refactor(tvix): use composition & registry for from_addr

Change-Id: I3c94ecb5958294b5973c6fcdf5ee9c0d37fa54ad
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11976
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Autosubmit: yuka <yuka@yuka.dev>

tvix/build/src/bin/tvix-build.rs

@@ -1,5 +1,3 @@
-use std::sync::Arc;
-
 use clap::Parser;
 use clap::Subcommand;
 use tokio_listener::Listener;
@@ -51,7 +49,7 @@ enum Commands {
 }
 
 #[tokio::main]
-async fn main() -> Result<(), Box<dyn std::error::Error>> {
+async fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     let cli = Cli::parse();
 
     let _ = tvix_tracing::TracingBuilder::default()
@@ -69,11 +67,8 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
             let blob_service = blobservice::from_addr(&blob_service_addr).await?;
             let directory_service = directoryservice::from_addr(&directory_service_addr).await?;
 
-            let build_service = buildservice::from_addr(
+            let build_service =
-                &build_service_addr,
+                buildservice::from_addr(&build_service_addr, blob_service, directory_service)
-                Arc::from(blob_service),
-                Arc::from(directory_service),
-            )
                     .await?;
 
             let listen_address = listen_address

tvix/castore/src/blobservice/combinator.rs

@@ -6,7 +6,7 @@ use tonic::async_trait;
 use tracing::{instrument, warn};
 
 use crate::composition::{CompositionContext, ServiceBuilder};
-use crate::B3Digest;
+use crate::{B3Digest, Error};
 
 use super::{naive_seeker::NaiveSeeker, BlobReader, BlobService, BlobWriter};
 
@@ -103,6 +103,16 @@ pub struct CombinedBlobServiceConfig {
     remote: String,
 }
 
+impl TryFrom<url::Url> for CombinedBlobServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(_url: url::Url) -> Result<Self, Self::Error> {
+        Err(Error::StorageError(
+            "Instantiating a CombinedBlobService from a url is not supported".into(),
+        )
+        .into())
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for CombinedBlobServiceConfig {
     type Output = dyn BlobService;

tvix/castore/src/blobservice/from_addr.rs

@@ -1,8 +1,12 @@
+use std::sync::Arc;
+
 use url::Url;
 
-use crate::{proto::blob_service_client::BlobServiceClient, Error};
+use crate::composition::{
+    with_registry, CompositionContext, DeserializeWithRegistry, ServiceBuilder, REG,
+};
 
-use super::{BlobService, GRPCBlobService, MemoryBlobService, ObjectStoreBlobService};
+use super::BlobService;
 
 /// Constructs a new instance of a [BlobService] from an URI.
 ///
@@ -12,53 +16,19 @@ use super::{BlobService, GRPCBlobService, MemoryBlobService, ObjectStoreBlobServ
 /// - `objectstore+*://` ([ObjectStoreBlobService])
 ///
 /// See their `from_url` methods for more details about their syntax.
-pub async fn from_addr(uri: &str) -> Result<Box<dyn BlobService>, crate::Error> {
+pub async fn from_addr(
+    uri: &str,
+) -> Result<Arc<dyn BlobService>, Box<dyn std::error::Error + Send + Sync>> {
     let url = Url::parse(uri)
         .map_err(|e| crate::Error::StorageError(format!("unable to parse url: {}", e)))?;
 
