping customers build and optimize their AI/ML solutions on Amazon SageMaker. In his spare time, he loves traveling and writing.

Read More

If you’ve had the opportunity to build a search application for unstructured data (i.e., wiki, informational web sites, self-service help pages, internal documentation, etc.) using open source or commercial-off-the-shelf search engines, then you’re probably familiar with the inherent accuracy challenges involved in getting relevant search results. The intended meaning of both query and document can be lost because the search is reduced to matching component keywords and terms. Consequently, while you get results that may contain the right words, they aren’t always relevant to the user. You need your search engine to be smarter so it can rank documents based on matching the meaning or semantics of the content to the intention of the user’s query.

Amazon Kendra provides a fully managed intelligent search service that automates document ingestion and provides highly accurate search and FAQ results based on content across many data sources. If you haven’t migrated to Amazon Kendra and would like to improve the quality of search results, you can use Amazon Kendra Intelligent Ranking for self-managed OpenSearch on your existing search solution.

We’re delighted to introduce the new Amazon Kendra Intelligent Ranking for self-managed OpenSearch, and its companion plugin for the OpenSearch search engine! Now you can easily add intelligent ranking to your OpenSearch document queries, with no need to migrate, duplicate your OpenSearch indexes, or rewrite your applications. The difference between Amazon Kendra Intelligent Ranking for self-managed OpenSearch and the fully managed Amazon Kendra service is that while the former provides powerful semantic re-ranking for the search results, the later provides additional search accuracy improvements and functionality such as incremental learning, question answering, FAQ matching, and built-in connectors. For more information about the fully managed service, please visit the Amazon Kendra service page.

With Amazon Kendra Intelligent Ranking for self-managed OpenSearch, previous results like this:

Query: What is the address of the White House?

Hit1 (best): The president delivered an address to the nation from the White House today.

Hit2: The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500

become like this: 

Query: What is the address of the White House?

Hit1 (best): The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500

Hit2: The president delivered an address to the nation from the White House today.

In this post, we show you how to get started with Amazon Kendra Intelligent Ranking for self-managed OpenSearch, and we provide a few examples that demonstrate the power and value of this feature.

Components of Amazon Kendra Intelligent Ranking for self-managed OpenSearch

• Amazon Kendra Intelligent Ranking application programming interface (API) – The functions from this API are used to perform tasks related to provisioning execution plans and semantic re-ranking of your search results.
• Amazon Kendra Intelligent Ranking plugin for self-managed OpenSearch – This is installed along with your OpenSearch deployment and uses the Rescore function of the Amazon Kendra Intelligent Ranking API to semantically re-rank the search results.
• OpenSearch Dashboards Compare Search Results Plugin – This lets you compare search results from two queries side by side, for example, where one query is a keyword search, while the other query uses Amazon Kendra Intelligent Ranking for self-managed OpenSearch.

Prerequisites

For this tutorial, you’ll need a bash terminal on Linux, Mac, or Windows Subsystem for Linux, and an AWS account. Hint: consider using an Amazon Cloud9 instance or an Amazon Elastic Compute Cloud (Amazon EC2) instance.

You will:

• Install Docker, if it’s not already installed on your system.
• Install the latest AWS Command Line Interface (AWS CLI), if it’s not already installed.
• Create and start OpenSearch containers, with the Amazon Kendra Intelligent Ranking plugin enabled.
• Create test indexes, and load some sample documents.
• Run some queries, with and without intelligent ranking, and be suitably impressed by the differences!

Install Docker

If Docker (i.e., docker and docker-compose) is not already installed in your environment, then install it. See Get Docker for directions.

Install the AWS CLI

If you don’t already have the latest version of the AWS CLI installed, then install and configure it now (see AWS CLI Getting Started). Your default AWS user credentials must have administrator access, or ask your AWS administrator to add the following policy to your user permissions:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "VisualEditor0",
            "Effect": "Allow",
            "Action": "kendra-ranking:*",
            "Resource": "*"
        }
    ]
}

Create and start OpenSearch using the Quickstart script

Download the search_processing_kendra_quickstart.sh script:

wget https://raw.githubusercontent.com/msfroh/search-relevance/quickstart-script/helpers/search_processing_kendra_quickstart.sh
chmod +x search_processing_kendra_quickstart.sh

chmod +x ./search_processing_kendra_quickstart.sh

The quickstart script:

1. Creates an Amazon Kendra Intelligent Ranking Rescore Execution Plan in your AWS account.
2. Creates Docker containers for OpenSearch and its Dashboards.
3. Configures OpenSearch to use the Kendra Intelligent Ranking Service.
4. Starts the OpenSearch services.
5. Provides helpful guidance for using the service.

Use the --help option to see the command line options:

./search_processing_kendra_quickstart.sh --help

Now, execute the script to automate the Amazon Kendra and OpenSearch setup:

./search_processing_kendra_quickstart.sh --create-execution-plan

That’s it! OpenSearch and OpenSearch Dashboard containers are now up and running.

Read the output message from the quickstart script, and make a note of the directory where you can run the handy docker-compose commands, and the cleanup_resources.sh script.

