AWS S3 Best Practices

S3 Best Practices

Performance

Multiple Concurrent PUTs/GETs

  • S3 scales to support very high request rates. If the request rate grows steadily, S3 automatically partitions the buckets as needed to support higher request rates.
  • S3 can achieve at least 3,500 PUT/COPY/POST/DELETE and 5,500 GET/HEAD requests per second per prefix in a bucket.
  • If the typical workload involves only occasional bursts of 100 requests per second and less than 800 requests per second, AWS scales and handle it.
  • If the typical workload involves a request rate for a bucket to more than 300 PUT/LIST/DELETE requests per second or more than 800 GET requests per second, it’s recommended to open a support case to prepare for the workload and avoid any temporary limits on your request rate.
  • S3 best practice guidelines can be applied only if you are routinely processing 100 or more requests per second
  • Workloads that include a mix of request types
    • If the request workload is typically a mix of GET, PUT, DELETE, or GET Bucket (list objects), choosing appropriate key names for the objects ensures better performance by providing low-latency access to the S3 index
    • This behavior is driven by how S3 stores key names.
      • S3 maintains an index of object key names in each AWS region.
      • Object keys are stored lexicographically (UTF-8 binary ordering) across multiple partitions in the index i.e. S3 stores key names in alphabetical order.
      • Object keys are stored in across multiple partitions in the index and the key name dictates which partition the key is stored in
      • Using a sequential prefix, such as timestamp or an alphabetical sequence, increases the likelihood that S3 will target a specific partition for a large number of keys, overwhelming the I/O capacity of the partition.
    • Introduce some randomness in the key name prefixes, the key names, and the I/O load, will be distributed across multiple index partitions.
    • It also ensures scalability regardless of the number of requests sent per second.

Transfer Acceleration

  • S3 Transfer Acceleration enables fast, easy, and secure transfers of files over long distances between the client and an S3 bucket.
  • Transfer Acceleration takes advantage of CloudFront’s globally distributed edge locations. As the data arrives at an edge location, data is routed to S3 over an optimized network path.

GET-intensive Workloads

  • CloudFront can be used for performance optimization and can help by
    • distributing content with low latency and high data transfer rate.
    • caching the content and thereby reducing the number of direct requests to S3
    • providing multiple endpoints (Edge locations) for data availability
    • available in two flavors as Web distribution or RTMP distribution
  • To fast data transport over long distances between a client and an S3 bucket, use S3 Transfer Acceleration. Transfer Acceleration uses the globally distributed edge locations in CloudFront to accelerate data transport over geographical distances

PUTs/GETs for Large Objects

  • AWS allows Parallelizing the PUTs/GETs request to improve the upload and download performance as well as the ability to recover in case it fails
  • For PUTs, Multipart upload can help improve the uploads by
    • performing multiple uploads at the same time and maximizing network bandwidth utilization
    • quick recovery from failures, as only the part that failed to upload, needs to be re-uploaded
    • ability to pause and resume uploads
    • begin an upload before the Object size is known
  • For GETs, the range HTTP header can help to improve the downloads by
    • allowing the object to be retrieved in parts instead of the whole object
    • quick recovery from failures, as only the part that failed to download needs to be retried.

List Operations

  • Object key names are stored lexicographically in S3 indexes, making it hard to sort and manipulate the contents of LIST
  • S3 maintains a single lexicographically sorted list of indexes
  • Build and maintain Secondary Index outside of S3 for e.g. DynamoDB or RDS to store, index and query objects metadata rather than performing operations on S3

Security

  • Use Versioning
    • can be used to protect from unintended overwrites and deletions
    • allows the ability to retrieve and restore deleted objects or rollback to previous versions
  • Enable additional security by configuring a bucket to enable MFA (Multi-Factor Authentication) to delete
  • Versioning does not prevent Bucket deletion and must be backed up as if accidentally or maliciously deleted the data is lost
  • Use Same Region Replication or Cross Region replication feature to backup data to a different region
  • When using VPC with S3, use VPC S3 endpoints as
    • are horizontally scaled, redundant, and highly available VPC components
    • help establish a private connection between VPC and S3 and the traffic never leaves the Amazon network

Refer blog post @ S3 Security Best Practices

Cost

  • Optimize S3 storage cost by selecting an appropriate storage class for objects
  • Configure appropriate lifecycle management rules to move objects to different storage classes and expire them