-    let blob_service: Box<dyn BlobService> = match url.scheme() {
+    let blob_service_config = with_registry(&REG, || {
-        "memory" => {
+        <DeserializeWithRegistry<Box<dyn ServiceBuilder<Output = dyn BlobService>>>>::try_from(url)
-            // memory doesn't support host or path in the URL.
+    })?
-            if url.has_host() || !url.path().is_empty() {
+    .0;
-                return Err(Error::StorageError("invalid url".to_string()));
+    let blob_service = blob_service_config
-            }
+        .build("anonymous", &CompositionContext::blank())
-            Box::<MemoryBlobService>::default()
+        .await?;
-        }
-        scheme if scheme.starts_with("grpc+") => {
-            // schemes starting with grpc+ go to the GRPCPathInfoService.
-            //   That's normally grpc+unix for unix sockets, and grpc+http(s) for the HTTP counterparts.
-            // - In the case of unix sockets, there must be a path, but may not be a host.
-            // - In the case of non-unix sockets, there must be a host, but no path.
-            // Constructing the channel is handled by tvix_castore::channel::from_url.
-            Box::new(GRPCBlobService::from_client(
-                BlobServiceClient::with_interceptor(
-                    crate::tonic::channel_from_url(&url).await?,
-                    tvix_tracing::propagate::tonic::send_trace,
-                ),
-            ))
-        }
-        scheme if scheme.starts_with("objectstore+") => {
-            // We need to convert the URL to string, strip the prefix there, and then
-            // parse it back as url, as Url::set_scheme() rejects some of the transitions we want to do.
-            let trimmed_url = {
-                let s = url.to_string();
-                let mut url = Url::parse(s.strip_prefix("objectstore+").unwrap()).unwrap();
-                // trim the query pairs, they might contain credentials or local settings we don't want to send as-is.
-                url.set_query(None);
-                url
-            };
-            Box::new(
-                ObjectStoreBlobService::parse_url_opts(&trimmed_url, url.query_pairs())
-                    .map_err(|e| Error::StorageError(e.to_string()))?,
-            )
-        }
-        scheme => {
-            return Err(crate::Error::StorageError(format!(
-                "unknown scheme: {}",
-                scheme
-            )))
-        }
-    };
 
     Ok(blob_service)
 }

tvix/castore/src/blobservice/grpc.rs

@@ -187,6 +187,19 @@ pub struct GRPCBlobServiceConfig {
     url: String,
 }
 
+impl TryFrom<url::Url> for GRPCBlobServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        //   normally grpc+unix for unix sockets, and grpc+http(s) for the HTTP counterparts.
+        // - In the case of unix sockets, there must be a path, but may not be a host.
+        // - In the case of non-unix sockets, there must be a host, but no path.
+        // Constructing the channel is handled by tvix_castore::channel::from_url.
+        Ok(GRPCBlobServiceConfig {
+            url: url.to_string(),
+        })
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for GRPCBlobServiceConfig {
     type Output = dyn BlobService;

tvix/castore/src/blobservice/memory.rs

@@ -7,7 +7,7 @@ use tracing::instrument;
 
 use super::{BlobReader, BlobService, BlobWriter};
 use crate::composition::{CompositionContext, ServiceBuilder};
-use crate::B3Digest;
+use crate::{B3Digest, Error};
 
 #[derive(Clone, Default)]
 pub struct MemoryBlobService {
@@ -42,6 +42,17 @@ impl BlobService for MemoryBlobService {
 #[serde(deny_unknown_fields)]
 pub struct MemoryBlobServiceConfig {}
 
+impl TryFrom<url::Url> for MemoryBlobServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        // memory doesn't support host or path in the URL.
+        if url.has_host() || !url.path().is_empty() {
+            return Err(Error::StorageError("invalid url".to_string()).into());
+        }
+        Ok(MemoryBlobServiceConfig {})
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for MemoryBlobServiceConfig {
     type Output = dyn BlobService;

tvix/castore/src/blobservice/object_store.rs

@@ -21,21 +21,11 @@ use url::Url;
 use crate::{
     composition::{CompositionContext, ServiceBuilder},
     proto::{stat_blob_response::ChunkMeta, StatBlobResponse},
-    B3Digest, B3HashingReader,
+    B3Digest, B3HashingReader, Error,
 };
 
 use super::{BlobReader, BlobService, BlobWriter, ChunkedReader};
 
