Object Versioning and Concurrency Control

1. Object Versioning and Concurrency Control
   1. Overview
   2. Object Versioning
   3. Copying Versioned Buckets
   4. Concurrency Control
   5. For More Information

Object Versioning and Concurrency Control

Overview

Versioning-enabled buckets maintain an archive of objects, providing a way to un-delete data that you accidentally deleted, or to retrieve older versions of your data. You can turn versioning on or off for a bucket at any time. Turning versioning off leaves existing object versions in place, and simply causes the bucket to stop accumulating new object versions. In this case, if you upload to an existing object the current version is overwritten instead of creating a new version.

Regardless of whether you have enabled versioning on a bucket, every object has two associated positive integer fields:

• the generation, which is updated when the content of an object is overwritten.
• the metageneration, which identifies the metadata generation. It starts at 1; is updated every time the metadata (e.g., ACL or Content-Type) for a given content generation is updated; and gets reset when the generation number changes.

Of these two integers, only the generation is used when working with versioned data. Both generation and metageneration can be used with concurrency control (discussed in a later section).

To work with object versioning in gsutil, you can use a flavor of storage URLs that embed the object generation, which we refer to as version-specific URLs. For example, the version-less object URL:

gs://bucket/object

might have have two versions, with these version-specific URLs:

gs://bucket/object#1360383693690000
gs://bucket/object#1360383802725000

The following sections discuss how to work with versioning and concurrency control.

Object Versioning

You can view, enable, and disable object versioning on a bucket using the 'versioning get' and 'versioning set' commands. For example:

gsutil versioning set on gs://bucket

will enable versioning for the named bucket. See gsutil help versioning for additional details.

To see all object versions in a versioning-enabled bucket along with their generation.metageneration information, use gsutil ls -a:

gsutil ls -a gs://bucket

You can also specify particular objects for which you want to find the version-specific URL(s), or you can use wildcards:

gsutil ls -a gs://bucket/object1 gs://bucket/images/*.jpg

The generation values form a monotonically increasing sequence as you create additional object versions. Because of this, the latest object version is always the last one listed in the gsutil ls output for a particular object. For example, if a bucket contains these three versions of gs://bucket/object:

gs://bucket/object#1360035307075000
gs://bucket/object#1360101007329000
gs://bucket/object#1360102216114000

then gs://bucket/object#1360102216114000 is the latest version and gs://bucket/object#1360035307075000 is the oldest available version.

If you specify version-less URLs with gsutil, you will operate only on the live version of an object, for example:

gsutil cp gs://bucket/object ./dir

or:

gsutil rm gs://bucket/object

The same is true when using wildcards like * and **. These will operate only on the live version of matching objects. For example, this command will remove the live version and create an archived version for each object in a bucket:

gsutil rm gs://bucket/**

To operate on a specific object version, use a version-specific URL. For example, suppose the output of the above gsutil ls -a command is:

gs://bucket/object#1360035307075000
gs://bucket/object#1360101007329000

In this case, the command:

gsutil cp gs://bucket/object#1360035307075000 ./dir

will retrieve the second most recent version of the object.

Note that version-specific URLs cannot be the target of the gsutil cp command (trying to do so will result in an error), because writing to a versioned object always creates a new version.

Note also that some shells treat "#" as a special character (e.g., zsh with the extendedglob option enabled). If you're using a shell that treats "#" as a special character you will need to surround the argument with quotes, such as:

gsutil cp 'gs://bucket/object#1360035307075000' ./dir

If an object has been deleted, it will not show up in a normal gsutil ls listing (i.e., ls without the -a option). You can restore a deleted object by running gsutil ls -a to find the available versions, and then copying one of the version-specific URLs to the version-less URL, for example:

gsutil cp gs://bucket/object#1360101007329000 gs://bucket/object

Note that when you do this it creates a new object version, which will incur additional charges. You can get rid of the extra copy by deleting the older version-specfic object:

gsutil rm gs://bucket/object#1360101007329000

Or you can combine the two steps by using the gsutil mv command:

gsutil mv gs://bucket/object#1360101007329000 gs://bucket/object

If you remove the live version of an object in a versioning-enabled bucket, an archived version will be preserved:

gsutil rm gs://bucket/object

If you remove a version-specific URL for an object (even if it is the live version), that version will be deleted permanently:

gsutil rm gs://bucket/object#1360101007329000

If you want to remove all versions of an object, use the gsutil rm -a option:

gsutil rm -a gs://bucket/object

If you want to remove all versions of all objects in a bucket (and the bucket itself), use the rm -r option (-r implies the -a option):

gsutil rm -r gs://bucket

Note that there is no limit to the number of older versions of an object you will create if you continue to upload to the same object in a versioning- enabled bucket. It is your responsibility to delete versions beyond the ones you want to retain.