Tracking

  • Use Event Notifications to be notified for any put or delete request on the S3 objects
  • Use CloudTrail, which helps capture specific API calls made to S3 from the AWS account and delivers the log files to an S3 bucket
  • Use CloudWatch to monitor the Amazon S3 buckets, tracking metrics such as object counts and bytes stored, and configure appropriate actions

S3 Monitoring and Auditing Best Practices

Refer blog post @ S3 Monitoring and Auditing Best Practices

AWS Certification Exam Practice Questions

  • Questions are collected from Internet and the answers are marked as per my knowledge and understanding (which might differ with yours).
  • AWS services are updated everyday and both the answers and questions might be outdated soon, so research accordingly.
  • AWS exam questions are not updated to keep up the pace with AWS updates, so even if the underlying feature has changed the question might not be updated
  • Open to further feedback, discussion and correction.
  1. A media company produces new video files on-premises every day with a total size of around 100GB after compression. All files have a size of 1-2 GB and need to be uploaded to Amazon S3 every night in a fixed time window between 3am and 5am. Current upload takes almost 3 hours, although less than half of the available bandwidth is used. What step(s) would ensure that the file uploads are able to complete in the allotted time window?
    1. Increase your network bandwidth to provide faster throughput to S3
    2. Upload the files in parallel to S3 using multipart upload
    3. Pack all files into a single archive, upload it to S3, then extract the files in AWS
    4. Use AWS Import/Export to transfer the video files
  2. You are designing a web application that stores static assets in an Amazon Simple Storage Service (S3) bucket. You expect this bucket to immediately receive over 150 PUT requests per second. What should you do to ensure optimal performance?
    1. Use multi-part upload.
    2. Add a random prefix to the key names.
    3. Amazon S3 will automatically manage performance at this scale.
    4. Use a predictable naming scheme, such as sequential numbers or date time sequences, in the key names
  3. You have an application running on an Amazon Elastic Compute Cloud instance, that uploads 5 GB video objects to Amazon Simple Storage Service (S3). Video uploads are taking longer than expected, resulting in poor application performance. Which method will help improve performance of your application?
    1. Enable enhanced networking
    2. Use Amazon S3 multipart upload
    3. Leveraging Amazon CloudFront, use the HTTP POST method to reduce latency.
    4. Use Amazon Elastic Block Store Provisioned IOPs and use an Amazon EBS-optimized instance
  4. Which of the following methods gives you protection against accidental loss of data stored in Amazon S3? (Choose 2)
    1. Set bucket policies to restrict deletes, and also enable versioning
    2. By default, versioning is enabled on a new bucket so you don’t have to worry about it (Not enabled by default)
    3. Build a secondary index of your keys to protect the data (improves performance only)
    4. Back up your bucket to a bucket owned by another AWS account for redundancy
  5. A startup company hired you to help them build a mobile application that will ultimately store billions of image and videos in Amazon S3. The company is lean on funding, and wants to minimize operational costs, however, they have an aggressive marketing plan, and expect to double their current installation base every six months. Due to the nature of their business, they are expecting sudden and large increases to traffic to and from S3, and need to ensure that it can handle the performance needs of their application. What other information must you gather from this customer in order to determine whether S3 is the right option?
    1. You must know how many customers that company has today, because this is critical in understanding what their customer base will be in two years. (No. of customers do not matter)
    2. You must find out total number of requests per second at peak usage.
    3. You must know the size of the individual objects being written to S3 in order to properly design the key namespace. (Size does not relate to the key namespace design but the count does)
    4. In order to build the key namespace correctly, you must understand the total amount of storage needs for each S3 bucket. (S3 provided unlimited storage the key namespace design would depend on the number)
  6. A document storage company is deploying their application to AWS and changing their business model to support both free tier and premium tier users. The premium tier users will be allowed to store up to 200GB of data and free tier customers will be allowed to store only 5GB. The customer expects that billions of files will be stored. All users need to be alerted when approaching 75 percent quota utilization and again at 90 percent quota use. To support the free tier and premium tier users, how should they architect their application?
