swift/doc/source/overview_policies.rst
Alistair Coles 71a4aea31a Update docs to discourage policy names being numbers
There are times when it is convenient to specify a policy by name or
by index (see Related-Change), but policy names can unfortunately
collide with indexes. Using a number as a policy name should at least
be discouraged.

Change-Id: I0cdd3b86b527d6656b7fb50c699e3c0cc566e732
Related-Change: Icf1517bd930c74e9552b88250a7b4019e0ab413e
2021-03-26 09:17:34 +00:00

34 KiB

Storage Policies

Storage Policies allow for some level of segmenting the cluster for various purposes through the creation of multiple object rings. The Storage Policies feature is implemented throughout the entire code base so it is an important concept in understanding Swift architecture.

As described in overview_ring, Swift uses modified hashing rings to determine where data should reside in the cluster. There is a separate ring for account databases, container databases, and there is also one object ring per storage policy. Each object ring behaves exactly the same way and is maintained in the same manner, but with policies, different devices can belong to different rings. By supporting multiple object rings, Swift allows the application and/or deployer to essentially segregate the object storage within a single cluster. There are many reasons why this might be desirable:

  • Different levels of durability: If a provider wants to offer, for example, 2x replication and 3x replication but doesn't want to maintain 2 separate clusters, they would setup a 2x and a 3x replication policy and assign the nodes to their respective rings. Furthermore, if a provider wanted to offer a cold storage tier, they could create an erasure coded policy.
  • Performance: Just as SSDs can be used as the exclusive members of an account or database ring, an SSD-only object ring can be created as well and used to implement a low-latency/high performance policy.
  • Collecting nodes into group: Different object rings may have different physical servers so that objects in specific storage policies are always placed in a particular data center or geography.
  • Different Storage implementations: Another example would be to collect together a set of nodes that use a different Diskfile (e.g., Kinetic, GlusterFS) and use a policy to direct traffic just to those nodes.
  • Different read and write affinity settings: proxy-servers can be configured to use different read and write affinity options for each policy. See proxy_server_per_policy_config for more details.

Note

Today, Swift supports two different policy types: Replication and Erasure Code. See overview_erasure_code for details.

Also note that Diskfile refers to backend object storage plug-in architecture. See development_ondisk_backends for details.

Containers and Policies

Policies are implemented at the container level. There are many advantages to this approach, not the least of which is how easy it makes life on applications that want to take advantage of them. It also ensures that Storage Policies remain a core feature of Swift independent of the auth implementation. Policies were not implemented at the account/auth layer because it would require changes to all auth systems in use by Swift deployers. Each container has a new special immutable metadata element called the storage policy index. Note that internally, Swift relies on policy indexes and not policy names. Policy names exist for human readability and translation is managed in the proxy. When a container is created, one new optional header is supported to specify the policy name. If no name is specified, the default policy is used (and if no other policies defined, Policy-0 is considered the default). We will be covering the difference between default and Policy-0 in the next section.

Policies are assigned when a container is created. Once a container has been assigned a policy, it cannot be changed (unless it is deleted/recreated). The implications on data placement/movement for large datasets would make this a task best left for applications to perform. Therefore, if a container has an existing policy of, for example 3x replication, and one wanted to migrate that data to an Erasure Code policy, the application would create another container specifying the other policy parameters and then simply move the data from one container to the other. Policies apply on a per container basis allowing for minimal application awareness; once a container has been created with a specific policy, all objects stored in it will be done so in accordance with that policy. If a container with a specific name is deleted (requires the container be empty) a new container may be created with the same name without any restriction on storage policy enforced by the deleted container which previously shared the same name.

Containers have a many-to-one relationship with policies meaning that any number of containers can share one policy. There is no limit to how many containers can use a specific policy.

The notion of associating a ring with a container introduces an interesting scenario: What would happen if 2 containers of the same name were created with different Storage Policies on either side of a network outage at the same time? Furthermore, what would happen if objects were placed in those containers, a whole bunch of them, and then later the network outage was restored? Well, without special care it would be a big problem as an application could end up using the wrong ring to try and find an object. Luckily there is a solution for this problem, a daemon known as the Container Reconciler works tirelessly to identify and rectify this potential scenario.

