S3COMPAT

AIS supports Amazon S3 in two distinct and different ways:

1. On the back, via backend abstraction. Specifically for S3 the corresponding backend implementation currently utilizes AWS SDK for Go;
2. On the client-facing front, AIS provides S3 compatible API, so that existing S3 applications could use AIStore out of the box and without the need to change their (existing) code.

This document talks about the latter - about AIS providing S3 compatible API.

There’s a separate, albeit closely related, document that explains how to configure s3cmd and then maybe tweak AIStore configuration to work with it:

• Getting Started with s3cmd - also contains configuration, tips, usage examples, and more.

For additional background, see:

• High-level AIS block diagram that shows frontend and backend APIs and capabilities.
• Setting custom S3 endpoint can come in handy when a bucket is hosted by an S3 compliant backend (such as, e.g., minio).

Table of Contents

• Table of Contents
• Quick example using aws CLI
  • PUT(object)
  • GET(object)
  • HEAD(object)
• Quick example using Internet Browser
• s3cmd command line
• ETag and MD5
• Last Modification Time
• Multipart Upload using aws
• More Usage Examples
  • Create bucket
  • Remove bucket
  • Upload large object
• TensorFlow Demo
• S3 Compatibility
  • Supported S3
  • Unsupported S3
• Boto3 Compatibility
• Amazon CLI tools

Quick example using aws CLI

The following was tested with an older version of aws CLI, namely:

$ aws --version
aws-cli/1.15.58 Python/3.5.2 Linux/5.4.0-124-generic botocore/1.10.57

You can create buckets and execute PUT/GET verbs, etc.

$ aws --endpoint-url http://localhost:8080/s3 s3 mb s3://abc
make_bucket: abc

PUT(object)

# PUT using AIS CLI:
$ ais put README.md ais://abc

# The same via `aws` CLI:
$ aws --endpoint-url http://localhost:8080/s3 s3api put-object --bucket abc --key LICENSE --body LICENSE
$ ais ls ais://abc
NAME             SIZE
LICENSE          1.05KiB
README.md        10.44KiB

GET(object)

# GET using `aws` CLI:
$ aws --endpoint-url http://localhost:8080/s3 s3api get-object --bucket abc --key README.md /tmp/readme
{
    "ContentType": "text/plain; charset=utf-8",
    "Metadata": {},
    "ContentLength": 10689
}
$ diff -uN README.md /tmp/readme

HEAD(object)

# Get object metadata using `aws` CLI:
$ aws s3api --endpoint-url http://localhost:8080/s3 head-object --bucket abc --key LICENSE
{
    "Metadata": {},
    "ContentLength": 1075,
    "ETag": "f70a21a0c5fa26a93820b0bef5be7619",
    "LastModified": "Mon, 19 Dec 2022 22:23:05 GMT"
}

Quick example using Internet Browser

AIStore gateways provide HTTP/HTTPS interface, which is also why it is maybe sometimes convenient (and very fast) to use your Browser to execute GET type queries.

Specifically - since in this document we are talking about s3-compatible API - here’s an example that utilizes /s3 endpoint to list all buckets:

<ListBucketResult xmlns="http://s3.amazonaws.com/doc/2006-03-01">
  <Owner>
     <ID>1</ID>
     <DisplayName>ListAllMyBucketsResult</DisplayName>
  </Owner>
<Buckets>
  <Bucket>
    <Name>bucket-111</Name>
    <CreationDate>2022-08-23T09:16:40-04:00</CreationDate>
    <String>Provider: ais</String>
  </Bucket>
  <Bucket>
    <Name>bucket-222</Name>
    <CreationDate>2022-08-23T13:47:00-04:00</CreationDate>
    <String>Provider: ais</String>
  </Bucket>
  <Bucket>
  <Name>bucket-222</Name>
    <CreationDate>2022-08-23T13:21:21-04:00</CreationDate>
    <String>Provider: aws (WARNING: {bucket-222, Provider: ais} and {bucket-222, Provider: aws} share the same name)</String>
  </Bucket>
  <Bucket>
    <Name>bucket-333</Name>
    <CreationDate>2022-08-23T13:26:32-04:00</CreationDate>
    <String>Provider: gcp</String>
  </Bucket>
</Buckets>
</ListBucketResult>

Notice the “sharing the same name” warning above. For background, please refer to backend providers.

