Rust bindings to the Dropbox APIv2, generated by Stone from the official spec.
The Stone SDK and Dropbox API spec used to generate the code are in the stone
and dropbox-api-spec
submodules, respectively. Use git submodule init
and
git submodule update
to fetch them.
The generated code is checked in under src/generated
in order to simplify
building. To regenerate or update it, run python generate.py
. Doing so
requires a working Python environment and some dependencies. See the Stone
documentation for details.
This SDK is not yet official. What does this mean?
- There is no formal Dropbox support for the SDK at this point.
- Bugs may or may not get fixed.
- Not all SDK features may be implemented.
However, that said,
- The SDK is usable!
- We are happy to get feedback and/or pull requests from the community! See contributing for more information.
To actually use the API calls, you need a HTTP client -- all functions take a
type that implements HttpClient
as their first argument. This trait is
located at dropbox_sdk::client_trait::HttpClient
. Implement this trait and
pass it as the client argument.
If you don't want to implement your own, this SDK comes with an optional
default client that uses ureq
and rustls
. To use it, build with the
default_client
feature flag, and then there will be a set of clents in the
dropbox_sdk::default_client
module that you can use, corresponding to each of
the authentication types Dropbox uses (see below). The default client needs a
Dropbox API token; how you get one is up to you and your program. See the
programs under examples/ for examples, and see the helper code in
the oauth2 module.
The Dropbox API has a number of different authentication types. Each route
requires a HTTP client compatible with the specific authentication type needed.
The authentication type is designated by implementing a marker trait in
addition to the base HttpClient
trait: one of NoauthClient
,
UserAuthClient
, TeamAuthClient
, or AppAuthClient
.
The default client has implementations of all of these (except for
AppAuthClient
currently). They all share a common implementation and differ
only in which HTTP headers they add to the request.
If you only use a subset of the API, and you want to cut down on the compile
time, you can explicitly specify features corresponding to the namespaces you
need. For each namespace there is a corresponding feature dbx_{whatever}
. The
set of features can be updated if needed using the update_manifest.py
script.
An example that only needs the 'files' and 'users' namespaces:
[dependencies.dropbox-sdk]
version = "*"
default_features = false
features = ["dbx_files", "dbx_users"]
Routes return a nested result type: Result<Result<T, E>, dropbox_sdk::Error>
.
The outer Result
is Err
if something went wrong in the course of actually
making the request, such as network I/O errors or failure to serialize or
deserialize the request data. This Result
's Ok
variant is another Result
where the Ok
value is the deserialized successful result of the call, and the
Err
value is the strongly-typed error returned by the API. This inner error
indicates some problem with the request, such as file not found, lacking
permissions, etc.
The rationale for splitting the errors this way is that the former category
usually can't be handled in any way other than by retrying the request, whereas
the latter category indicate problems with the actual request itself and
probably should not be retried. Since most callers can't handle I/O errors in
any sensible way, this allows them to use the ?
syntax to pass it up the
stack, while still handling errors returned by the server.
The tests are auto-generated from the spec as well, but unlike the main code,
are not checked in. Run python generate.py
to generate the tests, and cargo test
to run them.
The test generator starts by generating a reference Python SDK and loading that code. It then generates an instance of every type in the SDK and uses the Python code to serialize them to JSON. Then it emits Rust tests that contain the JSON as a string, deserialize it, assert that all fields contain the expected values, re-serialize it, deserialize it again, and assert the fields again. Thus we have reasonably good coverage of the serialization and deserialization logic that the Rust generator emits, checked against the Python implementation (which is what Dropbox uses server-side).
Some implementation notes, limitations, and TODOs:
- Stone allows structures to inherit from other structures and be polymorphic.
Rust doesn't have these paradigms, so instead this SDK represents
polymorphic parent structs as enums, and the inherited fields are put in all
variants. See
dropbox_sdk::files::Metadata
for an example. - This crate only supports synchronous I/O. Eventually we probably want to support async I/O, which will require making incompatible changes to the types returned by routes. This should probably wait until the futures ecosystem and async/await have stabilized some more.
- This code does not use
serde_derive
for the most part, and instead uses manually-emitted serialization code. Previous work on this crate did attempt to useserde_derive
, but the way the Dropbox API serializes unions containing structs (by collapsing their fields into the union) isn't supported byserde_derive
. It also took an extremely long time to compile (~30 minutes for release build) and huge (~190MB) .rlib files. The hand-written code is more versatile, compiles faster, and produces a smaller binary, at the expense of making the generated source code much larger. - Types with constraints (such as strings with patterns or min/max lengths, or integers with a range) do not check that the data being stored in them meets the constraints.