On a ranged download the metadata decoder stored the response's
Content-Length (the length of the range, not the blob) in the object's
size and only corrected it from the Content-Range total afterwards.
Object.Size() is read concurrently by the VFS cache and chunked reader
while a download is in progress, so with --vfs-read-chunk-size a reader
could observe the chunk length (e.g. 67108864 for 64M chunks) as the
object size. The VFS cache then logged
vfs cache: cached file (N) is unexpectedly larger than the remote
object (67108864). The cached file is likely corrupted after an
unclean shutdown; recovering ...
and truncated the read request against the bogus size, breaking
sequential reads of large blobs with --vfs-cache-mode full.
This applies the Content-Range correction before the size is stored so
the range length is never published as the object size.
When using Microsoft Entra ID credentials, Azure Blob server-side copy
uses a user delegation SAS URL for the private copy source.
The SAS start time was set to the current local time. Azure Storage
validates the copy source from the service side, and small clock
differences can make that SAS appear not yet valid. The service then
returns 403 CannotVerifyCopySource with AuthenticationFailed.
Start the copy-source SAS 15 minutes in the past, matching Microsoft SAS
guidance for clock skew.
Add a new "list_parallelism" advanced option (requires use_arrow_list) which
splits a directory's blob-name keyspace into ranges and lists them
concurrently using the Arrow startFrom/endBefore range parameters. Both
recursive (ListR) and single-directory listings benefit.
The keyspace is sharded with a single-case (digit+lowercase) character
ladder. Single case is deliberate: the service validates startFrom /
endBefore case-insensitively (while listing case-sensitively), so a
mixed-case boundary pair is rejected with a 400; its endBefore filtering is
byte-ordered, so the listing is exact.
The service only honours endBefore on the Arrow listing path of accounts with
Blob Listing with Apache Arrow enabled: elsewhere it either rejects the
request with 400 OperationNotSupportedWithFeatureMissing or, if it fell back
to XML, would ignore the shards' upper bounds. Both cases are detected and
the parallel listing falls back to a sequential listing.
Listing 100,000 objects from a well-connected server:
config time vs XML
xml 23.9s 1.0x
use_arrow_list 6.9s 3.5x
list_parallelism=2 4.4s 5.5x
list_parallelism=4 2.8s 8.4x
list_parallelism=6 2.2s 10.9x
list_parallelism=10 1.5s 15.4x
list_parallelism=20 1.4s 17.6x
list_parallelism=30 1.1s 21.7x
So Apache Arrow listing alone is ~3.5x faster than XML, and parallel listing
takes that to ~20x. The ladder caps at 37 shards, so the benefit plateaus
beyond list_parallelism=30.
The option is hidden from the configurator, command line help and docs
until the feature is officially released.
Add a new "use_arrow_list" advanced option which fetches directory listings
using the ListBlobs Apache Arrow response format instead of XML. This can be
substantially faster for very large containers.
This is EXPERIMENTAL and requires the "Blob Listing with Apache Arrow" preview
feature to be enabled on the storage account. It is not supported on accounts
with a hierarchical namespace (ADLS Gen2). If the feature is not enabled the
service returns XML and the listing transparently falls back to it. Connection
string auth is not supported - it also falls back to the normal XML listing.
The option is hidden from the configurator, command line help and docs until the
feature is officially released.
This adds a temporary subpackage implementing the experimental "Blob
Listing with Apache Arrow" feature on top of the released Azure azblob
SDK.
The Azure SDK for Go supports Arrow listing on its unreleased
feature/storage/bifrost branch (commit c6fa341ca22b) but that can't be
used in a release. This package exposes the same options and pager
interface as the experimental SDK, built on the released SDK's public
API plus verbatim copies of its Arrow decoder and of the two internal
auth policies (shared key signing and storage bearer challenge) which
the released SDK does not export. Copied files carry MIT attribution
headers pinned to their upstream source.
