Amazon AWS S3

This guide describes the instructions to configure Amazon AWS S3 as Alluxio's under storage system.

Amazon AWS S3, or Amazon Simple Storage Service, is an object storage service offering industry-leading scalability, data availability, security, and performance.

For more information about Amazon AWS S3, please read its documentation.

Prerequisites

If you haven't already, please see Prerequisites before you get started.

In preparation for using Amazon AWS S3 with Alluxio:

<S3_BUCKET>

<S3_DIRECTORY>

The directory you want to use in that container, either by creating a new directory or using an existing one.

<S3_ACCESS_KEY_ID>

<S3_SECRET_KEY>

Basic Setup

To use Amazon AWS S3 as the UFS of Alluxio root mount point, you need to configure Alluxio to use under storage systems by modifying conf/alluxio-site.properties. If it does not exist, create the configuration file from the template.

$ cp conf/alluxio-site.properties.template conf/alluxio-site.properties

Specify an existing S3 bucket and directory as the underfs address by modifying conf/alluxio-site.properties to include:

alluxio.dora.client.ufs.root=s3://<S3_BUCKET>/<S3_DIRECTORY>

Note that if you want to mount the whole s3 bucket, add a trailing slash after the bucket name (e.g. s3://S3_BUCKET/).

Specify the AWS credentials for S3 access by setting s3a.accessKeyId and s3a.secretKey in alluxio-site.properties.

s3a.accessKeyId=<S3_ACCESS_KEY_ID>
s3a.secretKey=<S3_SECRET_KEY>

For other methods of setting AWS credentials, see the credentials section in Advanced Setup.

Advanced Setup

Configure S3 Region

Configure S3 region when accessing S3 buckets to improve performance. Otherwise, global S3 bucket access will be enabled which introduces extra requests. S3 region can be set in conf/alluxio-site.properties

alluxio.underfs.s3.region=us-west-1

Advanced Credentials Setup

You can specify credentials in different ways, from highest to lowest priority:

  1. s3a.accessKeyId and s3a.secretKey specified as mount options

  2. s3a.accessKeyId and s3a.secretKey specified as Java system properties

  3. s3a.accessKeyId and s3a.secretKey in alluxio-site.properties

  4. Environment Variables AWS_ACCESS_KEY_ID or AWS_ACCESS_KEY (either is acceptable) and AWS_SECRET_ACCESS_KEY or AWS_SECRET_KEY (either is acceptable) on the Alluxio servers

  5. Profile file containing credentials at ~/.aws/credentials

  6. AWS Instance profile credentials, if you are using an EC2 instance

When using an AWS Instance profile as the credentials' provider:

  • Create an IAM Role with access to the mounted bucket

  • Create an Instance profile as a container for the defined IAM Role

  • Launch an EC2 instance using the created profile

Note that the IAM role will need access to both the files in the bucket as well as the bucket itself in order to determine the bucket's owner. Automatically assigning an owner to the bucket can be avoided by setting the property alluxio.underfs.s3.inherit.acl=false.

See Amazon's documentation for more details.

Enabling Server Side Encryption

You may encrypt your data stored in S3. The encryption is only valid for data at rest in S3 and will be transferred in decrypted form when read by clients. Note, enabling this will also enable HTTPS to comply with requirements for reading/writing objects.

Enable this feature by configuring conf/alluxio-site.properties:

alluxio.underfs.s3.server.side.encryption.enabled=true

DNS-Buckets

By default, a request directed at the bucket named "mybucket" will be sent to the host name "mybucket.s3.amazonaws.com". You can enable DNS-Buckets to use path style data access, for example: "http://s3.amazonaws.com/mybucket" by setting the following configuration:

alluxio.underfs.s3.disable.dns.buckets=true

Accessing S3 through a proxy

To communicate with S3 through a proxy, modify conf/alluxio-site.properties to include:

alluxio.underfs.s3.proxy.host=<PROXY_HOST>
alluxio.underfs.s3.proxy.port=<PROXY_PORT>

<PROXY_HOST> and <PROXY_PORT> should be replaced by the host and port of your proxy.