Try a test query to validate you can connect to your OpenSearch container:

curl -XGET --insecure -u 'admin:admin' 'https://localhost:9200'

Note that if you get the error curl(35):OpenSSL SSL_connect: SSL_ERROR_SYSCALL in connection to localhost:9200, it means that OpenSearch is still coming up. Please wait for a couple of minutes for OpenSearch to be ready and try again.

Create test indexes and load sample documents

The script below is used to create an index and load sample documents. Save it on your computer as bulk_post.sh:

#!/bin/bash
curl -u admin:admin -XPOST https://localhost:9200/_bulk --insecure --data-binary @$1 -H 'Content-Type: application/json'

Save the data files below as tinydocs.jsonl:

{ "create" : { "_index" : "tinydocs",  "_id" : "tdoc1" } }
{"title": "WhiteHouse1", "body": "The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500"}
{ "create" : { "_index" : "tinydocs",  "_id" : "tdoc2" } }
{"title": "WhiteHouse2", "body": "The president delivered an address to the nation from the White House today."}

And save the data file below as dstinfo.jsonl:

(This data is adapted from Daylight Saving Time article).

{ "create" : { "_index" : "dstinfo",  "_id" : "dst1" } }
{"title": "Daylight Saving Time", "body": "Daylight saving time begins on the second Sunday in March at 2 a.m., and clocks are set an hour ahead, according to the Farmers’ Almanac. It lasts for eight months and ends on the first Sunday in November, when clocks are set back an hour at 2 a.m."}
{ "create" : { "_index" : "dstinfo",  "_id" : "dst2" } }
{"title":"History of daylight saving time", "body": "Founding Father Benjamin Franklin is often deemed the brain behind daylight saving time after a letter he wrote in 1784 to a Parisian newspaper, according to the Farmers’ Almanac. But Franklin’s letter suggested people simply change their routines and schedules — not the clocks — to the sun’s cycles. Perhaps surprisingly, daylight saving time had a soft rollout in the United States in 1883 to solve issues with railroad accidents, according to the U.S. Bureau of Transportation Services. It was instituted across the United States in 1918, according to the Congressional Research Service. In 2005, Congress changed it to span from March to November instead of its original timeframe of April to October."}
{ "create" : { "_index" : "dstinfo",  "_id" : "dst3" } }
{"title": "Daylight saving time participants", "body":"The United States is one of more than 70 countries that follow some form of daylight saving time, according to World Data. States can individually decide whether or not to follow it, according to the Farmers’ Almanac. Arizona and Hawaii do not, nor do parts of northeastern British Columbia in Canada. Puerto Rico and the Virgin Islands, both U.S. territories, also don’t follow daylight saving time, according to the Congressional Research Service."}
{ "create" : { "_index" : "dstinfo",  "_id" : "dst4" } }
{"title":"Benefits of daylight saving time", "body":"Those in favor of daylight saving time, whether eight months long or permanent, also vouch that it increases tourism in places such as parks or other public attractions, according to National Geographic. The longer days can keep more people outdoors later in the day."}

Make the script executable:

chmod +x ./bulk_post.sh

Now use the bulk_post.sh script to create indexes and load the data by running the two commands below:

./bulk_post.sh tinydocs.jsonl
./bulk_post.sh dstinfo.jsonl

Run sample queries

Prepare query scripts

OpenSearch queries are defined in JSON using the OpenSearch query domain specific language (DSL). For this post, we use the Linux curl command to send queries to our local OpenSearch server using HTTPS.

To make this easy, we’ve defined two small scripts to construct our query DSL and send it to OpenSearch.

The first script creates a regular OpenSearch text match query on two document fields – title and body. See OpenSearch documentation for more on the multi-match query syntax. We’ve kept the query very simple, but you can experiment later with defining alternate types of queries.

Save the script below as query_nokendra.sh:

#!/bin/bash
curl -XGET "https://localhost:9200/$1/_search?pretty" -u 'admin:admin' --insecure -H 'Content-Type: application/json' -d'
  {
    "query": {
      "multi_match": {
        "fields": ["title", "body"],
        "query": "'"$2"'"
      }
    },
    "size": 20
  }
  '

The second script is similar to the first one, but this time we add a query extension to instruct OpenSearch to invoke the Amazon Kendra Intelligent Ranking plugin as a post-processing step to re-rank the original results using the Amazon Kendra Intelligent Ranking service.

The size property determines how many OpenSearch result documents are sent to Kendra for re-ranking. Here, we specify a maximum of 20 results for re-ranking. Two properties, title_field (optional) and body_field (required), specify the document fields used for intelligent ranking.

Save the script below as query_kendra.sh:

#!/bin/bash
curl -XGET "https://localhost:9200/$1/_search?pretty" -u 'admin:admin' --insecure -H 'Content-Type: application/json' -d'
  {
    "query": {
      "multi_match": {
        "fields": ["title", "body"],
        "query": "'"$2"'"
      }
    },
    "size": 20,
    "ext": {
      "search_configuration": {
        "result_transformer": {
          "kendra_intelligent_ranking": {
            "order": 1,
            "properties": {
              "title_field": "title",
              "body_field": "body"
            }
          }
        }
      }
    }
  }
  '

Make both scripts executable:

chmod +x ./query_*kendra.sh

Run initial queries

Start with a simple query on the tinydocs index, to reproduce the example used in the post introduction.