Container Reconciler

Because atomicity of container creation cannot be enforced in a distributed eventually consistent system, object writes into the wrong storage policy must be eventually merged into the correct storage policy by an asynchronous daemon. Recovery from object writes during a network partition which resulted in a split brain container created with different storage policies are handled by the swift-container-reconciler daemon.

The container reconciler works off a queue similar to the object-expirer. The queue is populated during container-replication. It is never considered incorrect to enqueue an object to be evaluated by the container-reconciler because if there is nothing wrong with the location of the object the reconciler will simply dequeue it. The container-reconciler queue is an indexed log for the real location of an object for which a discrepancy in the storage policy of the container was discovered.

To determine the correct storage policy of a container, it is necessary to update the status_changed_at field in the container_stat table when a container changes status from deleted to re-created. This transaction log allows the container-replicator to update the correct storage policy both when replicating a container and handling REPLICATE requests.

Because each object write is a separate distributed transaction it is not possible to determine the correctness of the storage policy for each object write with respect to the entire transaction log at a given container database. As such, container databases will always record the object write regardless of the storage policy on a per object row basis. Object byte and count stats are tracked per storage policy in each container and reconciled using normal object row merge semantics.

The object rows are ensured to be fully durable during replication using the normal container replication. After the container replicator pushes its object rows to available primary nodes any misplaced object rows are bulk loaded into containers based off the object timestamp under the .misplaced_objects system account. The rows are initially written to a handoff container on the local node, and at the end of the replication pass the .misplaced_objects containers are replicated to the correct primary nodes.

The container-reconciler processes the .misplaced_objects containers in descending order and reaps its containers as the objects represented by the rows are successfully reconciled. The container-reconciler will always validate the correct storage policy for enqueued objects using direct container HEAD requests which are accelerated via caching.

Because failure of individual storage nodes in aggregate is assumed to be common at scale, the container-reconciler will make forward progress with a simple quorum majority. During a combination of failures and rebalances it is possible that a quorum could provide an incomplete record of the correct storage policy - so an object write may have to be applied more than once. Because storage nodes and container databases will not process writes with an X-Timestamp less than or equal to their existing record when objects writes are re-applied their timestamp is slightly incremented. In order for this increment to be applied transparently to the client a second vector of time has been added to Swift for internal use. See ~swift.common.utils.Timestamp.

As the reconciler applies object writes to the correct storage policy it cleans up writes which no longer apply to the incorrect storage policy and removes the rows from the .misplaced_objects containers. After all rows have been successfully processed it sleeps and will periodically check for newly enqueued rows to be discovered during container replication.

Default versus 'Policy-0'

Storage Policies is a versatile feature intended to support both new and pre-existing clusters with the same level of flexibility. For that reason, we introduce the Policy-0 concept which is not the same as the "default" policy. As you will see when we begin to configure policies, each policy has a single name and an arbitrary number of aliases (human friendly, configurable) as well as an index (or simply policy number). Swift reserves index 0 to map to the object ring that's present in all installations (e.g., /etc/swift/object.ring.gz). You can name this policy anything you like, and if no policies are defined it will report itself as Policy-0, however you cannot change the index as there must always be a policy with index 0.

Another important concept is the default policy which can be any policy in the cluster. The default policy is the policy that is automatically chosen when a container creation request is sent without a storage policy being specified. configure-policy describes how to set the default policy. The difference from Policy-0 is subtle but extremely important. Policy-0 is what is used by Swift when accessing pre-storage-policy containers which won't have a policy - in this case we would not use the default as it might not have the same policy as legacy containers. When no other policies are defined, Swift will always choose Policy-0 as the default.

In other words, default means "create using this policy if nothing else is specified" and Policy-0 means "use the legacy policy if a container doesn't have one" which really means use object.ring.gz for lookups.

Note

With the Storage Policy based code, it's not possible to create a container that doesn't have a policy. If nothing is provided, Swift will still select the default and assign it to the container. For containers created before Storage Policies were introduced, the legacy Policy-0 will be used.

