Currently all blobs are uploaded serially when ingesting NARs. If a NAR
contains many, small blobs, ingestion may become slow if there is a lot
of round-trip latency to the blob service.
This makes the NAR ingester use the ConcurrentBlobUploader which allows
for buffering small blobs in memory so they can be uploaded concurrently
to the blob service without blocking further deserialization.
Change-Id: I093a73770232df12d9a11e5d901b99c08505c3cb
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11694
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
This shouldn't be part of the PathInfoService trait.
Pretty much none of the PathInfoServices do implement it, and requiring
them to implement it means they also cannot make use of this calculation
already being done by other PathInfoServices.
Move it out into its own NarCalculationService trait, defined somewhere
at tvix_store::nar, and have everyone who wants to trigger nar
calculation use nar_calculation_service directly, which now is an
additional field in TvixStoreIO for example.
It being moved outside the PathInfoService trait doesn't prohibit
specific implementations to implement it (like the GRPC client for the
`PathInfoService` does.
This is currently wired together in a bit of a hacky fashion - as of
now, everything uses the naive implementation that traverses blob and
directoryservice, rather than composing it properly. I want to leave
that up to a later CL, dealing with other parts of store composition
too.
Change-Id: I18d07ea4301d4a07651b8218bc5fe95e4e307208
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11619
Reviewed-by: Connor Brewster <cbrewster@hey.com>
Autosubmit: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
We reuse the prev_name allocation for Entry, instead of allocating and
returning a separate Vec.
We encode the `prev_name: None` case as an empty vector, since we don't
allow empty names anyway, and the sorting is equivalent.
Change-Id: I975b37ff873805f5ff099bc82128706891052247
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11607
Reviewed-by: Brian Olsen <me@griff.name>
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
If our underlying reader supports AsyncBufRead, then we can too.
Change-Id: If4b948c983400ca591c1c475bbcf7dc00d562040
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11545
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Rust 1.77 supports async recursion as long as there is some form of
indirection (ie. `Box::pin`). This removes the need to use the
async-recursion crate.
Change-Id: Ic9613ab7f32016f0103032a861edff92e2fb8b41
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11596
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Connor Brewster <cbrewster@hey.com>
Tested-by: BuildkiteCI
Rename read_nar to ingest_nar, and have it use the async nar reader
version, and the ingest_entries machinery.
This means we can now drop all code dealing with manually assembling
castore nodes.
Update our consumer, NixHTTPPathInfoService to use the new API.
As we now accept an AsyncRead, we don't need to do any blocking here
anymore, and can use the same async-compression crate as in the fetching
logic (and support some more compression formats out of the box).
Change-Id: I8646d20bd8603f8da47b5c84bc9e4ac236eb7f1a
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11580
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: Connor Brewster <cbrewster@hey.com>
We implement DirectoryService for Arc<DirectoryService> and
Box<DirectoryService>, this is sufficient.
Change-Id: I0a5a81cbc4782764406b5bca57f908ace6090737
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11586
Tested-by: BuildkiteCI
Reviewed-by: Connor Brewster <cbrewster@hey.com>
There's little reason to keep the nar writer using Async{Read,Write}
traits from futures, while everything else async in tvix (and
nix-compat) uses tokio.
Change-Id: I8cd1efcd0dd5bb76471de997603c7b701a5095de
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11391
Tested-by: BuildkiteCI
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Reviewed-by: Brian Olsen <me@griff.name>
Now, it supports almost everything except `recursive = false;`, i.e. `flat`-ingestion
because we have no knob exposed in the tvix store import side to do it.
This has been tested to work.
Change-Id: I2e9da10ceccdfbf45b43c532077ed45d6306aa98
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10597
Tested-by: BuildkiteCI
Autosubmit: raitobezarius <tvl@lahfa.xyz>
Reviewed-by: flokli <flokli@flokli.de>
Previous CLs ensured BlobService and DirectoryService is implemented on
AsRef<dyn BlobService> and AsRef<dyn DirectoryService> respectively, so
there's no need to require AsRef in here.