Use the query_nokendra.sh script to search for the address of the White House:

./query_nokendra.sh tinydocs "what is the address of White House"

You see the results shown below. Observe the order of the two results, which are ranked by the score assigned by the OpenSearch text match query. Although the top scoring result does contain the keywords address and White House, it’s clear the meaning doesn’t match the intent of the question. The keywords match, but the semantics do not.

{
  "took" : 2,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 0,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 2,
      "relation" : "eq"
    },
    "max_score" : 1.1619741,
    "hits" : [
      {
        "_index" : "tinydocs",
        "_id" : "tdoc2",
        "_score" : 1.1619741,
        "_source" : {
          "title" : "Whitehouse2",
          "body" : "The president delivered an address to the nation from the White House today."
        }
      },
      {
        "_index" : "tinydocs",
        "_id" : "tdoc1",
        "_score" : 1.0577903,
        "_source" : {
          "title" : "Whitehouse1",
          "body" : "The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500"
        }
      }
    ]
  }
}

Now let’s run the query with Amazon Kendra Intelligent Ranking, using the query_kendra.sh script:

./query_kendra.sh tinydocs "what is the address of White House"

This time, you see the results in a different order as shown below. The Amazon Kendra Intelligent Ranking service has re-assigned the score values, and assigned a higher score to the document that more closely matches the intention of the query. From a keyword perspective, this is a poorer match because it doesn’t contain the word address; however, from a semantic perspective it’s the better response. Now you see the benefit of using the Amazon Kendra Intelligent Ranking plugin!

{
  "took" : 522,
  "timed_out" : false,
  "_shards" : {
    "total" : 1,
    "successful" : 1,
    "skipped" : 2,
    "failed" : 0
  },
  "hits" : {
    "total" : {
      "value" : 2,
      "relation" : "eq"
    },
    "max_score" : 0.3798389,
    "hits" : [
      {
        "_index" : "tinydocs",
        "_id" : "tdoc1",
        "_score" : 0.3798389,
        "_source" : {
          "title" : "Whitehouse1",
          "body" : "The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500"
        }
      },
      {
        "_index" : "tinydocs",
        "_id" : "tdoc2",
        "_score" : 0.25906953,
        "_source" : {
          "title" : "Whitehouse2",
          "body" : "The president delivered an address to the nation from the White House today."
        }
      }
    ]
  }
}

Run additional queries and compare search results

Try the dstinfo index now, to see how the same concept works with different data and queries. While you can use the scripts query_nokendra.sh and query_kendra.sh to make queries from the command line, let’s use instead the OpenSearch Dashboards Compare Search Results Plugin to run queries and compare search results.

Paste the local Dashboards URL into your browser: http://localhost:5601/app/searchRelevance – / to access the dashboard comparison tool. Use the default credentials: Username: admin, Password: admin.

In the search bar, enter: what is daylight saving time?

For the Query 1 and Query 2 index, select dstinfo.

Copy the DSL query below and paste it in the Query panel under Query 1. This is a keyword search query.

{
  "query": { "multi_match": { "fields": ["title", "body"], "query": "%SearchText%" } }, 
  "size": 20
}

Now copy the DSL query below and paste it in the Query panel under Query 2. This query invokes the Amazon Kendra Intelligent Ranking plugin for self-managed OpenSearch to perform semantic re-ranking of the search results.

{
  "query": { "multi_match": { "fields": ["title", "body"], "query": "%SearchText%" } },
  "size": 20,
  "ext": {
    "search_configuration": {
      "result_transformer": {
        "kendra_intelligent_ranking": {
          "order": 1,
          "properties": { "title_field": "title", "body_field": "body" }
        }
      }
    }
  }
}

Choose the Search button to run the queries and observe the search results. In Result 1, the hit ranked last is probably actually the most relevant response to this query. In Result 2, the output from Amazon Kendra Intelligent Ranking has the most relevant answer correctly ranked first.

Now that you have experienced Amazon Kendra Intelligent Ranking for self-managed OpenSearch, experiment with a few queries of your own. Use the data we have already loaded or use the bulk_post.sh script to load your own data.

Explore the Amazon Kendra ranking rescore API

As you’ve seen from this post, the Amazon Kendra Intelligent Ranking plugin for OpenSearch can be conveniently used for semantic re-ranking of your search results. However, if you use a search service that doesn’t support the Amazon Kendra Intelligent Ranking plugin for self-managed OpenSearch, then you can use the Rescore function from the Amazon Kendra Intelligent Ranking API directly.

Try this API using the search results from the example query we used above: what is the address of the White House?

First, find your Execution Plan Id by running:

aws kendra-ranking list-rescore-execution-plans

The JSON below contains the search query, and the two results that were returned by the original OpenSearch match query, with their original OpenSearch scores. Replace {kendra-execution-plan_id} with your Execution Plan Id (from above) and save it as rescore_input.json:

{
    "RescoreExecutionPlanId": "{kendra-execution-plan_id}", 
    "SearchQuery": "what is the address of White House", 
    "Documents": [
        { "Id": "tdoc1",  "Title": "Whitehouse1",  "Body": "The president delivered an address to the nation from the White House today.",  "OriginalScore": 1.4484794 },
        { "Id": "tdoc2",  "Title": "Whitehouse2",  "Body": "The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500",  "OriginalScore": 1.2401118 }
    ]
}

Run the CLI command below to re-score this list of documents using the Amazon Kendra Intelligent Ranking service:

aws kendra-ranking rescore --cli-input-json "`cat rescore_input.json`"

The output of a successful execution of this will look as below.