Deprecating Policies

There will be times when a policy is no longer desired; however simply deleting the policy and associated rings would be problematic for existing data. In order to ensure that resources are not orphaned in the cluster (left on disk but no longer accessible) and to provide proper messaging to applications when a policy needs to be retired, the notion of deprecation is used. configure-policy describes how to deprecate a policy.

Swift's behavior with deprecated policies is as follows:

  • The deprecated policy will not appear in /info
  • PUT/GET/DELETE/POST/HEAD are still allowed on the pre-existing containers created with a deprecated policy
  • Clients will get an ''400 Bad Request'' error when trying to create a new container using the deprecated policy
  • Clients still have access to policy statistics via HEAD on pre-existing containers

Note

A policy cannot be both the default and deprecated. If you deprecate the default policy, you must specify a new default.

You can also use the deprecated feature to rollout new policies. If you want to test a new storage policy before making it generally available you could deprecate the policy when you initially roll it the new configuration and rings to all nodes. Being deprecated will render it innate and unable to be used. To test it you will need to create a container with that storage policy; which will require a single proxy instance (or a set of proxy-servers which are only internally accessible) that has been one-off configured with the new policy NOT marked deprecated. Once the container has been created with the new storage policy any client authorized to use that container will be able to add and access data stored in that container in the new storage policy. When satisfied you can roll out a new swift.conf which does not mark the policy as deprecated to all nodes.

Configuring Policies

Note

See policies_saio for a step by step guide on adding a policy to the SAIO setup.

It is important that the deployer have a solid understanding of the semantics for configuring policies. Configuring a policy is a three-step process:

  1. Edit your /etc/swift/swift.conf file to define your new policy.
  2. Create the corresponding policy object ring file.
  3. (Optional) Create policy-specific proxy-server configuration settings.

Defining a policy

Each policy is defined by a section in the /etc/swift/swift.conf file. The section name must be of the form [storage-policy:<N>] where <N> is the policy index. There's no reason other than readability that policy indexes be sequential but the following rules are enforced:

  • If a policy with index 0 is not declared and no other policies are defined, Swift will create a default policy with index 0.
  • The policy index must be a non-negative integer.
  • Policy indexes must be unique.

Warning

The index of a policy should never be changed once a policy has been created and used. Changing a policy index may cause loss of access to data.

Each policy section contains the following options:

  • name = <policy_name> (required)
    • The primary name of the policy.
    • Policy names are case insensitive.
    • Policy names must contain only letters, digits or a dash.
    • Policy names must be unique.
    • Policy names can be changed.
    • The name Policy-0 can only be used for the policy with index 0.
    • To avoid confusion with policy indexes it is strongly recommended that policy names are not numbers (e.g. '1'). However, for backwards compatibility, names that are numbers are supported.
  • aliases = <policy_name>[, <policy_name>, ...] (optional)
    • A comma-separated list of alternative names for the policy.
    • The default value is an empty list (i.e. no aliases).
    • All alias names must follow the rules for the name option.
    • Aliases can be added to and removed from the list.
    • Aliases can be useful to retain support for old primary names if the primary name is changed.
  • default = [true|false] (optional)
    • If true then this policy will be used when the client does not specify a policy.
    • The default value is false.
    • The default policy can be changed at any time, by setting default = true in the desired policy section.
    • If no policy is declared as the default and no other policies are defined, the policy with index 0 is set as the default;
    • Otherwise, exactly one policy must be declared default.
    • Deprecated policies cannot be declared the default.
    • See default-policy for more information.
  • deprecated = [true|false] (optional)
    • If true then new containers cannot be created using this policy.
    • The default value is false.
    • Any policy may be deprecated by adding the deprecated option to the desired policy section. However, a deprecated policy may not also be declared the default. Therefore, since there must always be a default policy, there must also always be at least one policy which is not deprecated.
    • See deprecate-policy for more information.
  • policy_type = [replication|erasure_coding] (optional)
    • The option policy_type is used to distinguish between different policy types.
    • The default value is replication.
    • When defining an EC policy use the value erasure_coding.
  • diskfile_module = <entry point> (optional)
    • The option diskfile_module is used to load an alternate backend object storage plug-in architecture.
    • The default value is egg:swift#replication.fs or egg:swift#erasure_coding.fs depending on the policy type. The scheme and package name are optionals and default to egg and swift.