-#[derive(Clone)]
-pub struct ObjectStoreBlobService {
-    object_store: Arc<dyn ObjectStore>,
-    base_path: Path,
-
-    /// Average chunk size for FastCDC, in bytes.
-    /// min value is half, max value double of that number.
-    avg_chunk_size: u32,
-}
-
 /// Uses any object storage supported by the [object_store] crate to provide a
 /// tvix-castore [BlobService].
 ///
@@ -72,31 +62,14 @@ pub struct ObjectStoreBlobService {
 /// It also allows signalling any compression of chunks in the content-type.
 /// Migration *should* be possible by simply adding the right content-types to
 /// all keys stored so far, but no promises ;-)
-impl ObjectStoreBlobService {
+#[derive(Clone)]
-    /// Constructs a new [ObjectStoreBlobService] from a [Url] supported by
+pub struct ObjectStoreBlobService {
-    /// [object_store].
+    object_store: Arc<dyn ObjectStore>,
-    /// Any path suffix becomes the base path of the object store.
+    base_path: Path,
-    /// additional options, the same as in [object_store::parse_url_opts] can
-    /// be passed.
-    pub fn parse_url_opts<I, K, V>(url: &Url, options: I) -> Result<Self, object_store::Error>
-    where
-        I: IntoIterator<Item = (K, V)>,
-        K: AsRef<str>,
-        V: Into<String>,
-    {
-        let (object_store, path) = object_store::parse_url_opts(url, options)?;
 
-        Ok(Self {
+    /// Average chunk size for FastCDC, in bytes.
-            object_store: Arc::new(object_store),
+    /// min value is half, max value double of that number.
-            base_path: path,
+    avg_chunk_size: u32,
-            avg_chunk_size: 256 * 1024,
-        })
-    }
-
-    /// Like [Self::parse_url_opts], except without the options.
-    pub fn parse_url(url: &Url) -> Result<Self, object_store::Error> {
-        Self::parse_url_opts(url, Vec::<(String, String)>::new())
-    }
 }
 
 #[instrument(level=Level::TRACE, skip_all,fields(base_path=%base_path,blob.digest=%digest),ret(Display))]
@@ -281,10 +254,41 @@ pub struct ObjectStoreBlobServiceConfig {
     object_store_url: String,
     #[serde(default = "default_avg_chunk_size")]
     avg_chunk_size: u32,
-    #[serde(default)]
     object_store_options: HashMap<String, String>,
 }
 
+impl TryFrom<url::Url> for ObjectStoreBlobServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    /// Constructs a new [ObjectStoreBlobService] from a [Url] supported by
+    /// [object_store].
+    /// Any path suffix becomes the base path of the object store.
+    /// additional options, the same as in [object_store::parse_url_opts] can
+    /// be passed.
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        // We need to convert the URL to string, strip the prefix there, and then
+        // parse it back as url, as Url::set_scheme() rejects some of the transitions we want to do.
+        let trimmed_url = {
+            let s = url.to_string();
+            let mut url = Url::parse(
+                s.strip_prefix("objectstore+")
+                    .ok_or(Error::StorageError("Missing objectstore uri".into()))?,
+            )?;
+            // trim the query pairs, they might contain credentials or local settings we don't want to send as-is.
+            url.set_query(None);
+            url
+        };
+        Ok(ObjectStoreBlobServiceConfig {
+            object_store_url: trimmed_url.into(),
+            object_store_options: url
+                .query_pairs()
+                .into_iter()
+                .map(|(k, v)| (k.to_string(), v.to_string()))
+                .collect(),
+            avg_chunk_size: 256 * 1024,
+        })
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for ObjectStoreBlobServiceConfig {
     type Output = dyn BlobService;
@@ -548,7 +552,7 @@ where
 