{
    "ResultItems": [
        {
            "Score": 0.39321771264076233, 
            "DocumentId": "tdoc2"
        }, 
        {
            "Score": 0.328217089176178, 
            "DocumentId": "tdoc1"
        }
    ], 
    "RescoreId": "991459b0-ca9e-4ba8-b0b3-1e8e01f2ad15"
}

As expected, the document tdoc2 (containing the text body “The White House is located at: 1600 Pennsylvania Avenue NW, Washington, DC 20500”) now has the higher ranking, as it’s the semantically more relevant response for the query. The ResultItems list in the output contains each input DocumentId with its new Score, ranked in descending order of Score.

Clean up

When you’re done experimenting, shut down, and remove your Docker containers and Rescore Execution Plan by running the cleanup_resources.sh script created by the Quickstart script, e.g.:

./opensearch-kendra-ranking-docker.xxxx/cleanup_resources.sh

Conclusion

In this post, we showed you how to use Amazon Kendra Intelligent Ranking plugin for self-managed OpenSearch to easily add intelligent ranking to your OpenSearch document queries to dramatically improve the relevance ranking of the results, while using your existing OpenSearch search engine deployments.

You can also use the Amazon Kendra Intelligent Ranking Rescore API directly to intelligently re-score and rank results from your own applications.

Read the Amazon Kendra Intelligent Ranking for self-managed OpenSearch documentation to learn more about this feature, and start planning to apply it in your production applications.

About the Authors

Abhinav Jawadekar is a Principal Solutions Architect focused on Amazon Kendra in the AI/ML language services team at AWS. Abhinav works with AWS customers and partners to help them build intelligent search solutions on AWS.

Bob Strahan is a Principal Solutions Architect in the AWS Language AI Services team.

Read More

Machine learning (ML) applications are complex to deploy and often require the ability to hyper-scale, and have ultra-low latency requirements and stringent cost budgets. Use cases such as fraud detection, product recommendations, and traffic prediction are examples where milliseconds matter and are critical for business success. Strict service level agreements (SLAs) need to be met, and a typical request may require multiple steps such as preprocessing, data transformation, feature engineering, model selection logic, model aggregation, and postprocessing.

Deploying ML models at scale with optimized cost and compute efficiencies can be a daunting and cumbersome task. Each model has its own merits and dependencies based on the external data sources as well as runtime environment such as CPU/GPU power of the underlying compute resources. An application may require multiple ML models to serve a single inference request. In certain scenarios, a request may flow across multiple models. There is no one-size-fits-all approach, and it’s important for ML practitioners to look for tried-and-proven methods to address recurring ML hosting challenges. This has led to the evolution of design patterns for ML model hosting.

In this post, we explore common design patterns for building ML applications on Amazon SageMaker.

Design patterns for building ML applications

Let’s look at the following design patterns to use for hosting ML applications.

Single-model based ML applications

This is a great option when your ML use case requires a single model to serve a request. The model is deployed on a dedicated compute infrastructure with the ability to scale based on the input traffic. This option is also ideal when the client application has a low-latency (in the order of milliseconds or seconds) inference requirement.

Multi-model based ML applications

To make hosting more cost-effective, this design pattern allows you to host multiple models on the same tenant infrastructure. Multiple ML models can share the host or container resources, including caching the most-used ML models in the memory, resulting in better utilization of memory and compute resources. Depending on the types of the models you chose to deploy, model co-hosting may use the following methods:

• Multi-model hosting – This option allows you to host multiple models using a shared serving container on a single endpoint. This feature is ideal when you have a large number of similar models that you can serve through a shared serving container and don’t need to access all the models at the same time.
• Multi-container hosting – This option is ideal when you have multiple models running on different serving stacks with similar resource needs, and when individual models don’t have sufficient traffic to utilize the full capacity of the endpoint instances. Multi-container hosting allows you to deploy multiple containers that use different models or frameworks on a single endpoint. The models can be completely heterogenous, with their own independent serving stack.
• Model ensembles – In a lot of production use cases, there can often be many upstream models feeding inputs to a given downstream model. This is where ensembles are useful. Ensemble patterns involve mixing output from one or more base models in order to reduce the generalization error of the prediction. The base models can be diverse and trained by different algorithms. Model ensembles can out-perform single models because the prediction error of the model decreases when the ensemble approach is used.

The following are common use cases of ensemble patterns and their corresponding design pattern diagrams:

• Scatter-gather – In a scatter-gather pattern, a request for inference is routed to a number of models. An aggregator is then used to collect the responses and distill them into a single inference response. For example, an image classification use case may use three different models to perform the task. The scatter-gather pattern allows you to combine results from inferences run on three different models and pick the most probable classification model.

• Model aggregate – In an aggregation pattern, outputs from multiple models are averaged. For classification models, multiple models’ predictions are evaluated to determine the class that received the most votes and is treated as the final output of the ensemble. For example, in a two-class classification problem to classify a set of fruits as oranges or apples, if two models vote for an orange and one model votes for an apple, then the aggregated output will be an orange. Aggregation helps combat inaccuracy in individual models and makes the output more accurate.