In re /s3 endpoint mentioned above, the corresponding request URL in the browser’s address bar would look something like ais-gateway-host:port/s3.

s3cmd command line

The following table enumerates some of the s3cmd options that may appear to be useful:

Options	Usage	Example
--host	Define an AIS cluster endpoint	--host=10.10.0.1:51080/s3
--host-bucket	Define URL path to access a bucket of an AIS cluster	--host-bucket="10.10.0.1:51080/s3/%(bucket)"
--no-ssl	Use HTTP instead of HTTPS
--no-check-certificate	Disable checking server’s certificate in case of self-signed ones
--region	Define a bucket region	--region=us-west-1

ETag and MD5

When you are reading an object from Amazon S3, the response will contain ETag.

Amazon S3 ETag is the object’s checksum. Amazon computes those checksums using md5.

On the other hand, the default checksum type that AIS uses is xxhash.

Therefore, it is advisable to:

1. keep in mind this dichotomy, and
2. possibly, configure AIS bucket in question with md5.

Here’s a simple scenario:

Say, an S3-based client performs a GET or a PUT operation and calculates md5 of an object that’s being GET (or PUT). When the operation finishes, the client then compares the checksum with the ETag value in the response header. If checksums differ, the client raises the error “MD5 sum mismatch.”

To enable MD5 checksum at bucket creation time:

$ ais create ais://bck --props="checksum.type=md5"
"ais://bck2" bucket created

$ ais show bucket ais://bck | grep checksum
checksum         Type: md5 | Validate: ColdGET

Or, you can change bucket’s checksum type at any later time:

$ ais bucket props ais://bck checksum.type=md5
Bucket props successfully updated
"checksum.type" set to:"md5" (was:"xxhash")

Please note that changing the bucket’s checksum does not trigger updating (existing) checksums of existing objects - only new writes will be checksummed with the newly configured checksum.

Last Modification Time

AIS tracks object last access time and returns it as LastModified for S3 clients. If an object has never been accessed, which can happen when AIS bucket uses a Cloud bucket as a backend one, zero Unix time is returned.

Example when access time is undefined (not set):

# create AIS bucket with AWS backend bucket (for supported backends and details see docs/providers.md)
$ ais create ais://bck
$ ais bucket props ais://bck backend_bck=aws://bckaws
$ ais bucket props ais://bck checksum.type=md5

# put an object using native ais API and note access time (same as creation time in this case)
$ ais put object.txt ais://bck/obj-ais

# put object with s3cmd - the request bypasses ais, so no access time in the `ls` results
$ s3cmd put object.txt s3://bck/obj-aws --host=localhost:51080 --host-bucket="localhost:51080/s3/%(bucket)"

$ ais ls ais://bck --props checksum,size,atime
NAME            CHECKSUM                                SIZE            ATIME
obj-ais         a103a20a4e8a207fe7ba25eeb2634c96        69.99KiB        08 Dec 20 11:25 PST
obj-aws         a103a20a4e8a207fe7ba25eeb2634c96        69.99KiB

$ s3cmd ls s3://bck --host=localhost:51080 --host-bucket="localhost:51080/s3/%(bucket)"
2020-12-08 11:25     71671   s3://test/obj-ais
1969-12-31 16:00     71671   s3://test/obj-aws

Multipart Upload using aws

Example below reproduces the following Amazon Knowledge-Center instruction.

Used aws-cli/1.15.58 Python/3.5.2 Linux/5.15.0-46-generic botocore/1.10.57

Compare with (user-friendly and easy-to-execute) multipart examples from the s3cmd companion doc.