 #[cfg(test)]
 mod test {
-    use super::chunk_and_upload;
+    use super::{chunk_and_upload, default_avg_chunk_size};
     use crate::{
         blobservice::{BlobService, ObjectStoreBlobService},
         fixtures::{BLOB_A, BLOB_A_DIGEST},
@@ -559,13 +563,18 @@ mod test {
     /// Tests chunk_and_upload directly, bypassing the BlobWriter at open_write().
     #[tokio::test]
     async fn test_chunk_and_upload() {
-        let blobsvc = Arc::new(
+        let (object_store, base_path) =
-            ObjectStoreBlobService::parse_url(&Url::parse("memory:///").unwrap()).unwrap(),
+            object_store::parse_url(&Url::parse("memory:///").unwrap()).unwrap();
-        );
+        let object_store: Arc<dyn object_store::ObjectStore> = Arc::from(object_store);
+        let blobsvc = Arc::new(ObjectStoreBlobService {
+            object_store: object_store.clone(),
+            avg_chunk_size: default_avg_chunk_size(),
+            base_path,
+        });
 
         let blob_digest = chunk_and_upload(
             &mut Cursor::new(BLOB_A.to_vec()),
-            blobsvc.object_store.clone(),
+            object_store,
             object_store::path::Path::from("/"),
             1024 / 2,
             1024,

tvix/castore/src/composition.rs

@@ -31,6 +31,13 @@
 //!     }
 //! }
 //!
+//! impl TryFrom<url::Url> for MyBlobServiceConfig {
+//!     type Error = Box<dyn std::error::Error + Send + Sync>;
+//!     fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+//!         todo!()
+//!     }
+//! }
+//!
 //! pub fn add_my_service(reg: &mut Registry) {
 //!     reg.register::<Box<dyn ServiceBuilder<Output = dyn BlobService>>, MyBlobServiceConfig>("myblobservicetype");
 //! }
@@ -100,7 +107,7 @@ use tonic::async_trait;
 // This is really ugly. Really we would want to store this as a generic static field:
 //
 // ```
-// struct Registry<T>(BTreeMap<(&'static str), BoxSeedFn<T>);
+// struct Registry<T>(BTreeMap<(&'static str), RegistryEntry<T>);
 // static REG<T>: Registry<T>;
 // ```
 //
@@ -116,6 +123,12 @@ use tonic::async_trait;
 // I said it was ugly...
 #[derive(Default)]
 pub struct Registry(BTreeMap<(TypeId, &'static str), Box<dyn Any + Sync>>);
+pub type FromUrlSeed<T> =
+    Box<dyn Fn(url::Url) -> Result<T, Box<dyn std::error::Error + Send + Sync>> + Sync>;
+pub struct RegistryEntry<T> {
+    serde_deserialize_seed: BoxFnSeed<DeserializeWithRegistry<T>>,
+    from_url_seed: FromUrlSeed<DeserializeWithRegistry<T>>,
+}
 
 struct RegistryWithFakeType<'r, T>(&'r Registry, PhantomData<T>);
 
@@ -137,7 +150,9 @@ impl<'r, 'de: 'r, T: 'static> SeedFactory<'de, TagString<'de>> for RegistryWithF
             .ok_or_else(|| serde::de::Error::custom("Unknown tag"))?
             .1;
 
-        Ok(<dyn Any>::downcast_ref(&**seed).unwrap())
+        let entry: &RegistryEntry<T> = <dyn Any>::downcast_ref(&**seed).unwrap();
+
+        Ok(&entry.serde_deserialize_seed)
     }
 }
 