Change-Id: I1e8b93188ca784e1b9a592f3d35fcf6b9df75c75
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11275
Tested-by: BuildkiteCI
Reviewed-by: Connor Brewster <cbrewster@hey.com>
This stops using the gen_{blob,directory}_service() functions from
tvix_castore::utils in favor of the rstest-based fixtures.
Change-Id: I4c2fff2926f34af63d3bd5c29c8e939dbdf7e062
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11274
Reviewed-by: Connor Brewster <cbrewster@hey.com>
Tested-by: BuildkiteCI
Ensure consistent formatting of blake3 digests.
It looks like we don't actually construct these error types anywhere, so
no code needs to be refactored to it currently, but still good to be
consistent.
Change-Id: I49dc8a7f3cb4245ac06b9a6a44b72060434a3d32
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11193
Reviewed-by: Connor Brewster <cbrewster@hey.com>
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
To render NARs, we're fine with a simple AsRef to a BlobService and
DirectoryService. We just need to have the function pass back the
references, so we can reuse it after the recursion.
Change-Id: I8a1b899134ddda26cf14aa829a08383986101850
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10605
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
We need to be a bit careful and pass the BlobService around (similar to
how we already do with the directory_putter), but that allows getting
rid of a bunch of annoying trait bounds.
We also stop spawning additional tasks where we can just use block_on.
Change-Id: If36de0ee947d2c779d20a384308241d2262d4764
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10580
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
Only convert to and reuse an Arc<…> where needed.
Change-Id: I2c1bc69cca5a4a3ebd3bdb33d6e28e1f5fb86cb9
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10514
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
For all these calls, the caller has enough context about what it did, so
it should be fine to use io::Result here.
We pretty much only constructed crate::Error::StorageError before
anyways, so this conveys *more* information.
Change-Id: I5cabb3769c9c2314bab926d34dda748fda9d3ccc
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10328
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
Allow taking advantage of the buffer of the underlying reader to avoid
unnecessary copies of file data.
We can't easily implement the methods of BufRead directly, since we
have some extra I/O to perform in the final consume() invocation.
That could be resolved at the cost of additional bookkeeping, but this
will suffice for now.
Change-Id: I8100cf0abd79e7469670b8596bd989be5db44a91
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10089
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
We rely on being able to make small reads cheaply, so this was already
an implicit practical requirement. Requiring it explicitly removes a
performance footgun, and makes further optimisations possible.
Change-Id: I7f65880a41b1d6b5e6bf2e52dfe47d4c49b34bcd
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10088
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
This adds a function that can ingest a NAR file into a given
blob_service and directory_service, returning the root node.
Passing around the directory_putter is a bit annoying, but at least for
this linear "reading through a NAR" case, we don't need it to be clone
or sync.
Change-Id: Ia26d596f5dffc6dbd69aa39a82b59bba41685a9c
Reviewed-on: https://cl.tvl.fyi/c/depot/+/9915
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
Having more than 4GiB files is quite possible (think about the NixOS
graphical installer, and an uncompressed iso of it).
No wire format changes.
Change-Id: Ia78a07e4c554e91b93c5b9f8533266e4bd7f22b6
Reviewed-on: https://cl.tvl.fyi/c/depot/+/9950
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
This splits the pure content-addressed layers from tvix-store into a
`castore` crate, and only leaves PathInfo related things, as well as the
CLI entrypoint in the tvix-store crate.
Notable changes:
- `fixtures` and `utils` had to be moved out of the `test` cfg, so they
can be imported from tvix-store.
- Some ad-hoc fixtures in the test were moved to proper fixtures in the
same step.
- The protos are now created by a (more static) recipe in the protos/
directory.
The (now two) golang targets are commented out, as it's not possible to
update them properly in the same CL. This will be done by a followup CL
once this is merged (and whitby deployed)
Bug: https://b.tvl.fyi/issues/301
Change-Id: I8d675d4bf1fb697eb7d479747c1b1e3635718107
Reviewed-on: https://cl.tvl.fyi/c/depot/+/9370
Reviewed-by: tazjin <tazjin@tvl.su>
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Reviewed-by: Connor Brewster <cbrewster@hey.com>
We've decided to asyncify all of the services to reduce some of the
pains going back and for between sync<->async. The end goal will be for
all the tvix-store internals to be async and then expose a sync
interface for things like tvix eval io.
