rgw: Document the layout of pools and objects

This has to be kept in sync as the layout changes. Like any other
document, really.

Signed-off-by: Pete Zaitcev <zaitcev@redhat.com>

diff --git a/doc/radosgw/layout.rst b/doc/radosgw/layout.rst
new file mode 100644 (file)
index 0000000..0ebb540
--- /dev/null
+++ b/doc/radosgw/layout.rst
@@ -0,0 +1,178 @@
+===========================
+ Rados Gateway Data Layout
+===========================
+
+Although the source code is the ultimate guide, this document helps
+new developers to get up to speed with the implementation details.
+
+Introduction
+------------
+
+Swift offers something called a container, that we use interchangeably with
+the term bucket. One may say that RGW's buckets implement Swift containers.
+
+This document does not consider how RGW operates on these structures,
+e.g. the use of encode() and decode() methods for serialization and so on.
+
+Conceptual View
+---------------
+
+Although RADOS only knows about pools and objects with their xattrs and
+omap[1], conceptually RGW organizes its data into three different kinds:
+metadata, bucket index, and data.
+
+* Metadata
+
+We have 3 'sections' of metadata: 'user', 'bucket', and 'bucket.instance'.
+You can use the following commands to introspect metadata entries:
+
+$ radosgw-admin metadata list
+
+$ radosgw-admin metadata list bucket
+$ radosgw-admin metadata list bucket.instance
+$ radosgw-admin metadata list user
+
+$ radosgw-admin metadata get bucket:<bucket>
+$ radosgw-admin metadata get bucket.instance:<bucket>:<bucket_id>
+$ radosgw-admin metadata get user:<user>   # or set
+
+user: Holds user information
+bucket: Holds a mapping between bucket name and bucket instance id
+bucket.instance: Holds bucket instance information[2]
+
+Every metadata entry is kept on a single rados object.
+See below for implementation defails.
+
+Note that the metadata is not indexed. When listing a metadata section we do a
+rados pgls operation on the containing pool.
+
+* Bucket Index
+
+It's a different kind of metadata, and kept separately. The bucket index holds
+a key-value map in rados objects. By default it is a single rados object per
+bucket, but it is possible since Hammer to shard that map over multiple rados
+objects. The map itself is kept in omap, associated with each rados object.
+The key of each omap is the name of the objects, and the value holds some basic
+metadata of that object -- metadata that shows up when listing the bucket.
+Also, each omap holds a header, and we keep some bucket accounting metadata
+in that header (number of objects, total size, etc.).
+
+Note that we also hold other information in the bucket index, and it's kept in
+other key namespaces. We can hold the bucket index log there, and for versioned
+objects there is more information that we keep on other keys.
+
+* Data
+
+Objects data is kept in one or more rados objects for each rgw object.
+
+Object Lookup Path
+------------------
+
+When accessing objects, ReST APIs come to RGW with three parameters:
+account information (access key in S3 or account name in Swift),
+bucket or container name, and object name (or key). At present, RGW only
+uses account information to find out the user ID and for access control.
+Only the bucket name and object key are used to address the object in a pool.
+
+The user ID in RGW is a string, typically the actual user name from the user
+credentials and not a hashed or mapped identifier.
+
+When accessing a user's data, the user record is loaded from an object
+"<user_id>" in pool ".users.uid".
+
+Bucket names are represented directly in the pool ".rgw". Bucket record is
+loaded in order to obtain so-called marker, which serves as a bucket ID.
+
+The object is located in pool ".rgw.buckets". Object name is "<marker>_<key>",
+for example "default.7593.4_image.png", where the marker is "default.7593.4"
+and the key is "image.png". Since these concatenated names are not parsed,
+only passed down to RADOS, the choice of the separator is not important and
+causes no ambiguity. For the same reason, slashes are permitted in object
+names (keys).
+
+It is also possible to create multiple data pools and make it so that
+different users buckets will be created in different rados pools by default,
+thus providing the necessary scaling. The layout and naming of these pools
+is controlled by a 'policy' setting.[3]
+
+An RGW object may consist of several RADOS objects, the first of which
+is the head that contains the metadata, such as manifest, ACLs, content type,
+ETag, and user-defined metadata. The metadata is stored in xattrs.
+The head may also contain up to 512 kilobytes of object data, for efficiency
+and atomicity. The manifest describes how each object is laid out in RADOS
+objects.
+
+Bucket and Object Listing
+-------------------------
+
+Buckets that belong to a given user are listed in an omap of an object named
+"<user_id>.buckets" (for example, "foo.buckets") in pool ".users.uid".
+These objects are accessed when listing buckets, when updating bucket
+contents, and updating and retrieving bucket statistics (e.g. for quota).
+
+See the user-visible, encoded class 'cls_user_bucket_entry' and its
+nested class 'cls_user_bucket' for the values of these omap entires.
+
+These listings are kept consistent with buckets in pool ".rgw".
+
+Objects that belong to a given bucket are listed in a bucket index,
+as discussed in sub-section 'Bucket Index' above. The default naming
+for index objects is ".dir.<marker>" in pool ".rgw.buckets.index".
+
+Footnotes
+---------
+
+[1] Omap is a key-value store, associated with an object, in a way similar
+to how Extended Attributes associate with a POSIX file. An object's omap
+is not physically located in the object's storage, but its precise
+implementation is invisible and immaterial to RADOS Gateway.
+In Hammer, one LevelDB is used to store omap in each OSD.
+
+[2] Before the Dumpling release, the 'bucket.instance' metadata did not
+exist and the 'bucket' metadata contained its information. It is possible
+to encounter such buckets in old installations.
+
+[3] In Infernalis, a pending commit exists that removes the need of prefixing
+all the rgw system pools with a period, and also renames all of these pools.
+See Github pull request #4944 "rgw noperiod".
+
+Appendix: Compendum
+-------------------
+
+Known pools:
+
+.rgw.root
+  Unspecified region, zone, and global information records, one per object.
+
+.rgw.control
+  notify.<N>
+
+.rgw
+  <bucket>
+  .bucket.meta.<bucket>:<marker>
+
+.rgw.gc
+  gc.<N>
+
+.users.uid
+  Contains _both_ per-user information (RGWUserInfo) in "<user>" objects
+  and per-user bucket lists in omaps of "<user>.buckets" objects.
+
+.users.email
+  Unimportant
+
+.users
+  47UA98JSTJZ9YAN3OS3O
+  It's unclear why user ID is not used to name objects in this pool.
+
+.users.swift
+  test:tester
+
+.rgw.buckets.index
+  Objects are named ".dir.<marker>", each contains a bucket index.
+
+.rgw.buckets
+  default.7593.4__shadow_.488urDFerTYXavx4yAd-Op8mxehnvTI_1
+  <marker>_<key>
+
+An example of a marker would be "default.16004.1" or "default.7593.4".