@@ -146,7 +161,7 @@ impl<'r, 'de: 'r, T: 'static> SeedFactory<'de, TagString<'de>> for RegistryWithF
 /// Wrap your type in this in order to deserialize it using a registry, e.g.
 /// `RegistryWithFakeType<Box<dyn MyTrait>>`, then the types registered for `Box<dyn MyTrait>`
 /// will be used.
-pub struct DeserializeWithRegistry<T>(T);
+pub struct DeserializeWithRegistry<T>(pub T);
 
 impl Registry {
     /// Registers a mapping from type tag to a concrete type into the registry.
@@ -156,14 +171,30 @@ impl Registry {
     /// deserializes into an input with the type tag "myblobservicetype" into a
     /// `Box<dyn FooTrait>`, it will first call the Deserialize imple of `FooStruct` and
     /// then convert it into a `Box<dyn FooTrait>` using From::from.
-    pub fn register<T: 'static, C: DeserializeOwned + Into<T>>(&mut self, type_name: &'static str) {
+    pub fn register<
-        let seed = BoxFnSeed::new(|x| {
+        T: 'static,
+        C: DeserializeOwned
+            + TryFrom<url::Url, Error = Box<dyn std::error::Error + Send + Sync>>
+            + Into<T>,
+    >(
+        &mut self,
+        type_name: &'static str,
+    ) {
+        self.0.insert(
+            (TypeId::of::<T>(), type_name),
+            Box::new(RegistryEntry {
+                serde_deserialize_seed: BoxFnSeed::new(|x| {
                     deserialize::<C>(x)
                         .map(Into::into)
                         .map(DeserializeWithRegistry)
-        });
+                }),
-        self.0
+                from_url_seed: Box::new(|url| {
-            .insert((TypeId::of::<T>(), type_name), Box::new(seed));
+                    C::try_from(url)
+                        .map(Into::into)
+                        .map(DeserializeWithRegistry)
+                }),
+            }),
+        );
     }
 }
 
@@ -180,6 +211,30 @@ impl<'de, T: 'static> serde::Deserialize<'de> for DeserializeWithRegistry<T> {
     }
 }
 
+#[derive(Debug, thiserror::Error)]
+enum TryFromUrlError {
+    #[error("Unknown tag: {0}")]
+    UnknownTag(String),
+}
+
+impl<T: 'static> TryFrom<url::Url> for DeserializeWithRegistry<T> {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        let tag = url.scheme().split('+').next().unwrap();
+        // same as in the SeedFactory impl: using find() and not get() because of https://github.com/rust-lang/rust/issues/80389
+        let seed = ACTIVE_REG
+            .get()
+            .unwrap()
+            .0
+            .iter()
+            .find(|(k, _)| *k == &(TypeId::of::<T>(), tag))
+            .ok_or_else(|| Box::new(TryFromUrlError::UnknownTag(tag.into())))?
+            .1;
+        let entry: &RegistryEntry<T> = <dyn Any>::downcast_ref(&**seed).unwrap();
+        (entry.from_url_seed)(url)
+    }
+}
+
 thread_local! {
     /// The active Registry is global state, because there is no convenient and universal way to pass state
     /// into the functions usually used for deserialization, e.g. `serde_json::from_str`, `toml::from_str`,
@@ -200,7 +255,7 @@ pub fn with_registry<R>(reg: &'static Registry, f: impl FnOnce() -> R) -> R {
 lazy_static! {
     /// The provided registry of tvix_castore, with all builtin BlobStore/DirectoryStore implementations
     pub static ref REG: Registry = {
-        let mut reg = Registry(Default::default());
+        let mut reg = Default::default();
         add_default_services(&mut reg);
         reg
     };

tvix/castore/src/directoryservice/bigtable.rs

@@ -361,6 +361,26 @@ impl ServiceBuilder for BigtableParameters {
     }
 }
 
+impl TryFrom<url::Url> for BigtableParameters {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(mut url: url::Url) -> Result<Self, Self::Error> {
+        // parse the instance name from the hostname.
+        let instance_name = url
+            .host_str()
+            .ok_or_else(|| Error::StorageError("instance name missing".into()))?
+            .to_string();
+
+        // … but add it to the query string now, so we just need to parse that.
+        url.query_pairs_mut()
+            .append_pair("instance_name", &instance_name);
+
+        let params: BigtableParameters = serde_qs::from_str(url.query().unwrap_or_default())
+            .map_err(|e| Error::InvalidRequest(format!("failed to parse parameters: {}", e)))?;
+
+        Ok(params)
+    }
+}
+
 fn default_app_profile_id() -> String {
     "default".to_owned()
 }