• Dynamic selection – Another pattern for ensemble models is to dynamically perform model selection for the given input attributes. For example, in a given input of images of fruits, if the input contains an orange, model A will be used because it’s specialized for oranges. If the input contains an apple, model B will be used because it’s specialized for apples.

• Serial inference ML applications – With a serial inference pattern, also known as an inference pipeline, use cases have requirements to preprocess incoming data before invoking a pre-trained ML model for generating inferences. Additionally, in some cases, the generated inferences may need to be processed further, so that they can be easily consumed by downstream applications. An inference pipeline allows you to reuse the same preprocessing code used during model training to process the inference request data used for predictions.

• Business logic – Productionizing ML always involves business logic. Business logic patterns involve everything that’s needed to perform an ML task that is not ML model inference. This includes loading the model from Amazon Simple Storage Service (Amazon S3), for example, database lookups to validate the input, obtaining pre-computed features from the feature store, and so on. After these business logic steps are complete, the inputs are passed through to ML models.

ML inference options

For model deployment, it’s important to work backward from your use case. What is the frequency of the prediction? Do you expect live traffic to your application and real-time response to your clients? Do you have many models trained for different subsets of data for the same use case? Does the prediction traffic fluctuate? Is latency of inference a concern? Based on these details, all the preceding design patterns can be implemented using the following deployment options:

• Real-time inference – Real-time inference is ideal for inference workloads where you have real-time, interactive, low-latency requirements. Real-time ML inference workloads may include a single-model based ML application, where an application requires only one ML model to serve a single request, or a multi-model based ML application, where an application requires multiple ML models to serve a single request.
• Near-real-time (asynchronous) inference – With-near-real time inference, you can queue incoming requests. This can be utilized for running inference on inputs that are hundreds of MBs. It operates in near-real time and allows users to use the input for inference, and read the output from the endpoint from an S3 bucket. It can especially be handy in cases with NLP and computer vision, where there are large payloads that require longer preprocessing times.
• Batch inference – Batch inference can be utilized for running inference offline on a large dataset. Because it runs offline, batch inference doesn’t offer the lowest latency. Here, the inference request is processed with either a scheduled or event-based trigger of a batch inference job.
• Serverless inference – Serverless inference is ideal for workloads that have idle periods between traffic spurts and can tolerate a few extra seconds of latency (cold start) for the first invocation after an idle period. For example, a chatbot service or an application to process forms or analyze data from documents. In this case, you might want an online inference option that is able to automatically provision and scale compute capacity based on the volume of inference requests. And during idle time, it should be able to turn off compute capacity completely so that you’re not charged. Serverless inference takes away the undifferentiated heavy lifting of selecting and managing servers by automatically launching compute resources and scaling them in and out depending on traffic.

Use fitness functions to select the right ML inference option

Deciding on the right hosting option is important because it impacts the end-users rendered by your applications. For this purpose, we’re borrowing the concept of fitness functions, which was coined by Neal Ford and his colleagues from AWS Partner ThoughtWorks in their work Building Evolutionary Architectures. Fitness functions provide a prescriptive assessment of various hosting options based on the customer’s objectives. Fitness functions help you obtain the necessary data to allow for the planned evolution of your architecture. They set measurable values to assess how close your solution is to achieving your set goals. Fitness functions can and should be adapted as the architecture evolves to guide a desired change process. This provides architects with a tool to guide their teams while maintaining team autonomy.

There are five main fitness functions that customers care about when it comes to selecting the right ML inference option for hosting their ML models and applications.

Deploying models with SageMaker

SageMaker is a fully managed AWS service that provides every developer and data scientist with the ability to quickly build, train, and deploy ML models at scale. With SageMaker inference, you can deploy your ML models on hosted endpoints and get inference results. SageMaker provides a wide selection of hardware and features to meet your workload requirements, allowing you to select over 70 instance types with hardware acceleration. SageMaker can also provide inference instance type recommendation using a new feature called SageMaker Inference Recommender, in case you’re not sure which one would be most optimal for your workload.

You can choose deployment options to best meet your use cases, such as real time inference, asynchronous, batch, and even serverless endpoints. In addition, SageMaker offers various deployment strategies such as canary, blue/green, shadow, and A/B testing for model deployment, along with cost-effective deployment with multi-model, multi-container endpoints, and elastic scaling. With SageMaker inference, you can view the performance metrics for your endpoints in Amazon CloudWatch, automatically scale endpoints based on traffic, and update your models in production without losing any availability.

SageMaker offers four options to deploy your model so you can start making predictions:

• Real-time inference – This is suitable for workloads with millisecond latency requirements, payload sizes up to 6 MB, and processing times of up to 60 seconds.
• Batch transform – This is ideal for offline predictions on large batches of data that are available up-front.
• Asynchronous inference – This is designed for workloads that don’t have sub-second latency requirements, payload sizes up to 1 GB, and processing times of up to 15 minutes.
• Serverless inference – With serverless inference, you can quickly deploy ML models for inference without having to configure or manage the underlying infrastructure. Additionally, you pay only for the compute capacity used to process inference requests, which is ideal for intermittent workloads.

The following diagram can help you understand the SageMaker hosting model deployment options along with the associated fitness function evaluations.

Let’s explore each of the deployment options in more detail.