The EC policy type has additional required options. See using_ec_policy for details.

The following is an example of a properly configured swift.conf file. See policies_saio for full instructions on setting up an all-in-one with this example configuration.:

[swift-hash]
# random unique strings that can never change (DO NOT LOSE)
# Use only printable chars (python -c "import string; print(string.printable)")
swift_hash_path_prefix = changeme
swift_hash_path_suffix = changeme

[storage-policy:0]
name = gold
aliases = yellow, orange
policy_type = replication
default = yes

[storage-policy:1]
name = silver
policy_type = replication
diskfile_module = replication.fs
deprecated = yes

Creating a ring

Once swift.conf is configured for a new policy, a new ring must be created. The ring tools are not policy name aware so it's critical that the correct policy index be used when creating the new policy's ring file. Additional object rings are created using swift-ring-builder in the same manner as the legacy ring except that -N is appended after the word object in the builder file name, where N matches the policy index used in swift.conf. So, to create the ring for policy index 1:

swift-ring-builder object-1.builder create 10 3 1

Continue to use the same naming convention when using swift-ring-builder to add devices, rebalance etc. This naming convention is also used in the pattern for per-policy storage node data directories.

Note

The same drives can indeed be used for multiple policies and the details of how that's managed on disk will be covered in a later section, it's important to understand the implications of such a configuration before setting one up. Make sure it's really what you want to do, in many cases it will be, but in others maybe not.

Proxy server configuration (optional)

The proxy-server configuration options related to read and write affinity may optionally be overridden for individual storage policies. See proxy_server_per_policy_config for more details.

Using Policies

Using policies is very simple - a policy is only specified when a container is initially created. There are no other API changes. Creating a container can be done without any special policy information:

curl -v -X PUT -H 'X-Auth-Token: <your auth token>' \
    http://127.0.0.1:8080/v1/AUTH_test/myCont0

Which will result in a container created that is associated with the policy name 'gold' assuming we're using the swift.conf example from above. It would use 'gold' because it was specified as the default. Now, when we put an object into this container, it will get placed on nodes that are part of the ring we created for policy 'gold'.

If we wanted to explicitly state that we wanted policy 'gold' the command would simply need to include a new header as shown below:

curl -v -X PUT -H 'X-Auth-Token: <your auth token>' \
    -H 'X-Storage-Policy: gold' http://127.0.0.1:8080/v1/AUTH_test/myCont0

And that's it! The application does not need to specify the policy name ever again. There are some illegal operations however:

  • If an invalid (typo, non-existent) policy is specified: 400 Bad Request
  • if you try to change the policy either via PUT or POST: 409 Conflict

If you'd like to see how the storage in the cluster is being used, simply HEAD the account and you'll see not only the cumulative numbers, as before, but per policy statistics as well. In the example below there's 3 objects total with two of them in policy 'gold' and one in policy 'silver':

curl -i -X HEAD -H 'X-Auth-Token: <your auth token>' \
    http://127.0.0.1:8080/v1/AUTH_test

and your results will include (some output removed for readability):

X-Account-Container-Count: 3
X-Account-Object-Count: 3
X-Account-Bytes-Used: 21
X-Storage-Policy-Gold-Object-Count: 2
X-Storage-Policy-Gold-Bytes-Used: 14
X-Storage-Policy-Silver-Object-Count: 1
X-Storage-Policy-Silver-Bytes-Used: 7

Under the Hood

Now that we've explained a little about what Policies are and how to configure/use them, let's explore how Storage Policies fit in at the nuts-n-bolts level.

Parsing and Configuring

The module, storage_policy, is responsible for parsing the swift.conf file, validating the input, and creating a global collection of configured policies via class .StoragePolicyCollection. This collection is made up of policies of class .StoragePolicy. The collection class includes handy functions for getting to a policy either by name or by index , getting info about the policies, etc. There's also one very important function, ~.StoragePolicyCollection.get_object_ring. Object rings are members of the .StoragePolicy class and are actually not instantiated until the ~.StoragePolicy.load_ring method is called. Any caller anywhere in the code base that needs to access an object ring must use the .POLICIES global singleton to access the ~.StoragePolicyCollection.get_object_ring function and provide the policy index which will call ~.StoragePolicy.load_ring if needed; however, when starting request handling services such as the proxy-server rings are proactively loaded to provide moderate protection against a mis-configuration resulting in a run time error. The global is instantiated when Swift starts and provides a mechanism to patch policies for the test code.

