* Automatically elide empty inputs and outputs to backend methods
* Refactor service generation
Despite the fact that the majority of our Protobuf service methods require
an open collection, they were not accessible with just a Collection
object. To access the methods (e.g. because we haven't gotten around to
exposing the correct API in Collection yet), you had to wrap the collection
in a Backend object, and pay a mutex-acquisition cost for each call, even
if you have exclusive access to the object.
This commit migrates the majority of service methods to the Collection, so
they can now be used directly, and improves the ergonomics a bit at the
same time.
The approach taken:
- The service generation now happens in rslib instead of anki_proto, which
avoids the need for trait constraints and associated types.
- Service methods are assumed to be collection-based by default. Instead of
implementing the service on Backend, we now implement it on Collection, which
means our methods no longer need to use self.with_col(...).
- We automatically generate methods in Backend which use self.with_col() to
delegate to the Collection method.
- For methods that are only appropriate for the backend, we add a flag in
the .proto file. The codegen uses this flag to write the method into a
BackendFooService instead of FooService, which the backend implements.
- The flag can also allows us to define separate implementations for collection
and backend, so we can e.g. skip the collection mutex in the i18n service
while also providing the service on a collection.
This PR replaces the existing Python-driven sync server with a new one in Rust.
The new server supports both collection and media syncing, and is compatible
with both the new protocol mentioned below, and older clients. A setting has
been added to the preferences screen to point Anki to a local server, and a
similar setting is likely to come to AnkiMobile soon.
Documentation is available here: <https://docs.ankiweb.net/sync-server.html>
In addition to the new server and refactoring, this PR also makes changes to the
sync protocol. The existing sync protocol places payloads and metadata inside a
multipart POST body, which causes a few headaches:
- Legacy clients build the request in a non-deterministic order, meaning the
entire request needs to be scanned to extract the metadata.
- Reqwest's multipart API directly writes the multipart body, without exposing
the resulting stream to us, making it harder to track the progress of the
transfer. We've been relying on a patched version of reqwest for timeouts,
which is a pain to keep up to date.
To address these issues, the metadata is now sent in a HTTP header, with the
data payload sent directly in the body. Instead of the slower gzip, we now
use zstd. The old timeout handling code has been replaced with a new implementation
that wraps the request and response body streams to track progress, allowing us
to drop the git dependencies for reqwest, hyper-timeout and tokio-io-timeout.
The main other change to the protocol is that one-way syncs no longer need to
downgrade the collection to schema 11 prior to sending.
The Rust community appear to have converged on tracing - it's used by
the Rust compiler, and receives close to 10x the number of downloads
that slog does. Its API is more ergonomic, and it does a much nicer
job with async rust.
To make this change, we no longer pass around explicit loggers, and rely
on a globally-registered one. The log file location has been changed
from one in each profile folder to a single one in the base folder. This
will remain empty for most users, since only errors are logged by default,
but may be useful for debugging future changes.
(for upgrading users, please see the notes at the bottom)
Bazel brought a lot of nice things to the table, such as rebuilds based on
content changes instead of modification times, caching of build products,
detection of incorrect build rules via a sandbox, and so on. Rewriting the build
in Bazel was also an opportunity to improve on the Makefile-based build we had
prior, which was pretty poor: most dependencies were external or not pinned, and
the build graph was poorly defined and mostly serialized. It was not uncommon
for fresh checkouts to fail due to floating dependencies, or for things to break
when trying to switch to an older commit.
For day-to-day development, I think Bazel served us reasonably well - we could
generally switch between branches while being confident that builds would be
correct and reasonably fast, and not require full rebuilds (except on Windows,
where the lack of a sandbox and the TS rules would cause build breakages when TS
files were renamed/removed).
Bazel achieves that reliability by defining rules for each programming language
that define how source files should be turned into outputs. For the rules to
work with Bazel's sandboxing approach, they often have to reimplement or
partially bypass the standard tools that each programming language provides. The
Rust rules call Rust's compiler directly for example, instead of using Cargo,
and the Python rules extract each PyPi package into a separate folder that gets
added to sys.path.
These separate language rules allow proper declaration of inputs and outputs,
and offer some advantages such as caching of build products and fine-grained
dependency installation. But they also bring some downsides:
- The rules don't always support use-cases/platforms that the standard language
tools do, meaning they need to be patched to be used. I've had to contribute a
number of patches to the Rust, Python and JS rules to unblock various issues.