Accessing non-global region S3 service

If you want to access a specific region in the AWS service, other than the default us-east-1 region, modify conf/alluxio-site.properties to include:

alluxio.underfs.s3.region=<S3_REGION>

If you want to access a specific endpoint(like AWS VPC endpoint) in a specific region in the AWS service, modify conf/alluxio-site.properties to include:

alluxio.underfs.s3.endpoint=<S3_ENDPOINT>
alluxio.underfs.s3.endpoint.region=<S3_ENDPOINT_REGION>

Both the endpoint and region value need to be updated to use non-global region. And after the setting, alluxio.underfs.s3.region=<S3_REGION> will no longer take effect.

Using a non-Amazon service provider

To use an S3 service provider other than "s3.amazonaws.com", modify conf/alluxio-site.properties to include:

alluxio.underfs.s3.endpoint=<S3_ENDPOINT>
alluxio.underfs.s3.endpoint.region=<S3_ENDPOINT_REGION>

Replace <S3_ENDPOINT> with the hostname and port of your S3 service, e.g., http://localhost:9000. Only use this parameter if you are using a provider other than s3.amazonaws.com.

Connecting to Oracle Cloud Infrastructure (OCI) object storage

Both the endpoint and region value need to be updated to use non-home region.

alluxio.underfs.s3.endpoint=<S3_ENDPOINT>
alluxio.underfs.s3.endpoint.region=<S3_ENDPOINT_REGION>

All OCI object storage regions need to use PathStyleAccess

alluxio.underfs.s3.disable.dns.buckets=true
alluxio.underfs.s3.inherit.acl=false

Using v2 S3 Signatures

Some S3 service providers only support v2 signatures. For these S3 providers, you can enforce using the v2 signatures by setting the alluxio.underfs.s3.signer.algorithm to S3SignerType.

[Experimental] S3 streaming upload

S3 is an object store and because of this feature, the whole file is sent from client to worker, stored in the local disk temporary directory, and uploaded in the close() method by default.

To enable S3 streaming upload, you need to modify conf/alluxio-site.properties to include:

alluxio.underfs.s3.streaming.upload.enabled=true

The default upload process is safer but has the following issues:

  • Slow upload time. The file has to be sent to Alluxio worker first and then Alluxio worker is responsible for uploading the file to S3. The two processes are sequential.

  • The temporary directory must have the capacity to store the whole file.

  • Slow close(). The execution time of close() method is proportional to the file size and inversely proportional to the bandwidth. That is O(FILE_SIZE/BANDWIDTH). Slow close() is unexpected and has already been a bottleneck in the Alluxio Fuse integration. Alluxio Fuse method which calls close() is asynchronous and thus if we write a big file through Alluxio Fuse to S3, the Fuse write operation will be returned much earlier than the file has been written to S3.

The S3 streaming upload feature addresses the above issues and is based on the S3 low-level multipart upload.

The S3 streaming upload has the following advantages:

  • Shorter upload time. Alluxio worker uploads buffered data while receiving new data. The total upload time will be at least as fast as the default method.

  • Smaller capacity requirement. Our data is buffered and uploaded according to partitions (alluxio.underfs.s3.streaming.upload.partition.size which is 64MB by default). When a partition is successfully uploaded, this partition will be deleted.

  • Faster close(). We begin uploading data when data buffered reaches the partition size instead of uploading the whole file in close().

If a S3 streaming upload is interrupted, there may be intermediate partitions uploaded to S3 and S3 will charge for those data.

To reduce the charges, users can modify conf/alluxio-site.properties to include:

alluxio.underfs.cleanup.enabled=true

Intermediate multipart uploads in all non-readonly S3 mount points older than the clean age (configured by alluxio.underfs.s3.intermediate.upload.clean.age) will be cleaned when a leading master starts or a cleanup interval (configured by alluxio.underfs.cleanup.interval) is reached.