When Arrow listing ships in a released azblob the whole package will be
deleted and callers pointed back at the SDK - the interface is
deliberately source compatible to make that a mechanical change.
Also exports auth.Transporter so the new package's pipeline can use
rclone's fshttp transport, and adds github.com/apache/arrow-go/v18 as a
direct dependency for decoding the Arrow IPC stream.
On a range request Azure returns the whole-blob MD5 in the
x-ms-blob-content-md5 header (BlobContentMD5) and leaves Content-MD5
empty. The download metadata decoder only read Content-MD5, so every
ranged read overwrote the object's MD5 with an empty string.
With --vfs-cache-mode full this changed the object's fingerprint between
opens, so the VFS cache judged every reopened file as stale and
re-downloaded it, defeating the cache and inflating egress. Prefer
BlobContentMD5 and never overwrite a known hash with an empty one.
Before this change, if an object compressed with "Content-Encoding:
gzip" was downloaded, a length and hash mismatch would occur since the
go runtime automatically decompressed the object on download.
If --azureblob-decompress is set, this change erases the length and hash on
compressed objects so they can be downloaded successfully, at the cost
of not being able to check the length or the hash of the downloaded
object.
If --azureblob-decompress is not set the compressed files will be downloaded
as-is providing compressed objects with intact size and hash
information.
Fixes#9337
Set the User-Agent to include the APN prefix for Azure backends
(azureblob, azurefiles, onelake) to identify rclone as a Microsoft
Partner. The User-Agent is now:
APN/1.0 rclone/1.0 rclone/<version>
This change adds first-class metadata support to the Azure Blob backend,
including headers, user metadata, tags, and modtime overrides, and wires
it through uploads and server-side copies.
There is a behavior change in that rclone will now set the "mtime"
custom metadata when doing server side copies to azure and the
`--metadata` argument is given.
- Map standard headers: cache-control, content-disposition, content-encoding,
content-language, content-type to corresponding x-ms-blob-* HTTP headers.
- Map user metadata: any non-reserved keys (excluding x-ms-*) are sent as
blob user metadata. Keys are normalized to lowercase for consistency.
- Support tags: parse `x-ms-tags` as a comma-separated list of key=value
pairs and apply them on uploads and copies.
- Support mtime override: accept `mtime` in metadata (RFC3339/RFC3339Nano)
to override the stored modtime persisted in user metadata.
Before this change multipart uploads using OpenChunkWriter would
account for twice the space used.
This fixes the problem by adjusting the accounting delay.
Before this change the azureblob backend could deadlock when using
--max-connections. This is because when it receives InvalidBlockOrBlob
error it attempts to clear the condition before retrying. This in turn
involved recursively calling the pacer. At this point the pacer can
easily have no connections left which causes a deadlock as all the
other pacer connections are waiting for the InvalidBlockOrBlob to be
resolved.
This fixes the problem by using a temporary pacer when resolving the
InvalidBlockOrBlob errors.
Before this change, if not using shared key or SAS URL authentication
for the source, rclone gave this error
ManagedIdentityCredential.GetToken() requires exactly one scope
when doing server side copies.
This was introduced in:
3a5ddfcd3c azureblob: implement multipart server side copy
This fixes the problem by creating a temporary SAS URL using user
delegation to read the source blob when copying.
Fixes#8662
Before this fix multipart server side copies would fail.
This problem was due to an incorrect calculation of the number of
parts to transfer - it calculated 1 part to transfer rather than 0.
This commit modernizes Go usage. This was done with:
go run golang.org/x/tools/gopls/internal/analysis/modernize/cmd/modernize@latest -fix -test ./...
Then files needed to be `go fmt`ed and a few comments needed to be
restored.