Middleware

Middleware can take advantage of policies through the .POLICIES global and by importing .get_container_info to gain access to the policy index associated with the container in question. From the index it can then use the .POLICIES singleton to grab the right ring. For example, list_endpoints is policy aware using the means just described. Another example is recon which will report the md5 sums for all of the rings.

Proxy Server

The proxy-server module's role in Storage Policies is essentially to make sure the correct ring is used as its member element. Before policies, the one object ring would be instantiated when the .Application class was instantiated and could be overridden by test code via init parameter. With policies, however, there is no init parameter and the .Application class instead depends on the .POLICIES global singleton to retrieve the ring which is instantiated the first time it's needed. So, instead of an object ring member of the .Application class, there is an accessor function, ~.Application.get_object_ring, that gets the ring from .POLICIES.

In general, when any module running on the proxy requires an object ring, it does so via first getting the policy index from the cached container info. The exception is during container creation where it uses the policy name from the request header to look up policy index from the .POLICIES global. Once the proxy has determined the policy index, it can use the ~.Application.get_object_ring method described earlier to gain access to the correct ring. It then has the responsibility of passing the index information, not the policy name, on to the back-end servers via the header X -Backend-Storage-Policy-Index. Going the other way, the proxy also strips the index out of headers that go back to clients, and makes sure they only see the friendly policy names.

On Disk Storage

Policies each have their own directories on the back-end servers and are identified by their storage policy indexes. Organizing the back-end directory structures by policy index helps keep track of things and also allows for sharing of disks between policies which may or may not make sense depending on the needs of the provider. More on this later, but for now be aware of the following directory naming convention:

  • /objects maps to objects associated with Policy-0
  • /objects-N maps to storage policy index #N
  • /async_pending maps to async pending update for Policy-0
  • /async_pending-N maps to async pending update for storage policy index #N
  • /tmp maps to the DiskFile temporary directory for Policy-0
  • /tmp-N maps to the DiskFile temporary directory for policy index #N
  • /quarantined/objects maps to the quarantine directory for Policy-0
  • /quarantined/objects-N maps to the quarantine directory for policy index #N

Note that these directory names are actually owned by the specific Diskfile implementation, the names shown above are used by the default Diskfile.

Object Server

The object-server is not involved with selecting the storage policy placement directly. However, because of how back-end directory structures are setup for policies, as described earlier, the object server modules do play a role. When the object server gets a .Diskfile, it passes in the policy index and leaves the actual directory naming/structure mechanisms to .Diskfile. By passing in the index, the instance of .Diskfile being used will assure that data is properly located in the tree based on its policy.

For the same reason, the object-updater also is policy aware. As previously described, different policies use different async pending directories so the updater needs to know how to scan them appropriately.

The object-replicator is policy aware in that, depending on the policy, it may have to do drastically different things, or maybe not. For example, the difference in handling a replication job for 2x versus 3x is trivial; however, the difference in handling replication between 3x and erasure code is most definitely not. In fact, the term 'replication' really isn't appropriate for some policies like erasure code; however, the majority of the framework for collecting and processing jobs is common. Thus, those functions in the replicator are leveraged for all policies and then there is policy specific code required for each policy, added when the policy is defined if needed.

The ssync functionality is policy aware for the same reason. Some of the other modules may not obviously be affected, but the back-end directory structure owned by .Diskfile requires the policy index parameter. Therefore ssync being policy aware really means passing the policy index along. See ~swift.obj.ssync_sender and ~swift.obj.ssync_receiver for more information on ssync.

For .Diskfile itself, being policy aware is all about managing the back-end structure using the provided policy index. In other words, callers who get a .Diskfile instance provide a policy index and .Diskfile's job is to keep data separated via this index (however it chooses) such that policies can share the same media/nodes if desired. The included implementation of .Diskfile lays out the directory structure described earlier but that's owned within .Diskfile; external modules have no visibility into that detail. A common function is provided to map various directory names and/or strings based on their policy index. For example .Diskfile defines ~swift.obj.diskfile.get_data_dir which builds off of a generic .get_policy_string to consistently build policy aware strings for various usage.

