This driver implements the storagedriver.StorageDriver interface and
uses Ceph Object Storage as storage backend.
Since RADOS is an object storage and no hierarchy notion, the
following convention is used to keep the filesystem notions stored in
this backend:
* All the objects data are stored with opaque UUID names prefixed
(e.g. "blob:d3d232ff-ab3a-4046-9ab7-930228d4c164).
* All the hierarchy information are stored in rados omaps, where the
omap object identifier is the virtual directory name, the keys in
a specific are the relative filenames and the values the blob
object identifier (or empty value for a sub directory).
e.g. For the following hierarchy:
/directory1
/directory1/object1
/directory1/object2
/directory1/directory2/object3
The omap "/directory1" will contains the following key / values:
- "object1" "blob:d3d232ff-ab3a-4046-9ab7-930228d4c164"
- "object2" "blob:db2e359d-4af0-4bfb-ba1d-d2fd029866a0"
- "directory2" ""
The omap "/directory1/directory2" will contains:
- "object3" "blob:9ae2371c-81fc-4945-80ac-8bf7f566a5d9"
* The MOVE is implemented by changing the reference to a specific
blob in its parent virtual directory omap.
This driver stripes rados objects to a fixed size (e.g. 4M). The idea
is to keep small objects (as done by RBD on the top of RADOS) that
will be easily synchronized accross OSDs. The information of the
original object (i.e total size of the chunks) is stored as a Xattr
in the first chunk object.
Signed-off-by: Vincent Giersch <vincent.giersch@ovh.net>
This PR refactors the blob service API to be oriented around blob descriptors.
Identified by digests, blobs become an abstract entity that can be read and
written using a descriptor as a handle. This allows blobs to take many forms,
such as a ReadSeekCloser or a simple byte buffer, allowing blob oriented
operations to better integrate with blob agnostic APIs (such as the `io`
package). The error definitions are now better organized to reflect conditions
that can only be seen when interacting with the blob API.
The main benefit of this is to separate the much smaller metadata from large
file storage. Many benefits also follow from this. Reading and writing has
been separated into discrete services. Backend implementation is also
simplified, by reducing the amount of metadata that needs to be picked up to
simply serve a read. This also improves cacheability.
"Opening" a blob simply consists of an access check (Stat) and a path
calculation. Caching is greatly simplified and we've made the mapping of
provisional to canonical hashes a first-class concept. BlobDescriptorService
and BlobProvider can be combined in different ways to achieve varying effects.
Recommend Review Approach
-------------------------
This is a very large patch. While apologies are in order, we are getting a
considerable amount of refactoring. Most changes follow from the changes to
the root package (distribution), so start there. From there, the main changes
are in storage. Looking at (*repository).Blobs will help to understand the how
the linkedBlobStore is wired. One can explore the internals within and also
branch out into understanding the changes to the caching layer. Following the
descriptions below will also help to guide you.
To reduce the chances for regressions, it was critical that major changes to
unit tests were avoided. Where possible, they are left untouched and where
not, the spirit is hopefully captured. Pay particular attention to where
behavior may have changed.
Storage
-------
The primary changes to the `storage` package, other than the interface
updates, were to merge the layerstore and blobstore. Blob access is now
layered even further. The first layer, blobStore, exposes a global
`BlobStatter` and `BlobProvider`. Operations here provide a fast path for most
read operations that don't take access control into account. The
`linkedBlobStore` layers on top of the `blobStore`, providing repository-
scoped blob link management in the backend. The `linkedBlobStore` implements
the full `BlobStore` suite, providing access-controlled, repository-local blob
writers. The abstraction between the two is slightly broken in that
`linkedBlobStore` is the only channel under which one can write into the global
blob store. The `linkedBlobStore` also provides flexibility in that it can act
over different link sets depending on configuration. This allows us to use the
same code for signature links, manifest links and blob links. Eventually, we
will fully consolidate this storage.
The improved cache flow comes from the `linkedBlobStatter` component
of `linkedBlobStore`. Using a `cachedBlobStatter`, these combine together to
provide a simple cache hierarchy that should streamline access checks on read
and write operations, or at least provide a single path to optimize. The
metrics have been changed in a slightly incompatible way since the former
operations, Fetch and Exists, are no longer relevant.
The fileWriter and fileReader have been slightly modified to support the rest
of the changes. The most interesting is the removal of the `Stat` call from
`newFileReader`. This was the source of unnecessary round trips that were only
present to look up the size of the resulting reader. Now, one must simply pass
in the size, requiring the caller to decide whether or not the `Stat` call is
appropriate. In several cases, it turned out the caller already had the size
already. The `WriterAt` implementation has been removed from `fileWriter`,
since it is no longer required for `BlobWriter`, reducing the number of paths
which writes may take.
Cache
-----
Unfortunately, the `cache` package required a near full rewrite. It was pretty
mechanical in that the cache is oriented around the `BlobDescriptorService`
slightly modified to include the ability to set the values for individual
digests. While the implementation is oriented towards caching, it can act as a
primary store. Provisions are in place to have repository local metadata, in
addition to global metadata. Fallback is implemented as a part of the storage
package to maintain this flexibility.