The modernizations include replacing
- if/else conditional assignment by a call to the built-in min or max functions added in go1.21
- sort.Slice(x, func(i, j int) bool) { return s[i] < s[j] } by a call to slices.Sort(s), added in go1.21
- interface{} by the 'any' type added in go1.18
- append([]T(nil), s...) by slices.Clone(s) or slices.Concat(s), added in go1.21
- loop around an m[k]=v map update by a call to one of the Collect, Copy, Clone, or Insert functions from the maps package, added in go1.21
- []byte(fmt.Sprintf...) by fmt.Appendf(nil, ...), added in go1.19
- append(s[:i], s[i+1]...) by slices.Delete(s, i, i+1), added in go1.21
- a 3-clause for i := 0; i < n; i++ {} loop by for i := range n {}, added in go1.22
Sometimes the Azure blob servers reply with 503 ServerBusy errors and
these should be retried.
Before this change, when testing to see if a single part server side
copy was done, ServerBusy errors were returned to the user rather than
being retried.
Wrapping the call in the pacer fixes the problem and ensures it is
retried properly using the --low-level-retries mechanism.
When doing a multipart upload or copy, if a InvalidBlobOrBlock error
is received, it can mean that there are uncomitted blocks from a
previous failed attempt with a different length of ID.
This patch makes rclone attempt to clear the uncomitted blocks and
retry if it receives this error.
This implements multipart server side copy to improve copying from one
azure region to another by orders of magnitude (from 30s for a 100M
file to 10s for a 10G file with --azureblob-upload-concurrency 500).
- Add `--azureblob-copy-cutoff` to control the cutoff from single to multipart copy
- Add `--azureblob-copy-concurrency` to control the copy concurrency
- Add ServerSideAcrossConfigs flag as this now works properly
- Implement multipart copy using put block list API
- Shortcut multipart copy for same storage account
- Override with `--azureblob-use-copy-blob`
Fixes#8249
This speeds up server side copies for small files which need the check
the copy status by using an exponential ramp up of time to check the
copy status endpoint.
Before this change, if a multipart upload was aborted, then rclone
would leave uncommitted blocks lying around. Azure has a limit of
100,000 uncommitted blocks per storage account, so when you then try
to upload other stuff into that account, or simply the same file
again, you can run into this limit. This causes errors like the
following:
BlockCountExceedsLimit: The uncommitted block count cannot exceed the
maximum limit of 100,000 blocks.
This change removes the uncommitted blocks if a multipart upload is
aborted or fails.
If there was an existing destination file, it takes care not to
overwrite it by recomitting already comitted blocks.
This means that the scheme for allocating block IDs had to change to
make them different for each block and each upload.
Fixes#5583
`remote` has been converted ToStandardPath a few lines above, so `directory`
needs to be converted the same way in order to be compared properly. This was
spotted on `TestBisyncRemoteRemote/extended_filenames` for
`TestS3,directory_markers:` and `TestGoogleCloudStorage,directory_markers:`
which tripped over a directory name containing a Line Feed symbol.
It was reported that rclone copy occasionally uploaded corrupted data
to azure blob.
This turned out to be a race condition updating the block count which
caused blocks to be duplicated.
This bug was introduced in this commit in v1.64.0 and will be fixed in v1.65.2
0427177857 azureblob: implement OpenChunkWriter and multi-thread uploads #7056
This race only seems to happen if `--checksum` is used but can happen otherwise.
Unfortunately Azure blob does not check the MD5 that we send them so
despite sending incorrect data this corruption is not detected. The
corruption is detected when rclone tries to download the file, so
attempting to copy the files back to local disk will result in errors
such as:
ERROR : file.pokosuf5.partial: corrupted on transfer: md5 hash differ "XXX" vs "YYY"
This adds a check to test the blocklist we upload is as we expected
which would have caught the problem had it been in place earlier.
Nick Craig-Wood
Edward Klesel
acoeur
Duncan F
Cliff Frey
russcoss
dougal
Vikas Bhansali
Trevor Starick
albertony
nielash