Container Server

The container-server plays a very important role in Storage Policies, it is responsible for handling the assignment of a policy to a container and the prevention of bad things like changing policies or picking the wrong policy to use when nothing is specified (recall earlier discussion on Policy-0 versus default).

The container-updater is policy aware, however its job is very simple, to pass the policy index along to the account-server via a request header.

The container-backend is responsible for both altering existing DB schema as well as assuring new DBs are created with a schema that supports storage policies. The "on-demand" migration of container schemas allows Swift to upgrade without downtime (sqlite's alter statements are fast regardless of row count). To support rolling upgrades (and downgrades) the incompatible schema changes to the container_stat table are made to a container_info table, and the container_stat table is replaced with a view that includes an INSTEAD OF UPDATE trigger which makes it behave like the old table.

The policy index is stored here for use in reporting information about the container as well as managing split-brain scenario induced discrepancies between containers and their storage policies. Furthermore, during split-brain, containers must be prepared to track object updates from multiple policies so the object table also includes a storage_policy_index column. Per-policy object counts and bytes are updated in the policy_stat table using INSERT and DELETE triggers similar to the pre-policy triggers that updated container_stat directly.

The container-replicator daemon will pro-actively migrate legacy schemas as part of its normal consistency checking process when it updates the reconciler_sync_point entry in the container_info table. This ensures that read heavy containers which do not encounter any writes will still get migrated to be fully compatible with the post-storage-policy queries without having to fall back and retry queries with the legacy schema to service container read requests.

The container-sync-daemon functionality only needs to be policy aware in that it accesses the object rings. Therefore, it needs to pull the policy index out of the container information and use it to select the appropriate object ring from the .POLICIES global.

Account Server

The account-server's role in Storage Policies is really limited to reporting. When a HEAD request is made on an account (see example provided earlier), the account server is provided with the storage policy index and builds the object_count and byte_count information for the client on a per policy basis.

The account servers are able to report per-storage-policy object and byte counts because of some policy specific DB schema changes. A policy specific table, policy_stat, maintains information on a per policy basis (one row per policy) in the same manner in which the account_stat table does. The account_stat table still serves the same purpose and is not replaced by policy_stat, it holds the total account stats whereas policy_stat just has the break downs. The backend is also responsible for migrating pre-storage-policy accounts by altering the DB schema and populating the policy_stat table for Policy-0 with current account_stat data at that point in time.

The per-storage-policy object and byte counts are not updated with each object PUT and DELETE request, instead container updates to the account server are performed asynchronously by the swift-container-updater.

Upgrading and Confirming Functionality

Upgrading to a version of Swift that has Storage Policy support is not difficult, in fact, the cluster administrator isn't required to make any special configuration changes to get going. Swift will automatically begin using the existing object ring as both the default ring and the Policy-0 ring. Adding the declaration of policy 0 is totally optional and in its absence, the name given to the implicit policy 0 will be 'Policy-0'. Let's say for testing purposes that you wanted to take an existing cluster that already has lots of data on it and upgrade to Swift with Storage Policies. From there you want to go ahead and create a policy and test a few things out. All you need to do is:

  1. Upgrade all of your Swift nodes to a policy-aware version of Swift
  2. Define your policies in /etc/swift/swift.conf
  3. Create the corresponding object rings
  4. Create containers and objects and confirm their placement is as expected

For a specific example that takes you through these steps, please see policies_saio

Note

If you downgrade from a Storage Policy enabled version of Swift to an older version that doesn't support policies, you will not be able to access any data stored in policies other than the policy with index 0 but those objects WILL appear in container listings (possibly as duplicates if there was a network partition and un-reconciled objects). It is EXTREMELY important that you perform any necessary integration testing on the upgraded deployment before enabling an additional storage policy to ensure a consistent API experience for your clients. DO NOT downgrade to a version of Swift that does not support storage policies once you expose multiple storage policies.