* Configure compressor for k and x prefixed content
Adds metadata compression setting to policy
Add support to configure compressor for k and x prefixed content
Set zstd-fastest as the default compressor for metadata in the policy
Adds support to set and show metadata compression to kopia policy commands
Adds metadata compression config to dir writer
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* Pass concatenate options with ConcatenateOptions struct
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* Move content compression handling to caller
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* Move handling manifests to manifest pkg
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* Correct const in server_test
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* Remove unnecessary whitespace
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* Disable metadata compression for < V2 format
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
---------
Signed-off-by: Prasad Ghangal <prasad.ganghal@veeam.com>
* feat(snapshots): improved performance when uploading huge files
This is controlled by an upload policy which specifies the size
threshold above which indvidual files are uploaded in parts
and concatenated.
This allows multiple threads to run splitting, hashing, compression
and encryption in parallel, which was previously only possible across
multiple files, but not when a single file was being uploaded.
The default is 2GiB for now, so this feature only kicks in for very
larger files. In the future we may lower this.
Benchmark involved uploading a single 42.1 GB file which was a VM disk
snapshot of fresh Ubuntu installation (fresh EXT4 partition with lots
of zero bytes) to a brand-new filesystem repository on local SSD of
M1 Pro Macbook Pro 2021.
* before: 59-63s (~700 MB/s)
* after: 15-17s (~2.6 GB/s)
* additional test to ensure files are really e2e readable
* refactor(repository): ensure we always parse content.ID and object.ID
This changes the types to be incompatible with string to prevent direct
conversion to and from string.
This has the additional benefit of reducing number of memory allocations
and bytes for all IDs.
content.ID went from 2 allocations to 1:
typical case 32 characters + 16 bytes per-string overhead
worst-case 65 characters + 16 bytes per-string overhead
now: 34 bytes
object.ID went from 2 allocations to 1:
typical case 32 characters + 16 bytes per-string overhead
worst-case 65 characters + 16 bytes per-string overhead
now: 36 bytes
* move index.{ID,IDRange} methods to separate files
* replaced index.IDFromHash with content.IDFromHash externally
* minor tweaks and additional tests
* Update repo/content/index/id_test.go
Co-authored-by: Julio Lopez <1953782+julio-lopez@users.noreply.github.com>
* Update repo/content/index/id_test.go
Co-authored-by: Julio Lopez <1953782+julio-lopez@users.noreply.github.com>
* pr feedback
* post-merge fixes
* pr feedback
* pr feedback
* fixed subtle regression in sortedContents()
This was actually not producing invalid results because of how base36
works, just not sorting as efficiently as it could.
Co-authored-by: Julio Lopez <1953782+julio-lopez@users.noreply.github.com>
From https://github.com/google/gvisor/tree/master/tools/checklocks
This will perform static verification that we're using
`sync.Mutex`, `sync.RWMutex` and `atomic` correctly to guard access
to certain fields.
This was mostly just a matter of adding annotations to indicate which
fields are guarded by which mutex.
In a handful of places the code had to be refactored to allow static
analyzer to do its job better or to not be confused by some
constructs.
In one place this actually uncovered a bug where a function was not
releasing a lock properly in an error case.
The check is part of `make lint` but can also be invoked by
`make check-locks`.
* cli: added a flag to create repository with v2 index features
* content: plumb through compression.ID parameter to content.Manager.WriteContent()
* content: expose content.Manager.SupportsContentCompression
This allows object manager to decide whether to create compressed object
or let the content manager do it.
* object: if compression is requested and the repo supports it, pass compression ID to the content manager
* cli: show compression status in 'repository status'
* cli: output compression information in 'content list' and 'content stats'
* content: compression and decompression support
* content: unit tests for compression
* object: compression tests
* testing: added integration tests against v2 index
* testing: run all e2e tests with and without content-level compression
* htmlui: added UI for specifying index format on creation
* cli: additional tests for 'content ls' and 'content stats'
* applied pr suggestions
Removed Warning, Notify and Fatal:
* `Warning` => `Error` or `Info`
* `Notify` => `Info`
* `Fatal` was never used.
Note that --log-level=warning is still supported for backwards
compatibility, but it is the same as --log-level=error.
Co-authored-by: Julio López <julio+gh@kasten.io>
* object: fixed race condition between Result() and Checkpoint()
This would sometimes result in indirect objects having empty object IDs.
Fixes#648
* upload: ensure checkpoints never containt empty object IDs.
* testing: reduce armhf test weight
* object: refactored writer to detect split points before writing
This introduces new primitive that will be moved into splitters
themselves in subsequent commits. I'm doing this in small steps to
ensure we don't regress at any time.
* splitter: refactored TestSplitters test
This is use slow (byte-by-byte) and fast (nextSplitPoint) methods of
determining split points.
Note nextSplitPoint is not implemented by splitters yet, but this
verifies that the test is expecting the right thing.
* object: splitter refactoring - replaced ShouldSplit() with NextSplitPoint() everywhere, still not optimized
* splitter: added additional dimension to splitter_test
We split either in large chunks or one byte at a time to catch
the corner cases in the splitter implementation.
* splitter: optimized splitters using NextSplitPoint primitive
This improves splitter performance by about 40% (buzhash) and makes
it virtually free for FIXED splitter.
When prefix is not specified on ObjectWriter, we force
'x' content prefix on intermediate contents, so object IDs
will look like:
Ix{hash}
This ensures the index contents will be stored in `q` blobs,
making `snapshot gc` easier.
* object: added AsyncWrites to ObjectWriter, which improves performance of uploading of a single file
Fixes#351
Co-Authored-By: Julio López <julio+gh@kasten.io>
* performance: added buf.Pool which can be used to manage ephemeral buffers for encryption and compression
* repo: switched object writer to buf.Pool
* content: switched encryption to use buf.Pool
* object: switched compression to use buf.Pool
* testing: added missing content manager Close()
- added pooled splitters and ability to reset them without having to recreate
- added support for caller-provided compressor output to be able to pool it
- added pooling of compressor instances, since those are costly
this effectively defeated the purpose of compression, caused high
memory usage and other kinds of bad behavior.
refactored the code to prevent this issue by resetting the buffer
at the caller not callee.
fixed previous e2e test to catch the issue mentioned in #166,
verified it fails against master and passes with this change.
Also introduced strongly typed content.ID and manifest.ID (instead of string)
This aligns identifiers across all layers of repository:
blob.ID
content.ID
object.ID
manifest.ID
The splitter in question was depending on
github.com/silvasur/buzhash which is not licensed according to FOSSA bot
Switched to new faster implementation of buzhash, which is
unfortunately incompatible and will split the objects in different
places.
This change is be semi-breaking - old repositories can be read, but
when uploading large objects they will be re-uploaded where previously
they would be de-duped.
Also added 'benchmark splitters' subcommand and moved 'block cryptobenchmark'
subcommand to 'benchmark crypto'.
completely rewrote password storage:
- by default passwords are kept in OS-specific keyring (Keychain on macOS,
Windows Credentials Manager on Windows), which can be optionally disabled
to store password in a local file.
- on Linux keychain is disabled by default (does not work reliably
in terminal sessions), but can be enabled using command-line flag.