But first and separately, we create ais:// bucket and configure it with MD5:

$ ais create ais://abc
"ais://abc" created (see https://github.com/NVIDIA/aistore/blob/master/docs/bucket.md#default-bucket-properties)
$ ais bucket props set ais://abc checksum.type=md5
Bucket props successfully updated
"checksum.type" set to: "md5" (was: "xxhash")

Next, the uploading sequence:

# 1. initiate multipart upload
$ aws s3api create-multipart-upload --bucket abc --key large-test-file --endpoint-url http://localhost:8080/s3                                   {
    "Key": "large-test-file",
    "UploadId": "uu3DuXsJG",
    "Bucket": "abc"
}

# 2. upload the first part (w/ upload-id copied from the previous command)
$ aws s3api upload-part --endpoint-url http://localhost:8080/s3 --bucket abc --key large-test-file --part-number 1 --body README.md --upload-id uu3DuXsJG
{
    "ETag": "9bc8111718e22a34f9fa6a099da1f3df"
}

# 3. upload the second, etc. parts
$ aws s3api upload-part --endpoint-url http://localhost:8080/s3 --bucket abc --key large-test-file --part-number 2 --body LICENSE --upload-id uu3DuXsJG
{
    "ETag": "f70a21a0c5fa26a93820b0bef5be7619"
}

# 4. list active upload by its ID (upload-id)
$ aws s3api list-parts --endpoint-url http://localhost:8080/s3 --bucket abc --key large-test-file --upload-id uu3DuXsJG                          {
    "Owner": null,
    "StorageClass": null,
    "Initiator": null,
    "Parts": [
        {
            "PartNumber": 1,
            "ETag": "9bc8111718e22a34f9fa6a099da1f3df",
            "Size": 10725
        },
        {
            "PartNumber": 2,
            "ETag": "f70a21a0c5fa26a93820b0bef5be7619",
            "Size": 1075
        }
    ]
}

And finally:

# 5. complete upload, and be done
$ aws s3api complete-multipart-upload --endpoint-url http://localhost:8080/s3 --bucket abc --key large-test-file --multipart-upload file://up.json  --upload-id uu3DuXsJG
{
    "Key": "large-test-file",
    "Bucket": "abc",
    "ETag": "799e69a43a00794a86eebffb5fbaf4e6-2"
}
$ s3cmd ls s3://abc
2022-08-31 20:36        11800  s3://abc/large-test-file

Notice file://up.json in the complete-multipart-upload command. It simply contains the “Parts” section(**) copied from the “list active upload” step (above).

(**) with no sizes

See https://aws.amazon.com/premiumsupport/knowledge-center/s3-multipart-upload-cli for details.

More Usage Examples

Use any S3 client to access an AIS bucket. Examples below use standard AWS CLI. To access an AIS bucket, one has to pass the correct endpoint to the client. The endpoint is the primary proxy URL and /s3 path, e.g, http://10.0.0.20:51080/s3.

Create bucket

# check that AIS cluster has no buckets, and create a new one
$ ais ls ais://
AIS Buckets (0)
$ s3cmd --host http://localhost:51080/s3 s3 mb s3://bck1
make_bucket: bck1

# list buckets via native CLI
$ ais ls ais://
AIS Buckets (1)

Remove bucket

$ s3cmd --host http://localhost:51080/s3 s3 ls s3://
2020-04-21 16:21:08 bck1

$ s3cmd --host http://localhost:51080/s3 s3 mb s3://bck1
remove_bucket: aws1
$ s3cmd --host http://localhost:51080/s3 s3 ls s3://

Upload large object

In this section, we use all 3 (three) clients:

1. s3cmd client pre-configured to communicate with (and via) AIS
2. aws CLI that sends requests directly to AWS S3 standard endpoint (with no AIS in-between)
3. and, finally, native AIS CLI