Real-time inference in SageMaker

SageMaker real-time inference is recommended if you have sustained traffic and need lower and consistent latency for your requests with payload sizes up to 6 MB, and processing times of up to 60 seconds. You deploy your model to SageMaker hosting services and get an endpoint that can be used for inference. These endpoints are fully managed and support auto scaling. Real-time inference is popular for use cases where you expect a low-latency, synchronous response with predictable traffic patterns, such as personalized recommendations for products and services or transactional fraud detection use cases.

Typically, a client application sends requests to the SageMaker HTTPS endpoint to obtain inferences from a deployed model. You can deploy multiple variants of a model to the same SageMaker HTTPS endpoint. This is useful for testing variations of a model in production. Auto scaling allows you to dynamically adjust the number of instances provisioned for a model in response to changes in your workload.

The following table provides guidance on evaluating SageMaker real-time inference based on the fitness functions.

Asynchronous inference in SageMaker

SageMaker asynchronous inference is a new capability in SageMaker that queues incoming requests and processes them asynchronously. This option is ideal for requests with large payload sizes (up to 1 GB), long processing times (up to 15 minutes), and near-real-time latency requirements. Example workloads for asynchronous inference include healthcare companies processing high-resolution biomedical images or videos like echocardiograms to detect anomalies. These applications receive bursts of incoming traffic at different times in the day and require near-real-time processing at low cost. Processing times for these requests can range in the order of minutes, eliminating the need to run real-time inference. Instead, input payloads can be processed asynchronously from an object store like Amazon S3 with automatic queuing and a predefined concurrency threshold. Upon processing, SageMaker places the inference response in the previously returned Amazon S3 location. You can optionally choose to receive success or error notifications via Amazon Simple Notification Service (Amazon SNS).

The following table provides guidance on evaluating SageMaker asynchronous inference based on the fitness functions.

Batch inference in SageMaker

SageMaker batch transform is ideal for offline predictions on large batches of data that are available up-front. The batch transform feature is a high-performance and high-throughput method for transforming data and generating inferences. It’s ideal for scenarios where you’re dealing with large batches of data, don’t need subsecond latency, or need to both preprocess and transform the training data. Customers in certain domains such as advertising and marketing or healthcare often need to make offline predictions on hyperscale datasets where high throughput is often the objective of the use case and latency isn’t a concern.

When a batch transform job starts, SageMaker initializes compute instances and distributes the inference workload between them. It releases the resources when the jobs are complete, so you pay only for what was used during the run of your job. When the job is complete, SageMaker saves the prediction results in an S3 bucket that you specify. Batch inference tasks are usually good candidates for horizontal scaling. Each worker within a cluster can operate on a different subset of data without the need to exchange information with other workers. AWS offers multiple storage and compute options that enable horizontal scaling. Example workloads for SageMaker batch transform include offline applications such as banking applications for predicting customer churn where an offline job can be scheduled to run periodically.

The following table provides guidance on evaluating SageMaker batch transform based on the fitness functions.

Serverless inference in SageMaker

SageMaker serverless inference allows you to deploy ML models for inference without having to configure or manage the underlying infrastructure. Based on the volume of inference requests your model receives, SageMaker serverless inference automatically provisions, scales, and turns off compute capacity. As a result, you pay for only the compute time to run your inference code and the amount of data processed, not for idle time. You can use SageMaker’s built-in algorithms and ML framework-serving containers to deploy your model to a serverless inference endpoint or choose to bring your own container. If traffic becomes predictable and stable, you can easily update from a serverless inference endpoint to a SageMaker real-time endpoint without the need to make changes to your container image. With serverless inference, you also benefit from other SageMaker features, including built-in metrics such as invocation count, faults, latency, host metrics, and errors in CloudWatch.

The following table provides guidance on evaluating SageMaker serverless inference based on the fitness functions.

Model hosting design patterns in SageMaker

SageMaker inference endpoints use Docker containers for hosting ML models. Containers allow you package software into standardized units that run consistently on any platform that supports Docker. This ensures portability across platforms, immutable infrastructure deployments, and easier change management and CI/CD implementations. SageMaker provides pre-built managed containers for popular frameworks such as Apache MXNet, TensorFlow, PyTorch, Sklearn, and Hugging Face. For a full list of available SageMaker container images, refer to Available Deep Learning Containers Images. In the case that SageMaker doesn’t have a supported container, you can also build your own container (BYOC) and push your own custom image, installing the dependencies that are necessary for your model.

To deploy a model on SageMaker, you need a container (SageMaker managed framework containers or BYOC) and a compute instance to host the container. SageMaker supports multiple advanced options for common ML model hosting design patterns where models can be hosted on a single container or co-hosted on a shared container.

A real-time ML application may use a single model or multiple models to serve a single prediction request. The following diagram shows various inference scenarios for an ML application.

Let’s explore a suitable SageMaker hosting option for each of the preceding inference scenarios. You can refer to the fitness functions to assess if it’s the right option for the given use case.

Hosting a single-model based ML application

There are several options to host single-model based ML applications using SageMaker hosting services depending on the deployment scenario.