Change-Id: I97c71f8db1d05a38bd8f625df5087d565705d52d
Reviewed-on: https://cl.tvl.fyi/c/depot/+/9369
Autosubmit: Connor Brewster <cbrewster@hey.com>
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
We previously kept the trait of a BlobService sync.
This however had some annoying consequences:
- It became more and more complicated to track when we're in a context
with an async runtime in the context or not, producing bugs like
https://b.tvl.fyi/issues/304
- The sync trait shielded away async clients from async worloads,
requiring manual block_on code inside the gRPC client code, and
spawn_blocking calls in consumers of the trait, even if they were
async (like the gRPC server)
- We had to write our own custom glue code (SyncReadIntoAsyncRead)
to convert a sync io::Read into a tokio::io::AsyncRead, which already
existed in tokio internally, but upstream ia hesitant to expose.
This now makes the BlobService trait async (via the async_trait macro,
like we already do in various gRPC parts), and replaces the sync readers
and writers with their async counterparts.
Tests interacting with a BlobService now need to have an async runtime
available, the easiest way for this is to mark the test functions
with the tokio::test macro, allowing us to directly .await in the test
function.
In places where we don't have an async runtime available from context
(like tvix-cli), we can pass one down explicitly.
Now that we don't provide a sync interface anymore, the (sync) FUSE
library now holds a pointer to a tokio runtime handle, and needs to at
least have 2 threads available when talking to a blob service (which is
why some of the tests now use the multi_thread flavor).
The FUSE tests got a bit more verbose, as we couldn't use the
setup_and_mount function accepting a callback anymore. We can hopefully
move some of the test fixture setup to rstest in the future to make this
less repetitive.
Co-Authored-By: Connor Brewster <cbrewster@hey.com>
Change-Id: Ia0501b606e32c852d0108de9c9016b21c94a3c05
Reviewed-on: https://cl.tvl.fyi/c/depot/+/9329
Reviewed-by: Connor Brewster <cbrewster@hey.com>
Tested-by: BuildkiteCI
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
This will save us some copies, because a clone will simply create an
additional pointer to the same data.
Change-Id: I017a5d6b4c85a861b5541ebad2858ad4fbf8e8fa
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8978
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Autosubmit: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Some paths might use names that are not valid UTF-8. We should be able
to represent them.
We don't actually need to touch the PathInfo structures, as they need to
represent StorePaths, which come with their own harder restrictions,
which can't encode non-UTF8 data.
While this doesn't change any of the wire format of the gRPC messages,
it does however change the interface of tvix_eval::EvalIO - its
read_dir() method does now return a list of Vec<u8>, rather than
SmolStr. Maybe this should be OsString instead?
Change-Id: I821016d9a58ec441ee081b0b9f01c9240723af0b
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8974
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
This allows us to blob services without closing them before putting them
in a box.
We currently need to use Arc<_>, not Rc<_>, because the GRPC wrappers
require Sync.
Change-Id: I679c5f06b62304f5b0456cfefe25a0a881de7c84
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8738
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
Once we support configuring services at runtime, we don't know what
DirectoryService we're using at compile time.
This also means, we can't explicitly use the is_closed method from
GRPCPutter, without making it part of the DirectoryPutter itself.
Change-Id: Icd2a1ec4fc5649a6cd15c9cc7db4c2b473630431
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8727
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
Putting this in the PathInfoService trait makes much more sense, we can
have direct control over where/how to cache the results in the
implementation.
This now requires each PathInfoService to hold pointers to BlobService
and DirectoryService.
Change-Id: I4faae780d43eae4beeb57bd5e190e6d1a5d3314e
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8724
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: tazjin <tazjin@tvl.su>
There's only one way to calculate NAR files, by walking through them.
Things like caching such replies should be done closer to where we use
these, composing NARCalculationService doesn't actually give us much.