- The dependencies we use with each language sometimes make assumptions that do
not hold in Bazel, meaning they either need to be pinned or patched, or the
language rules need to be adjusted to accommodate them.
I was hopeful that after the initial setup work, things would be relatively
smooth-sailing. Unfortunately, that has not proved to be the case. Things
frequently broke when dependencies or the language rules were updated, and I
began to get frustrated at the amount of Anki development time I was instead
spending on build system upkeep. It's now about 2 years since switching to
Bazel, and I think it's time to cut losses, and switch to something else that's
a better fit.
The new build system is based on a small build tool called Ninja, and some
custom Rust code in build/. This means that to build Anki, Bazel is no longer
required, but Ninja and Rust need to be installed on your system. Python and
Node toolchains are automatically downloaded like in Bazel.
This new build system should result in faster builds in some cases:
- Because we're using cargo to build now, Rust builds are able to take advantage
of pipelining and incremental debug builds, which we didn't have with Bazel.
It's also easier to override the default linker on Linux/macOS, which can
further improve speeds.
- External Rust crates are now built with opt=1, which improves performance
of debug builds.
- Esbuild is now used to transpile TypeScript, instead of invoking the TypeScript
compiler. This results in faster builds, by deferring typechecking to test/check
time, and by allowing more work to happen in parallel.
As an example of the differences, when testing with the mold linker on Linux,
adding a new message to tags.proto (which triggers a recompile of the bulk of
the Rust and TypeScript code) results in a compile that goes from about 22s on
Bazel to about 7s in the new system. With the standard linker, it's about 9s.
Some other changes of note:
- Our Rust workspace now uses cargo-hakari to ensure all packages agree on
available features, preventing unnecessary rebuilds.
- pylib/anki is now a PEP420 implicit namespace, avoiding the need to merge
source files and generated files into a single folder for running. By telling
VSCode about the extra search path, code completion now works with generated
files without needing to symlink them into the source folder.
- qt/aqt can't use PEP420 as it's difficult to get rid of aqt/__init__.py.
Instead, the generated files are now placed in a separate _aqt package that's
added to the path.
- ts/lib is now exposed as @tslib, so the source code and generated code can be
provided under the same namespace without a merging step.
- MyPy and PyLint are now invoked once for the entire codebase.
- dprint will be used to format TypeScript/json files in the future instead of
the slower prettier (currently turned off to avoid causing conflicts). It can
automatically defer to prettier when formatting Svelte files.
- svelte-check is now used for typechecking our Svelte code, which revealed a
few typing issues that went undetected with the old system.
- The Jest unit tests now work on Windows as well.
If you're upgrading from Bazel, updated usage instructions are in docs/development.md and docs/build.md. A summary of the changes:
- please remove node_modules and .bazel
- install rustup (https://rustup.rs/)
- install rsync if not already installed (on windows, use pacman - see docs/windows.md)
- install Ninja (unzip from https://github.com/ninja-build/ninja/releases/tag/v1.11.1 and
place on your path, or from your distro/homebrew if it's 1.10+)
- update .vscode/settings.json from .vscode.dist
* Add zstd dep
* Implement backend backup with zstd
* Implement backup thinning
* Write backup meta
* Use new file ending anki21b
* Asynchronously backup on collection close in Rust
* Revert "Add zstd dep"
This reverts commit 3fcb2141d2be15f907269d13275c41971431385c.
* Add zstd again
* Take backup col path from col struct
* Fix formatting
* Implement backup restoring on backend
* Normalize restored media file names
* Refactor `extract_legacy_data()`
A bit cumbersome due to borrowing rules.
* Refactor
* Make thinning calendar-based and gradual
* Consider last kept backups of previous stages
* Import full apkgs and colpkgs with backend
* Expose new backup settings
* Test `BackupThinner` and make it deterministic
* Mark backup_path when closing optional
* Delete leaky timer
* Add progress updates for restoring media
* Write restored collection to tempfile first
* Do collection compression in the background thread
This has us currently storing an uncompressed and compressed copy of
the collection in memory (not ideal), but means the collection can be
closed without waiting for compression to complete. On a large collection,
this takes a close and reopen from about 0.55s to about 0.07s. The old
backup code for comparison: about 0.35s for compression off, about
8.5s for zip compression.
* Use multithreading in zstd compression
On my system, this reduces the compression time of a large collection
from about 0.55s to 0.08s.