One unfortunate side-effect is that caching is now repository-scoped, rather
than global. This should have little effect on performance but may increase
memory usage.
Handlers
--------
The `handlers` package has been updated to leverage the new API. For the most
part, the changes are superficial or mechanical based on the API changes. This
did expose a bug in the handling of provisional vs canonical digests that was
fixed in the unit tests.
Configuration
-------------
One user-facing change has been made to the configuration and is updated in
the associated documentation. The `layerinfo` cache parameter has been
deprecated by the `blobdescriptor` cache parameter. Both are equivalent and
configuration files should be backward compatible.
Notifications
-------------
Changes the `notification` package are simply to support the interface
changes.
Context
-------
A small change has been made to the tracing log-level. Traces have been moved
from "info" to "debug" level to reduce output when not needed.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
- Set an Etag header
- Check If-None-Match and respond appropriately
- Set a Cache-Control header with a default of 1 week
Signed-off-by: Richard Scothern <richard.scothern@gmail.com>
This deals with a memory leak, caused by goroutines, experienced when using the
s3 driver. Unfortunately, this section of the code leaks goroutines like a
sieve. There is probably some refactoring that could be done to avoid this but
instead, we have a done channel that will cause waiting goroutines to exit.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
- Change driver interface to take a context as its first argument
- Make newFileReader take a context as its first argument
- Make newFileWriter take a context as its first argument
- Make blobstore exists and delete take a context as a first argument
- Pass the layerreader's context to the storage layer
- Pass the app's context to purgeuploads
- Store the app's context into the blobstore (was previously null)
- Pass the trace'd context to the storage drivers
Signed-off-by: Richard Scothern <richard.scothern@gmail.com>
The code using values from the yaml package wasn't careful enought with the
possible incoming types. Turns out, it is just an int but we've made this
section somewhat bulletproof in case that package changes the behavior.
This code likely never worked. The configuration system should be decoupled
from the object instantiation.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
Rather than accept the resulting of a layer validation, we retry up to three
times, backing off 100ms after each try. The thought is that we allow s3 files
to make their way into the correct location increasing the liklihood the
verification can proceed, if possible.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
When the registry starts a background timer will periodically
scan the upload directories on the file system every 24 hours
and delete any files older than 1 week. An initial jitter
intends to avoid contention on the filesystem where multiple
registries with the same storage driver are started
simultaneously.
Registry is intended to be used as a repository service than an abstract collection of repositories. Namespace better describes a collection of repositories retrievable by name.
The registry service serves any repository in the global scope.
Signed-off-by: Derek McGowan <derek@mcgstyle.net> (github: dmcgowan)
The original implementation wrote to different locations in a shared slice.
While this is theoretically okay, we end up thrashing the cpu cache since
multiple slice members may be on the same cache line. So, even though each
thread has its own memory location, there may be contention over the cache
line. This changes the code to aggregate to a slice in a single goroutine.
In reality, this change likely won't have any performance impact. The theory
proposed above hasn't really even been tested. Either way, we can consider it
and possibly go forward.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
Rather than enforce lowercase paths for all drivers, support for
case-sensitivity has been deferred to the driver. There are a few caveats to
this approach:
1. There are possible security implications for tags that only differ in their
case. For instance, a tag "A" may be equivalent to tag "a" on certain file
system backends.
2. All system paths should not use case-sensitive identifiers where possible.
This might be problematic in a blob store that uses case-sensitive ids. For
now, since digest hex ids are all case-insensitive, this will not be an issue.
The recommend workaround is to not run the registry on a case-insensitive
filesystem driver in security sensitive applications.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
To avoid compounded round trips leading to slow retrieval of manifests with a
large number of signatures, the fetch of signatures has been parallelized. This
simply spawns a goroutine for each path, coordinated with a sync.WaitGroup.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
For consistency with other systems, the redis and caching monitoring data has
been moved under the "registry" section in expvar. This ensures the entire
registry state is kept to a single section.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
This changeset integrates the layer info cache with the registry webapp and
storage backend. The main benefit is to cache immutable layer meta data,
reducing backend roundtrips. The cache can be configured to use either redis or
an inmemory cache.
This provides massive performance benefits for HEAD http checks on layer blobs
and manifest verification.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
This changeset defines the interface for layer info caches. Layer info caches
speed up access to layer meta data accessed in storage driver backends. The
two main operations are tests for repository membership and resolving path and
size information for backend blobs.
Two implementations are available. The main implementation leverages redis to
store layer info. An alternative implementation simply caches layer info in
maps, which should speed up resolution for less sophisticated implementations.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
By using a resumable digester and storing the state of upload digests between
subsequent upload chunks, finalizing an upload no longer requires reading back
all of the uploaded data to verify the client's expected digest.
Docker-DCO-1.1-Signed-off-by: Josh Hawn <josh.hawn@docker.com> (github: jlhawn)
This chnage prevents a crash when moving from a non-existent directory that has
a file as a parent. To prevent this, we simply check that the node is a
directory and throws an error if it is not.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
This enables Azure storage driver to be used with non-default
cloud endpoints like Azure China or Azure Government that does
not use `.blob.core.windows.net` FQDN suffix.
Signed-off-by: Ahmet Alp Balkan <ahmetalpbalkan@gmail.com>