Single-model endpoint

SageMaker single-model endpoints allow you to host one model on a container hosted on dedicated instances for low latency and high throughput. These endpoints are fully managed and support auto scaling. You can configure the single-model endpoint as a provisioned endpoint where you pass in endpoint infrastructure configuration such as the instance type and count, or a serverless endpoint where SageMaker automatically launches compute resources and scales them in and out depending on traffic, eliminating the need to choose instance types or manage scaling policies. Serverless endpoints are for applications with intermittent or unpredictable traffic.

The following diagram shows single-model endpoint inference scenarios.

The following table provides guidance on evaluating fitness functions for a provisioned single-model endpoint. For serverless endpoint fitness function evaluations, refer to the serverless endpoint section in this post.

Co-hosting multiple models

When you’re dealing with a large number of models, deploying each one on an individual endpoint with a dedicated container and instance can result in a significant increase in cost. Additionally, it also becomes difficult to manage so many models in production, specifically when you don’t need to invoke all the models at the same time but still need them to be available at all times. Co-hosting multiple models on the same underlying compute resources makes it easy to manage ML deployments at scale and lowers your hosting costs through increased usage of the endpoint and its underlying compute resources. SageMaker supports advanced model co-hosting options such as multi-model endpoint (MME) for homogenous models and multi-container endpoint (MCE) for heterogenous models. Homogeneous models use the same ML framework on a shared service container, whereas heterogenous models allow you to deploy multiple serving containers that use different models or frameworks on a single endpoint.

The following diagram shows model co-hosting options using SageMaker.

SageMaker multi-model endpoints

SageMaker MMEs allow you to host multiple models using a shared serving container on a single endpoint. This is a scalable and cost-effective solution to deploy a large number of models that cater to the same use case, framework, or inference logic. MMEs can dynamically serve requests based on the model invoked by the caller. It also reduces deployment overhead because SageMaker manages loading models in memory and scaling them based on the traffic patterns to them. This feature is ideal when you have a large number of similar models that you can serve through a shared serving container and don’t need to access all the models at the same time. Multi-model endpoints also enable time-sharing of memory resources across your models. This works best when the models are fairly similar in size and invocation latency, allowing MMEs to effectively use the instances across all models. SageMaker MMEs support hosting both CPU and GPU backed models. By using GPU backed models, you can lower your model deployment costs through increased usage of the endpoint and its underlying accelerated compute instances. For a real world use case of MMEs, refer to How to scale machine learning inference for multi-tenant SaaS use cases.

The following table provides guidance on evaluating the fitness functions for MMEs.

SageMaker multi-container endpoints

SageMaker MCEs support deploying up to 15 containers that use different models or frameworks on a single endpoint, and invoking them independently or in sequence for low-latency inference and cost savings. The models can be completely heterogenous, with their own independent serving stack. Securely hosting multiple models from different frameworks on a single instance could save you up to 90% in cost.

The MCE invocation patterns are as follows:

• Inference pipelines – Containers in an MME can be invoked in a linear sequence, also known as a serial inference pipeline. They are typically used to separate preprocessing, model inference, and postprocessing into independent containers. The output from the current container is passed as input to the next. They are represented as a single pipeline model in SageMaker. An inference pipeline can be deployed as an MME, where one of the containers in the pipeline can dynamically serve requests based on the model being invoked.
• Direct invocation – With direct invocation, a request can be sent to a specific inference container hosted on an MCE.

The following table provides guidance on evaluating the fitness functions for MCEs.

Hosting a multi-model based ML application

Many business applications need to use multiple ML models to serve a single prediction request to their consumers. For example, a retail company that wants to provide recommendations to its users. The ML application in this use case may want to use different custom models for recommending different categories of products. If the company wants to add personalization to the recommendations by using individual user information, the number of custom models further increases. Hosting each custom model on a distinct compute instance is not only cost prohibitive, but also leads to underutilization of the hosting resources if not all models are frequently used. SageMaker offers efficient hosting options for multi-model based ML applications.

The following diagram shows multi-model hosting options for a single endpoint using SageMaker.

Serial inference pipeline

An inference pipeline is a SageMaker model that is composed of a linear sequence of 2–15 containers that process requests for inferences on data. You use an inference pipeline to define and deploy any combination of pretrained SageMaker built-in algorithms and your own custom algorithms packaged in Docker containers. You can use an inference pipeline to combine preprocessing, predictions, and postprocessing data science tasks. The output from one container is passed as input to the next. When defining the containers for a pipeline model, you also specify the order in which the containers are run. They are represented as a single pipeline model in SageMaker. The inference pipeline can be deployed as an MME, where one of the containers in the pipeline can dynamically serve requests based on the model being invoked. You can also run a batch transform job with an inference pipeline. Inference pipelines are fully managed.

The following table provides guidance on evaluating the fitness functions for ML model hosting using a serial inference pipeline.

Deploying model ensembles (Triton DAG):

SageMaker offers integration with NVIDIA Triton Inference Server through Triton Inference Server Containers. These containers include NVIDIA Triton Inference Server, support for common ML frameworks, and useful environment variables that let you optimize performance on SageMaker. With NVIDIA Triton container images, you can easily serve ML models and benefit from the performance optimizations, dynamic batching, and multi-framework support provided by NVIDIA Triton. Triton helps maximize the utilization of GPU and CPU, further lowering the cost of inference.