[Experimental] S3 multipart upload

The default upload method uploads one file completely from start to end in one go. We use multipart-upload method to upload one file by multiple parts, every part will be uploaded in one thread. It won't generate any temporary files while uploading. It will consume more memory but faster than streaming upload mode.

To enable S3 multipart upload, you need to modify conf/alluxio-site.properties to include:

alluxio.underfs.s3.multipart.upload.enabled=true

There are other parameters you can specify in conf/alluxio-site.properties to make the process faster and better.

# Timeout for uploading part when using multipart upload.
alluxio.underfs.object.store.multipart.upload.timeout
# Multipart upload partition size for S3. The default partition size is `64MB`
alluxio.underfs.s3.multipart.upload.partition.size

Tuning for High Concurrency

When using Alluxio to access S3 with a great number of clients per Alluxio server, these parameters can be tuned so that Alluxio uses a configuration optimized for the S3 backend.

If the S3 connection is slow, a larger timeout is useful:

alluxio.underfs.s3.socket.timeout=500sec
alluxio.underfs.s3.request.timeout=5min

If we expect a great number of concurrent metadata operations:

alluxio.underfs.s3.admin.threads.max=80

If the total number of metadata + data operations is huge:

alluxio.underfs.s3.threads.max=160

For a worker, the number of concurrent writes to S3. For a master, the number of threads to concurrently rename files within a directory.

alluxio.underfs.s3.upload.threads.max=80

Thread-pool size to submit delete and rename operations to S3 on master:

alluxio.underfs.object.store.service.threads=80

Identity and Access Control of S3 Objects

S3 identity and access management is very different from the traditional POSIX permission model. For instance, S3 ACL does not support groups or directory-level settings. Alluxio makes the best effort to inherit permission information including file owner, group and permission mode from S3 ACL information.

Why is 403 Access Denied Error Returned

The S3 credentials set in Alluxio configuration corresponds to an AWS user. If this user does not have the required permissions to access an S3 bucket or object, a 403 permission denied error will be returned.

If you see a 403 error in Alluxio server log when accessing an S3 service, you should double-check

  1. You are using the correct AWS credentials. See credential setup.

  2. Your AWS user has permissions to access the buckets and objects mounted to Alluxio.

Read more AWS troubleshooting guidance for 403 error.

File Owner and Group

Alluxio file system sets the file owner based on the AWS account configured in Alluxio to connect to S3. Since there is no group in S3 ACL, the owner is reused as the group.

By default, Alluxio extracts the display name of this AWS account as the file owner. In case this display name is not available, this AWS user's canonical user ID will be used. This canonical user ID is typically a long string (like 79a59df900b949e55d96a1e698fbacedfd6e09d98eacf8f8d5218e7cd47ef2be), thus often inconvenient to read and use in practice. Optionally, the property alluxio.underfs.s3.owner.id.to.username.mapping can be used to specify a preset mapping from canonical user IDs to Alluxio usernames, in the format "id1=user1;id2=user2". For example, edit alluxio-site.properties to include

alluxio.underfs.s3.owner.id.to.username.mapping=\
79a59df900b949e55d96a1e698fbacedfd6e09d98eacf8f8d5218e7cd47ef2be=john

This configuration helps Alluxio recognize all objects owned by this AWS account as owned by the user john in Alluxio namespace. To find out the AWS S3 canonical ID of your account, check the console https://console.aws.amazon.com/iam/home?#/security_credentials, expand the "Account Identifiers" tab and refer to "Canonical User ID".

Changing Permissions

chown, chgrp, and chmod of Alluxio directories and files do NOT propagate to the underlying S3 buckets nor objects.

Troubleshooting

Enabling AWS-SDK Debug Level

If issues are encountered when running against your S3 backend, enable additional logging to track HTTP traffic. Modify conf/log4j.properties to add the following properties:

log4j.logger.com.amazonaws=WARN
log4j.logger.com.amazonaws.request=DEBUG
log4j.logger.org.apache.http.wire=DEBUG

See Amazon's documentation for more details.

Last updated