    1. The company should utilize an amazon simple workflow service activity worker that updates the users data counter in amazon dynamo DB. The activity worker will use simple email service to send an email if the counter increases above the appropriate thresholds.
    2. The company should deploy an amazon relational data base service relational database with a store objects table that has a row for each stored object along with size of each object. The upload server will query the aggregate consumption of the user in questions (by first determining the files store by the user, and then querying the stored objects table for respective file sizes) and send an email via Amazon Simple Email Service if the thresholds are breached. (Good Approach to use RDS but with so many objects might not be a good option)
    3. The company should write both the content length and the username of the files owner as S3 metadata for the object. They should then create a file watcher to iterate over each object and aggregate the size for each user and send a notification via Amazon Simple Queue Service to an emailing service if the storage threshold is exceeded. (List operations on S3 not feasible)
    4. The company should create two separated amazon simple storage service buckets one for data storage for free tier users and another for data storage for premium tier users. An amazon simple workflow service activity worker will query all objects for a given user based on the bucket the data is stored in and aggregate storage. The activity worker will notify the user via Amazon Simple Notification Service when necessary (List operations on S3 not feasible as well as SNS does not address email requirement)
  7. Your company host a social media website for storing and sharing documents. the web application allow users to upload large files while resuming and pausing the upload as needed. Currently, files are uploaded to your php front end backed by Elastic Load Balancing and an autoscaling fleet of amazon elastic compute cloud (EC2) instances that scale upon average of bytes received (NetworkIn) After a file has been uploaded. it is copied to amazon simple storage service(S3). Amazon Ec2 instances use an AWS Identity and Access Management (AMI) role that allows Amazon s3 uploads. Over the last six months, your user base and scale have increased significantly, forcing you to increase the auto scaling groups Max parameter a few times. Your CFO is concerned about the rising costs and has asked you to adjust the architecture where needed to better optimize costs. Which architecture change could you introduce to reduce cost and still keep your web application secure and scalable?
    1. Replace the Autoscaling launch Configuration to include c3.8xlarge instances; those instances can potentially yield a network throughput of 10gbps. (no info of current size and might increase cost)
    2. Re-architect your ingest pattern, have the app authenticate against your identity provider as a broker fetching temporary AWS credentials from AWS Secure token service (GetFederation Token). Securely pass the credentials and s3 endpoint/prefix to your app. Implement client-side logic to directly upload the file to amazon s3 using the given credentials and S3 Prefix. (will not provide the ability to handle pause and restarts)
    3. Re-architect your ingest pattern, and move your web application instances into a VPC public subnet. Attach a public IP address for each EC2 instance (using the auto scaling launch configuration settings). Use Amazon Route 53 round robin records set and http health check to DNS load balance the app request this approach will significantly reduce the cost by bypassing elastic load balancing. (ELB is not the bottleneck)
    4. Re-architect your ingest pattern, have the app authenticate against your identity provider as a broker fetching temporary AWS credentials from AWS Secure token service (GetFederation Token). Securely pass the credentials and s3 endpoint/prefix to your app. Implement client-side logic that used the S3 multipart upload API to directly upload the file to Amazon s3 using the given credentials and s3 Prefix. (multipart allows one to start uploading directly to S3 before the actual size is known or complete data is downloaded)
  8. If an application is storing hourly log files from thousands of instances from a high traffic web site, which naming scheme would give optimal performance on S3?
    1. Sequential
    2. instanceID_log-HH-DD-MM-YYYY
    3. instanceID_log-YYYY-MM-DD-HH
    4. HH-DD-MM-YYYY-log_instanceID (HH will give some randomness to start with instead of instaneId where the first characters would be i-)
    5. YYYY-MM-DD-HH-log_instanceID