In business use cases where ML applications use several models to serve a prediction request, if each model uses a different framework or is hosted on a separate instance, it may lead to increased workload and cost as well as an increase in overall latency. SageMaker NVIDIA Triton Inference Server supports deployment of models from all major frameworks, such as TensorFlow GraphDef, TensorFlow SavedModel, ONNX, PyTorch TorchScript, TensorRT, and Python/C++ model formats and more. Triton model ensemble represents a pipeline of one or more models or preprocessing and postprocessing logic, and the connection of input and output tensors between them. A single inference request to an ensemble triggers the run of the entire pipeline. Triton also has multiple built-in scheduling and batching algorithms that combine individual inference requests to improve inference throughput. These scheduling and batching decisions are transparent to the client requesting inference. The models can be run on CPUs or GPUs for maximum flexibility and to support heterogeneous computing requirements.

Hosting multiple GPU backed models on multi-model endpoints is supported through the SageMaker Triton Inference Server. The NVIDIA Triton Inference Server has been extended to implement an MME API contract, to integrate with MMEs. You can use the NVIDIA Triton Inference Server, which creates a model repository configuration for different framework backends, to deploy an MME with auto scaling. This feature allows you to scale hundreds of hyper-personalized models that are fine-tuned to cater to unique end-user experiences in AI applications. You can also use this feature to achieve needful price performance for your inference application using fractional GPUs. To learn more, refer to Run multiple deep learning models on GPU with Amazon SageMaker multi-model endpoints.

The following table provides guidance on evaluating the fitness functions for ML model hosting using MMEs with GPU support on Triton inference containers. For single-model endpoints and serverless endpoint fitness function evaluations, refer to the earlier sections in this post.

Best practices

Consider the following best practices:

• High cohesion and low coupling between models – Host the models in the same container that has high cohesion (drives single-business functionality) and encapsulate them together for ease of upgrade and manageability. At the same time, decouple those models from each other (host them in different container) so that you can easily upgrade one model without impacting other models. Host multiple models that use different containers behind one endpoint and invoke then independently or add model preprocessing and postprocessing logic as a serial inference pipeline.
• Inference latency – Group the models that are single-business functionality driven and host them in a single container to minimize the number of hops and therefore minimize the overall latency. There are other caveats, like if the grouped models use multiple frameworks; you might also choose to host in multiple containers but run on the same host to reduce latency and minimize cost.
• Logically group ML models with high cohesion – The logical group may consist of models that are homogeneous (for example, all XGBoost models) or heterogeneous (for example, a few XGBoost and a few BERT). It may consist of models that are shared across multiple business functionalities or may be specific to fulfilling only one business functionality.
  • Shared models – If the logical group consists of shared models, the ease of upgrading the models and latency will play a major role in architecting the SageMaker endpoints. For example, if latency is a priority, it’s better to place all the models in a single container behind a single SageMaker endpoint to avoid multiple hops. The downside is that if any of the models need to be upgraded, it will result in upgrading all the relevant SageMaker endpoints hosting this model.
  • Non-shared models – If the logical group consists of only business feature specific models and is not shared with other groups, the packaging complexity and latency dimensions will become key to achieve. It’s advisable to host these models in a single container behind a single SageMaker endpoint.
• Efficient use of hardware (CPU, GPU) – Group CPU-based models together and host them on the same host so that you can efficiently use the CPU. Similarly, group GPU-based models together so that you can efficiently use and scale them. There are hybrid workloads that require both CPU and GPU on the same host. Hosting the CPU-only and GPU-only models on the same host should be driven by high cohesion and application latency requirements. Additionally, cost, ability to scale, and blast radius on impact in case of failure are the key dimensions to look into.
• Fitness functions – Use fitness functions as a guideline for selecting an ML hosting option.

Conclusion

When it comes to ML hosting, there is no one-size-fits-all approach. ML practitioners need to choose the right design pattern to address their ML hosting challenges. Evaluating the fitness functions provides prescriptive guidance on selecting the right ML hosting option.

For more details on each of the hosting options, refer to the following posts in this series:

• Part 2: Getting started with deploying real-time models on Amazon SageMaker
• Part 3: Run and optimize multi-model inference with Amazon SageMaker multi-model endpoints
• Part 4: Design patterns for serial inference on Amazon SageMaker
• Part 5: Cost efficient ML inference with multi-framework models on Amazon SageMaker
• Part 6: Best practices in testing and updating models on SageMaker
• Part 7: Run ensemble ML models on Amazon SageMaker

About the authors

Dhawal Patel is a Principal Machine Learning Architect at AWS. He has worked with organizations ranging from large enterprises to mid-sized startups on problems related to distributed computing, and Artificial Intelligence. He focuses on Deep learning including NLP and Computer Vision domains. He helps customers achieve high performance model inference on SageMaker.

Deepali Rajale is AI/ML Specialist Technical Account Manager at Amazon Web Services. She works with enterprise customers providing technical guidance on implementing machine learning solutions with best practices. In her spare time, she enjoys hiking, movies and hanging out with family and friends.

Saurabh Trikande is a Senior Product Manager for Amazon SageMaker Inference. He is passionate about working with customers and is motivated by the goal of democratizing machine learning. He focuses on core challenges related to deploying complex ML applications, multi-tenant ML models, cost optimizations, and making deployment of deep learning models more accessible. In his spare time, Saurabh enjoys hiking, learning about innovative technologies, following TechCrunch and spending time with his family.

Read More