Instead, expose two functions, `nar::calculate_size_and_sha256` and
`nar::writer_nar`, the latter writing NAR to a writer, the former using
write_nar to only keeping the NAR size and digest.
Change-Id: Ie5d2cfea35470fdbb5cbf9da1136b0cdf0250266
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8723
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
Autosubmit: flokli <flokli@flokli.de>
To construct various stores at runtime, we need to eliminate associated
types from the BlobService trait, and return Box<dyn …> instead of
specific types.
This also means we can't consume self in the close() method, so
everything we write to is put in an Option<>, and during the first close
we take from there.
Change-Id: Ia523b6ab2f2a5276f51cb5d17e81a5925bce69b6
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8647
Autosubmit: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Reviewed-by: tazjin <tazjin@tvl.su>
This asks a remote tvix-store for the nar size and digest of a given
root node.
Change-Id: If9f916d9bfc5f8dc3166e2c6c1671c0f0124d1c1
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8611
Autosubmit: flokli <flokli@flokli.de>
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
This introduces a new struct, B3Digest, which internally holds a
Vec<u8>, but only allows construction with 32 bytes.
It also implements display, which will print the base64 representation.
This should reduce some boilerplate when parsing Vec<u8>.
Change-Id: Ia91aa40cb691916773abc8f93e6ed79a5fd34863
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8592
Reviewed-by: tazjin <tazjin@tvl.su>
Autosubmit: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Whether chunking is involved or not, is an implementation detail of each
Blobstore. Consumers of a whole blob shouldn't need to worry about that.
It currently is not visible in the gRPC interface either. It
shouldn't bleed into everything.
Let the BlobService trait provide `open_read` and `open_write` methods,
which return handles providing io::Read or io::Write, and leave the
details up to the implementation.
This means, our custom BlobReader module can go away, and all the
chunking bits in there, too.
In the future, we might still want to add more chunking-aware syncing,
but as a syncing strategy some stores can expose, not as a fundamental
protocol component.
This currently needs "SyncReadIntoAsyncRead", taken and vendored in from
https://github.com/tokio-rs/tokio/pull/5669.
It provides a AsyncRead for a sync Read, which is necessary to connect
our (sync) BlobReader interface to a GRPC server implementation.
As an alternative, we could also make the BlobReader itself async, and
let consumers of the trait (EvalIO) deal with the async-ness, but this
is less of a change for now.
In terms of vendoring, I initially tried to move our tokio crate to
these commits, but ended up in version incompatibilities, so let's
vendor it in for now.
Change-Id: I5969ebbc4c0e1ceece47981be3b9e7cfb3f59ad0
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8551
Tested-by: BuildkiteCI
Reviewed-by: tazjin <tazjin@tvl.su>
Passing in a &proto::node::Node into all this allows us consumers to
keep ownership of the proto::node::Node.
Change-Id: I44882a86c46826b06a8a8a0b24c18adfc7052662
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8316
Tested-by: BuildkiteCI
Reviewed-by: tazjin <tazjin@tvl.su>
Autosubmit: flokli <flokli@flokli.de>
This adds a NARCalculationService trait, which will take a root node,
and provide a proto::CalculateNarResponse in return.
It also adds a NonCachingNARCalculationService implementation, that will
simply always render the NAR in memory to calculate the size and sha256,
without any caching.
Change-Id: Id1ffb18559212fa6001f70f2634bbc3dfd0aa343
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8096
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
Make sure the blob size in the current proto node matches what we get
back from the blob backend.
Change-Id: I939fa18f37c7bc86ada8a495c7be622e69ec47f8
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8129
Tested-by: BuildkiteCI
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
This moves the logic rendering NARs to a struct using the
previously introduced, more granular BlobService, ChunkService and
DirectoryService.
Instead of passing them around to the helper functions, they're kept as
members of a struct.
Remove the async invocations in the nar_renderer tests, there's nothing
async in here.
Change-Id: Ic6d24aaad68a1fda46ce29f2cdb5f7b87f481d5c
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8095
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