Reference

S3_Optimizing_Performance

35 thoughts on “AWS S3 Best Practices

    1. Hi Vamsi, what are you specifically looking for S3 endpoints ?
      S3 endpoints as such allows instances within a VPC to interact with S3 internally, instead of being routed through Internet.
      S3 endpoints are created with VPC and you need to add a route table and associate it with subnet for routing the traffic.
      Also, the ultimate protection still lies with S3 permissions to control who can access the buckets and the objects covered @ S3 Permissions

  1. Hello Jeyander

    How this randomness would help to improve the performance on S3 bucket ?

  2. Hello Jeyandra
    I feel there is a typo in the following:-
    “For GETs, range http header”

    –> quick recovery from failures, as only the part that failed to upload needs to be re-uploaded

    I feel instead of upload, it needs to be download.
    Please confirm.

    1. Thanks Hemal, have corrected the same as GET Range header helps in download.

    1. List Operation is possible with S3 but is not feasible from the performance perspective. As the list operation would be too cumbersome, and it returns only a limited number of objects i.e. 1000 and you would need to do multiple queries.

  3. Question 8
    Correct answer is C as instance id in the front would give the S3 object randomness for improving performance.

    Refer AWS documentation – S3 Log Object Key Format

    Log Object Key Format
    Amazon S3 uses the following object key format for the log objects it uploads in the target bucket:

    TargetPrefixYYYY-mm-DD-HH-MM-SS-UniqueString

    PLEASE CLARIFY

  4. Currently I have an on premise application which writes data into nfs.
    I want to use S3 instead of nfs..is it wise to use S3 in this scenario with a direct connect?
    my data is video/audio contents generated from customers.will there be any performance impact, if we access S3 outside of vpc?what about EFS?

    1. Is any part of the application processing in AWS ? Or are you planning to us only S3 for storage and processing on premises?
      There is surely a performance impact in interacting with S3 instead of NFS which would be in the same data center.
      Direct connect may take time and may not cheap, if you are using it only for storage.

  5. yes, on AWS side there will be servers to consolidate video audio and mp4 conversion etc..and aws looks good for those things to me because of scaling when high load and scalability of S3.and customer get final output from there.. but data first comes to the on premise server..I cant migrate that specific server type to aws because of some dependencies..so my only issue is app should transfer raw data to s3 from on premise server first..

  6. Hi Jayendra,

    Really useful post on S3 🙂

    Can you please elloborate on strategies to validate the impl. of S3 buckets? My project requirement is to upload files via an URL using java SDK; this has been done and need to know what tools or techniques are used in the market to validate services and functioning..

    Thanks

  7. Hello Sir,

    Question No. 4,
    can you please tell me, which one is the second option for this question?
    Is it option d ?

    1. Questions are for the old exams, new exam questions have not been included yet.

  8. Hi jayendra,

    For question 2,why option A is not the right answer. Can you pls explain?

    Thanks Vaishnavi

    1. i would say that right answer is C as this is burst traffic.
      “If the typical workload involves only occasional bursts of 100 requests per second and less than 800 requests per second, AWS scales and handle it”
      and not B as adding hashes to the filenames is needed in other case
      “S3 best practice guidelines can be applied only if you are routinely processing 100 or more requests per second”

  9. Hi jayendra,

    Question # 8, the same question is there in braincert question bank, but the answer is given option c. You have highlighted correct ans as option D Can you confirm me whether it’s c or d.

  10. Hi Jayendar,

    please confirm which one is correct for the question 2 B or C, please clarify

  11. Question 8

    D is useless for PUTS

    Why?

    Because the only time you have to worry about keys and randomness is when the puts exceed 100 per second for PUTS. That means, that when you choose HH-DD-MM as your prefix they are all going to hit the same partition because they are all coming in at the same time.

    Besides it is little better than sequential.

    1. Thanks Drake, i think i agree and HH-DD would not perform better. Instance id would provide the randomness as it is an hourly upload from thousands of instances.

  12. AWS does not recommend adding randomness for performance optimization,
    From AWS developer documentation,
    ” previously Amazon S3 performance guidelines recommended randomizing prefix naming with hashed characters to optimize performance for frequent data retrievals. You no longer have to randomize prefix naming for performance, and can use sequential date-based naming for your prefixes.”

    1. Thats right Anuj, however the exams were not yet updated for the latest enhancements.

  13. Hello, jayendrapatil! Can I translate this post in Russia and post it in my blog with link on your original post?

  14. Hi Jayendra,

    Your scenario based explanation helped me in clarifying myself but I do have a doubt in my scenario.
    I have x number of users and storing images and videos in my s3 bucket through my APIs but my concern is S3 returning me with default url. I have set my s3 bucket as private but object inside is set as public-read for everyone (to show the object to all users) here I’m practically talking about a social network. Is there any any security issue in showing the default s3 url when a user access the object?

    1. you can change the default url by routing it through Route 53 using custom domain, else it will be the direct S3 url.

  15. Regarding policy, if a user need read and write permission to an s3 bucket, is “ListBucket” permission necessary?

    1. from CLI or SDK no, but from console its needed – cause it will list before you can perform any operations.

    1. yes almost 98% content is uptodate. There are certain limitations that are removed, which I am trying to sync as well.

Comments are closed.