* Stream compressed collection data into zip file
* Tweak backup explanation
+ Fix incorrect tab order for ignore accents option
* Decouple restoring backup and full import
In the first case, no profile is opened, unless the new collection
succeeds to load.
In the second case, either the old collection is reloaded or the new one
is loaded.
* Fix number gap in Progress message
* Don't revert backup when media fails but report it
* Tweak error flow
* Remove native BackupLimits enum
* Fix type annotation
* Add thinning test for whole year
* Satisfy linter
* Await async backup to finish
* Move restart disclaimer out of backup tab
Should be visible regardless of the current tab.
* Write restored collection in chunks
* Refactor
* Write media in chunks and refactor
* Log error if removing file fails
* join_backup_task -> await_backup_completion
* Refactor backup.rs
* Refactor backup meta and collection extraction
* Fix wrong error being returned
* Call sync_all() on new collection
* Add ImportError
* Store logger in Backend, instead of creating one on demand
init_backend() accepts a Logger rather than a log file, to allow other
callers to customize the logger if they wish.
In the future we may want to explore using the tracing crate as an
alternative; it's a bit more ergonomic, as a logger doesn't need to be
passed around, and it plays more nicely with async code.
* Sync file contents prior to rename; sync folder after rename.
* Limit backup creation to once per 30 min
* Use zstd::stream::copy_decode
* Make importing abortable
* Don't revert if backup media is aborted
* Set throttle implicitly
* Change force flag to minimum_backup_interval
* Don't attempt to open folders on Windows
* Join last backup thread before starting new one
Also refactor.
* Disable auto sync and backup when restoring again
* Force backup on full download
* Include the reason why a media file import failed, and the file path
- Introduce a FileIoError that contains a string representation of
the underlying I/O error, and an associated path. There are a few
places in the code where we're currently manually including the filename
in a custom error message, and this is a step towards a more consistent
approach (but we may be better served with a more general approach in
the future similar to Anyhow's .context())
- Move the error message into importing.ftl, as it's a bit neater
when error messages live in the same file as the rest of the messages
associated with some functionality.
* Fix importing of media files
* Minor wording tweaks
* Save an allocation
I18n strings with replacements are already strings, so we can skip the
extra allocation. Not that it matters here at all.
* Terminate import if file missing from archive
If a third-party tool is creating invalid archives, the user should know
about it. This should be rare, so I did not attempt to make it
translatable.
* Skip multithreaded compression on small collections
Co-authored-by: Damien Elmes <gpg@ankiweb.net>
Rust requires all methods of impl Trait to be in a single file, which
means we had a giant backend/mod.rs covering all exposed methods. By
using separate service definitions for the separate areas, and updating
the code generation, we can split it into more manageable chunks -
this commit starts with the scheduling code.
In the long run, we'll probably want to split up the protobuf file into
multiple files as well.
Also dropped want_release_gil() from rsbridge, and the associated method
enum. While it allows us to skip the thread save/restore and mutex unlock/
lock, it looks to only be buying about 2.5% extra performance in the
best case (tested with timeit+format_timespan), and the majority of
the backend methods deal with I/O, and thus were already releasing the
GIL.
- SearchTerm -> SearchNode
- Operator -> Joiner; share between messages
- build_search_string() supports specifying AND/OR as a convenience
- group_searches() makes it easier to negate
While implementing the overdue search, I realised it would be nice to
be able to construct a search string with OR and NOT searches without
having to construct each part individually with build_search_string().
Changes:
- Extends SearchTerm to support a text search, which will be parsed
by the backend. This allows us to do things like wrap text in a group
or NOT node.
- Because SearchTerm->Node conversion can now fail with a parsing error,
it's switched over to TryFrom
- Switch concatenate_searches and replace_search_term to use SearchTerms,
so that they too don't require separate string building steps.
- Remove the unused normalize_search()
- Remove negate_search, as this is now an operation on a Node, and
users can wrap their search in SearchTerm(negated=...)
- Remove the match_any and negate args from build_search_string
Having done all this work, I've just realised that perhaps the original
JSON idea was more feasible than I first thought - if we wrote it out
to a string and re-parsed it, we would be able to leverage the existing
checks that occur at parsing stage.
- use the TimestampSecs newtype instead of raw i64s
- use FixedOffset instead of a minutes_west offset
- check localOffset each time the timing is calculated, and set it
if it's stale - even for v1.
- check for and fix missing rollover when calculating timing
- stop explicitly passing localOffset in the sync/start call
