AWS solutions documentation change
Summary
Added documentation for live race events including SageMaker warm pools, IoT Core connections, Kinesis Video Streams, and pre-event checklist
Security assessment
Changes focus on performance optimization (warm pools), scalability (connection quotas), and event planning guidance. No security vulnerabilities, configurations, or security features are mentioned. The additions are operational best practices without security implications.
Diff
diff --git a/solutions/latest/deepracer-on-aws/quotas.md b/solutions/latest/deepracer-on-aws/quotas.md index 0da8c7449..56a027e24 100644 --- a//solutions/latest/deepracer-on-aws/quotas.md +++ b//solutions/latest/deepracer-on-aws/quotas.md @@ -29 +29 @@ For some services, smaller increases are automatically approved, while larger re -Amazon SageMaker AI training jobs are responsible for running training and evaluation jobs, and also support community races. At the time of writing, the default applied account-level quota value is 8. This means that DeepRacer on AWS can dispatch up to 8 training jobs, evaluation jobs, or race submissions (i.e. evaluating one model in a given race) at a time. If demand exceeds this limit at any point, jobs will remain queued until an actively-running job is completed and capacity becomes available. The queue for jobs in DeepRacer on AWS is managed in FIFO (first-in-first-out) order, and the amount of time that a job spends in the queue depends on the number of jobs that can be processed currently (i.e. the service quota), and the number of jobs that have entered the queue ahead of it. +Amazon SageMaker AI training jobs are responsible for running training and evaluation jobs, and also support community and live races. At the time of writing, the default applied account-level quota value is 8. This means that DeepRacer on AWS can dispatch up to 8 training jobs, evaluation jobs, or race submissions (i.e. evaluating one model in a given race) at a time. If demand exceeds this limit at any point, jobs will remain queued until an actively-running job is completed and capacity becomes available. The queue for jobs in DeepRacer on AWS is managed in FIFO (first-in-first-out) order, and the amount of time that a job spends in the queue depends on the number of jobs that can be processed currently (i.e. the service quota), and the number of jobs that have entered the queue ahead of it. @@ -47,0 +48,23 @@ As a result, it is recommended to consider in advance the number of jobs that ma +#### SageMaker managed warm pools (live race events) + +If you plan to host live race events and want to reduce per-evaluation startup latency, requesting an increase for the SageMaker managed warm pool quota is recommended. The default quota is 0, meaning every evaluation incurs a cold start without an increase. If you decide that you would like to request a service quota increase, you may do so by: + + 1. Accessing the **AWS Management Console** + + 2. Searching for and selecting **Service Quotas** from the search bar at the top + + 3. Selecting **Amazon SageMaker** from the list of services + + 4. Searching for and selecting **ml.c7i.4xlarge for training warm pool usage** in the Service quotas table, and clicking **Request increase at account level** + + 5. Enter the desired warm pool capacity into the **Increase quota value** box. For most live race events, a value of 1–2 is sufficient. + + 6. When you are ready to submit the request, click **Request**. + + + + +###### Note + +Warm pool instances remain allocated for up to 1 hour after the last evaluation completes. If another evaluation starts within that window, the warm pool is reused and the timer resets. Standard SageMaker instance pricing applies while the warm pool is active. + @@ -248,0 +272,31 @@ Consider requesting your quota increase 3–4 weeks in advance. Increases of 500 +### Additional considerations for live race events + +Live races introduce real-time streaming and WebSocket connections on top of the standard SageMaker evaluation workload. The following guidance applies when you are planning a live race event in addition to, or instead of, standard model training and evaluation. + +**Concurrent viewers (AWS IoT Core WebSocket connections)** + +Each user watching a live race holds one persistent WebSocket connection to AWS IoT Core. AWS IoT Core has a default account-level quota of 500,000 concurrent connections, which is unlikely to be a constraint for typical event sizes. However, if your deployment is in a region where a lower limit applies, verify the quota in the [AWS IoT Core service quotas](https://docs.aws.amazon.com/general/latest/gr/iot-core.html) page. + +**Video stream (Amazon Kinesis Video Streams)** + +The live race video stream is delivered via Amazon Kinesis Video Streams as an HLS stream. Each live race uses one active stream. The number of concurrent viewers that a single stream can support is governed by the fragment-metadata and fragment-media quotas, which are soft limits that can be increased via a support request. Review the [Fragment-metadata and fragment-media quotas](https://docs.aws.amazon.com/kinesisvideostreams/latest/dg/limits.html#fragment_based_throttling) to understand the limits and worked examples for your stream configuration before hosting a large event. + +**SageMaker evaluation slots for live races** + +Live race model evaluations use the same SageMaker `ml.c7i.4xlarge` instance quota as training and community race submissions. However, live races run evaluations sequentially (one at a time), so only one SageMaker slot is consumed at any given moment during a live race. The queue size is not constrained by a separate service quota — it is bounded by the number of submitted models. + +For planning purposes, assume approximately 20 minutes per model evaluation. A live race with 30 participants will take roughly 10 hours to complete if run sequentially without pauses. + +**Recommended pre-event checklist** + + * Verify IoT Core concurrent connection quota (especially in opt-in regions). + + * Verify Kinesis Video Streams quota if running more than one simultaneous live race. + + * Confirm your SageMaker training job quota is sufficient to handle any concurrent training and community race activity alongside the live race. + + * Open the submission period well before the event to allow participants to queue models in advance, then close submissions once the event begins. + + + +