# 1. Upload via `s3cmd` => `aistore`

$ s3cmd put $(which aisnode) s3://ais-aa --multipart-chunk-size-mb=8
upload: 'bin/aisnode' -> 's3://ais-aa/aisnode'  [part 1 of 10, 8MB] [1 of 1]
 8388608 of 8388608   100% in    0s   233.84 MB/s  done
...
 8388608 of 8388608   100% in    0s   234.19 MB/s  done
upload: 'bin/aisnode' -> 's3://ais-aa/aisnode'  [part 10 of 10, 5MB] [1 of 1]
 5975140 of 5975140   100% in    0s   233.39 MB/s  done

# 2. View object metadata via native CLI
$ ais show object s3://ais-aa/aisnode --all
PROPERTY         VALUE
atime            30 Aug 54 17:47 LMT
cached           yes
checksum         md5[a38030ea13e1b59c...]
copies           1 [/tmp/ais/mp3/11]
custom           map[ETag:"e3be082db698af7c15b0502f6a88265d-16" source:aws version:3QEKSH7LowuRB2OnUHjWCFsp58aZpsC2]
ec               -
location         t[MKpt8091]:mp[/tmp/ais/mp3/11, nvme0n1]
name             s3://ais-aa/aisnode
size             77.70MiB
version          3QEKSH7LowuRB2OnUHjWCFsp58aZpsC2

# 3. View object metadata via `aws` CLI => directly to AWS (w/ no aistore in-between):
$ aws s3api head-object --bucket ais-aa --key aisnode
{
    "LastModified": "Tue, 20 Dec 2022 17:43:16 GMT",
    "ContentLength": 81472612,
    "Metadata": {
        "x-amz-meta-ais-cksum-type": "md5",
        "x-amz-meta-ais-cksum-val": "a38030ea13e1b59c529e888426001eed"
    },
    "ETag": "\"e3be082db698af7c15b0502f6a88265d-16\"",
    "AcceptRanges": "bytes",
    "ContentType": "binary/octet-stream",
    "VersionId": "3QEKSH7LowuRB2OnUHjWCFsp58aZpsC2"
}

# 4. Finally, view object metadata via `s3cmd` => `aistore`
$ s3cmd info s3://ais-aa/aisnode
s3://ais-aa/aisnode (object):
   File size: 81472612
   Last mod:  Fri, 30 Aug 1754 22:43:41 GMT
   MIME type: none
   Storage:   STANDARD
   MD5 sum:   a38030ea13e1b59c529e888426001eed
   SSE:       none
   Policy:    none
   CORS:      none
   ACL:       none

TensorFlow Demo

Setup S3_ENDPOINT and S3_USE_HTTPS environment variables prior to running a TensorFlow job. S3_ENDPOINT must be primary proxy hostname:port and URL path /s3 (e.g., S3_ENDPOINT=10.0.0.20:51080/s3). Secure HTTP is disabled by default, so S3_USE_HTTPS must be 0.

Example running a training task:

S3_ENDPOINT=10.0.0.20:51080/s3 S3_USE_HTTPS=0 python mnist.py

TensorFlow on AIS training screencast:

S3 Compatibility

AIStore fully supports Amazon S3 API with a few exceptions documented and detailed below. The functionality has been tested using native Amazon S3 clients:

• TensorFlow
• s3cmd
• aws CLI

Speaking of command-line tools, in addition to its own native CLI AIStore also supports Amazon’s s3cmd and aws CLIs. Python-based Amazon S3 clients that will often use Amazon Web Services (AWS) Software Development Kit for Python called Boto3 are also supported - see a note below on AIS <=> Boto3 compatibility.