Copying Versioned Buckets

You can copy data between two versioned buckets, using a command like:

gsutil cp -r -A gs://bucket1/* gs://bucket2

When run using versioned buckets, this command will cause every object version to be copied. The copies made in gs://bucket2 will have different generation numbers (since a new generation is assigned when the object copy is made), but the object sort order will remain consistent. For example, gs://bucket1 might contain:

% gsutil ls -la gs://bucket1 10  2013-06-06T02:33:11Z
53  2013-02-02T22:30:57Z  gs://bucket1/file#1359844257574000  metageneration=1
12  2013-02-02T22:30:57Z  gs://bucket1/file#1359844257615000  metageneration=1
97  2013-02-02T22:30:57Z  gs://bucket1/file#1359844257665000  metageneration=1

and after the copy, gs://bucket2 might contain:

% gsutil ls -la gs://bucket2
53  2013-06-06T02:33:11Z  gs://bucket2/file#1370485991580000  metageneration=1
12  2013-06-06T02:33:14Z  gs://bucket2/file#1370485994328000  metageneration=1
97  2013-06-06T02:33:17Z  gs://bucket2/file#1370485997376000  metageneration=1

Note that the object versions are in the same order (as can be seen by the same sequence of sizes in both listings), but the generation numbers (and timestamps) are newer in gs://bucket2.

Concurrency Control

If you are building an application using Cloud Storage, you may need to be careful about concurrency control. Normally gsutil itself isn't used for this purpose, but it's possible to write scripts around gsutil that perform concurrency control.

For example, suppose you want to implement a "rolling update" system using gsutil, where a periodic job computes some data and uploads it to the cloud. On each run, the job starts with the data that it computed from last run, and computes a new value. To make this system robust, you need to have multiple machines on which the job can run, which raises the possibility that two simultaneous runs could attempt to update an object at the same time. This leads to the following potential race condition:

• job 1 computes the new value to be written
• job 2 computes the new value to be written
• job 2 writes the new value
• job 1 writes the new value

In this case, the value that job 1 read is no longer current by the time it goes to write the updated object, and writing at this point would result in stale (or, depending on the application, corrupt) data.

To prevent this, you can find the version-specific name of the object that was created, and then use the information contained in that URL to specify an x-goog-if-generation-match header on a subsequent gsutil cp command. You can do this in two steps. First, use the gsutil cp -v option at upload time to get the version-specific name of the object that was created, for example:

gsutil cp -v file gs://bucket/object

might output:

Created: gs://bucket/object#1360432179236000

You can extract the generation value from this object and then construct a subsequent gsutil command like this:

gsutil -h x-goog-if-generation-match:1360432179236000 cp newfile \
    gs://bucket/object

This command requests Cloud Storage to attempt to upload newfile but to fail the request if the generation of newfile that is live at the time of the upload does not match that specified.

If the command you use updates object metadata, you will need to find the current metageneration for an object. To do this, use the gsutil ls -a and -l options. For example, the command:

gsutil ls -l -a gs://bucket/object

will output something like:

  64  2013-02-12T19:59:13Z  gs://bucket/object#1360699153986000  metageneration=3
1521  2013-02-13T02:04:08Z  gs://bucket/object#1360721048778000  metageneration=2

Given this information, you could use the following command to request setting the ACL on the older version of the object, such that the command will fail unless that is the current version of the data+metadata:

gsutil -h x-goog-if-generation-match:1360699153986000 -h \
  x-goog-if-metageneration-match:3 acl set public-read \
  gs://bucket/object#1360699153986000

Without adding these headers, the update would simply overwrite the existing ACL. Note that in contrast, the "gsutil acl ch" command uses these headers automatically, because it performs a read-modify-write cycle in order to edit ACLs.

If you want to experiment with how generations and metagenerations work, try the following. First, upload an object; then use gsutil ls -l -a to list all versions of the object, along with each version's metageneration; then re- upload the object and repeat the gsutil ls -l -a. You should see two object versions, each with metageneration=1. Now try setting the ACL, and rerun the gsutil ls -l -a. You should see the most recent object generation now has metageneration=2.

For More Information

For more details on how to use versioning and preconditions, see https://cloud.google.com/storage/docs/object-versioning