This change refactors the storage backend to use the new path layout. To
facilitate this, manifest storage has been separated into a revision store and
tag store, supported by a more general blob store. The blob store is a hybrid
object, effectively providing both small object access, keyed by content
address, as well as methods that can be used to manage and traverse links to
underlying blobs. This covers common operations used in the revision store and
tag store, such as linking and traversal. The blob store can also be updated to
better support layer reading but this refactoring has been left for another
day.
The revision store and tag store support the manifest store's compound view of
data. These underlying stores provide facilities for richer access models, such
as content-addressable access and a richer tagging model. The highlight of this
change is the ability to sign a manifest from different hosts and have the
registry merge and serve those signatures as part of the manifest package.
Various other items, such as the delegate layer handler, were updated to more
directly use the blob store or other mechanism to fit with the changes.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
To smooth initial implementation, uploads were spooled to local file storage,
validated, then pushed to remote storage. That approach was flawed in that it
present easy clustering of registry services that share a remote storage
backend. The original plan was to implement resumable hashes then implement
remote upload storage. After some thought, it was found to be better to get
remote spooling working, then optimize with resumable hashes.
Moving to this approach has tradeoffs: after storing the complete upload
remotely, the node must fetch the content and validate it before moving it to
the final location. This can double bandwidth usage to the remote backend.
Modifying the verification and upload code to store intermediate hashes should
be trivial once the layer digest format has settled.
The largest changes for users of the storage package (mostly the registry app)
are the LayerService interface and the LayerUpload interface. The LayerService
now takes qualified repository names to start and resume uploads. In corallry,
the concept of LayerUploadState has been complete removed, exposing all aspects
of that state as part of the LayerUpload object. The LayerUpload object has
been modified to work as an io.WriteSeeker and includes a StartedAt time, to
allow for upload timeout policies. Finish now only requires a digest, eliding
the requirement for a size parameter.
Resource cleanup has taken a turn for the better. Resources are cleaned up
after successful uploads and during a cancel call. Admittedly, this is probably
not completely where we want to be. It's recommend that we bolster this with a
periodic driver utility script that scans for partial uploads and deletes the
underlying data. As a small benefit, we can leave these around to better
understand how and why these uploads are failing, at the cost of some extra
disk space.
Many other changes follow from the changes above. The webapp needs to be
updated to meet the new interface requirements.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
To support clustered registry, upload UUIDs must be recognizable by
registries that did not issue the UUID. By creating an HMAC verifiable
upload state token, registries can validate upload requests that other
instances authorized. The tokenProvider interface could also use a redis
store or other system for token handling in the future.
Because manifests and their signatures are a discrete component of the
registry, we are moving the definitions into a separate package. This causes us
to lose some test coverage, but we can fill this in shortly. No changes have
been made to the external interfaces, but they are likely to come.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
This change implements the first pass at image manifest storage on top of the
storagedriver. Very similar to LayerService, its much simpler due to less
complexity of pushing and pulling images.
Various components are still missing, such as detailed error reporting on
missing layers during verification, but the base functionality is present.
This change separates out the remote file reader functionality from layer
reprsentation data. More importantly, issues with seeking have been fixed and
thoroughly tested.
This change contains the initial implementation of the LayerService to power
layer push and pulls on the storagedriver. The interfaces presented in this
package will be used by the http application to drive most features around
efficient pulls and resumable pushes.
The file storage/layer.go defines the interface interactions. LayerService is
the root type and supports methods to access Layer and LayerUpload objects.
Pull operations are supported with LayerService.Fetch and push operations are
supported with LayerService.Upload and LayerService.Resume. Reads and writes of
layers are split between Layer and LayerUpload, respectively.
LayerService is implemented internally with the layerStore object, which takes
a storagedriver.StorageDriver and a pathMapper instance.
LayerUploadState is currently exported and will likely continue to be as the
interaction between it and layerUploadStore are better understood. Likely, the
layerUploadStore lifecycle and implementation will be deferred to the
application.
Image pushes pulls will be implemented in a similar manner without the
discrete, persistent upload.
Much of this change is in place to get something running and working. Caveats
of this change include the following:
1. Layer upload state storage is implemented on the local filesystem, separate
from the storage driver. This must be replaced with using the proper backend
and other state storage. This can be removed when we implement resumable
hashing and tarsum calculations to avoid backend roundtrips.
2. Error handling is rather bespoke at this time. The http API implementation
should really dictate the error return structure for the future, so we
intend to refactor this heavily to support these errors. We'd also like to
collect production data to understand how failures happen in the system as
a while before moving to a particular edict around error handling.
3. The layerUploadStore, which manages layer upload storage and state is not
currently exported. This will likely end up being split, with the file
management portion being pointed at the storagedriver and the state storage
elsewhere.
4. Access Control provisions are nearly completely missing from this change.
There are details around how layerindex lookup works that are related with
access controls. As the auth portions of the new API take shape, these
provisions will become more clear.
Please see TODOs for details and individual recommendations.