By way of quick summary, Amazon S3 supports the following API categories:

• Create and delete a bucket
• HEAD bucket
• Get a list of buckets
• PUT, GET, HEAD, and DELETE objects
• Get a list of objects in a bucket (important options include name prefix and page size)
• Copy object within the same bucket or between buckets
• Multi-object deletion
• Get, enable, and disable bucket versioning

and a few more. The following table summarizes S3 APIs and provides the corresponding AIS (native) CLI, as well as s3cmd and aws CLI examples (along with comments on limitations, if any).

See also: a note on AIS <=> Boto3 compatibility.

Supported S3

API	AIS CLI and comments	s3cmd	aws CLI
Create bucket	ais create ais://bck (note: consider using S3 default md5 checksum - see discussion and examples below)	s3cmd mb	aws s3 mb
Head bucket	ais bucket show ais://bck	s3cmd info s3://bck	aws s3api head-bucket
Destroy bucket (aka “remove bucket”)	ais bucket rm ais://bck	s3cmd rb, aws s3 rb
List buckets	ais ls ais:// (or, same: ais ls ais:)	s3cmd ls s3://	aws s3 ls s3://
PUT object	ais put filename ais://bck/obj	s3cmd put ...	aws s3 cp ..
GET object	ais get ais://bck/obj filename	s3cmd get ...	aws s3 cp ..
GET object(range)	ais get ais://bck/obj --offset 0 --length 10	Not supported	aws s3api get-object --range= ..
HEAD object	ais object show ais://bck/obj	s3cmd info s3://bck/obj	aws s3api head-object
List objects in a bucket	ais ls ais://bck	s3cmd ls s3://bucket-name/	aws s3 ls s3://bucket-name/
Copy object in a given bucket or between buckets	S3 API is fully supported; we have yet to implement our native CLI to copy objects (we do copy buckets, though)	Limited support: s3cmd performs GET followed by PUT instead of AWS API call	aws s3api copy-object ... calls copy object API
Last modification time	AIS always stores only one - the last - version of an object. Therefore, we track creation and last access time but not “modification time”.	-	-
Bucket creation time	ais bucket show ais://bck	s3cmd displays creation time via ls subcommand: s3cmd ls s3://	-
Versioning	AIS tracks and updates versioning information but only for the latest object version. Versioning is enabled by default; to disable, run: ais bucket props ais://bck versioning.enabled=false	-	aws s3api get/put-bucket-versioning
ACL	Limited support; AIS provides an extensive set of configurable permissions - see ais bucket props ais://bck access and ais auth and the corresponding documentation	-	-
Multipart upload(**)	- (added in v3.12)	s3cmd put ... s3://bck --multipart-chunk-size-mb=5	aws s3api create-multipart-upload --bucket abc ...

(**) With the only exception of UploadPartCopy operation.

Unsupported S3

• Amazon Regions (us-east-1, us-west-1, etc.)
• Retention Policy
• CORS
• Website endpoints
• CloudFront CDN

Boto3 Compatibility

Arguably, extremely few HTTP client-side libraries do not follow HTTP redirects, and Amazon’s botocore, used by Boto3, just happens to be one of those (libraries).

AIStore provides a shim that you can use to alter botocore and boto3’s behavior to work as expected with AIStore.

To use boto3 or botocore as client libraries for AIStore:

• Install the aistore python package with the botocore extra.

$ pip install aistore[botocore]

• Import aistore.botocore_patch.botocore in your source code alongside botocore and / or boto3.

import boto3
from aistore.botocore_patch import botocore

For more context, see perhaps the following aws-cli ticket and discussion at:

• Support S3 HTTP redirects to non-Amazon URI’s

Amazon CLI tools

As far as existing Amazon-native CLI tools, s3cmd would be the preferred and recommended option. Please see s3cmd readme for usage examples and a variety of topics, including:

• s3cmd Configuration
• Getting Started
  • 1. AIS Endpoint
  • 2. How to have s3cmd calling AIS endpoint
  • 3. Alternatively
  • 4. Note and, possibly, update AIS configuration
  • 5. Create bucket and PUT/GET objects using s3cmd
  • 6. Multipart upload using s3cmd