tvix/castore/src/directoryservice/combinators.rs

@@ -151,6 +151,16 @@ pub struct CacheConfig {
     far: String,
 }
 
+impl TryFrom<url::Url> for CacheConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(_url: url::Url) -> Result<Self, Self::Error> {
+        Err(Error::StorageError(
+            "Instantiating a CombinedDirectoryService from a url is not supported".into(),
+        )
+        .into())
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for CacheConfig {
     type Output = dyn DirectoryService;

tvix/castore/src/directoryservice/from_addr.rs

@@ -1,12 +1,13 @@
+use std::sync::Arc;
+
 use url::Url;
 
-use crate::{proto::directory_service_client::DirectoryServiceClient, Error};
+use crate::composition::{
-
+    with_registry, CompositionContext, DeserializeWithRegistry, ServiceBuilder, REG,
-use super::{
-    DirectoryService, GRPCDirectoryService, MemoryDirectoryService, ObjectStoreDirectoryService,
-    SledDirectoryService,
 };
 
+use super::DirectoryService;
+
 /// Constructs a new instance of a [DirectoryService] from an URI.
 ///
 /// The following URIs are supported:
@@ -21,101 +22,23 @@ use super::{
 ///   Connects to a local tvix-store gRPC service via Unix socket.
 /// - `grpc+http://host:port`, `grpc+https://host:port`
 ///    Connects to a (remote) tvix-store gRPC service.
-pub async fn from_addr(uri: &str) -> Result<Box<dyn DirectoryService>, crate::Error> {
+pub async fn from_addr(
+    uri: &str,
+) -> Result<Arc<dyn DirectoryService>, Box<dyn std::error::Error + Send + Sync>> {
     #[allow(unused_mut)]
     let mut url = Url::parse(uri)
         .map_err(|e| crate::Error::StorageError(format!("unable to parse url: {}", e)))?;
 
-    let directory_service: Box<dyn DirectoryService> = match url.scheme() {
+    let directory_service_config = with_registry(&REG, || {
-        "memory" => {
+        <DeserializeWithRegistry<Box<dyn ServiceBuilder<Output = dyn DirectoryService>>>>::try_from(
-            // memory doesn't support host or path in the URL.
+            url,
-            if url.has_host() || !url.path().is_empty() {
-                return Err(Error::StorageError("invalid url".to_string()));
-            }
-            Box::<MemoryDirectoryService>::default()
-        }
-        "sled" => {
-            // sled doesn't support host, and a path can be provided (otherwise
-            // it'll live in memory only).
-            if url.has_host() {
-                return Err(Error::StorageError("no host allowed".to_string()));
-            }
-
-            if url.path() == "/" {
-                return Err(Error::StorageError(
-                    "cowardly refusing to open / with sled".to_string(),
-                ));
-            }
-
-            // TODO: expose compression and other parameters as URL parameters?
-
-            Box::new(if url.path().is_empty() {
-                SledDirectoryService::new_temporary()
-                    .map_err(|e| Error::StorageError(e.to_string()))?
-            } else {
-                SledDirectoryService::new(url.path())
-                    .map_err(|e| Error::StorageError(e.to_string()))?
-            })
-        }
-        scheme if scheme.starts_with("grpc+") => {
-            // schemes starting with grpc+ go to the GRPCPathInfoService.
-            //   That's normally grpc+unix for unix sockets, and grpc+http(s) for the HTTP counterparts.
-            // - In the case of unix sockets, there must be a path, but may not be a host.
-            // - In the case of non-unix sockets, there must be a host, but no path.
-            // Constructing the channel is handled by tvix_castore::channel::from_url.
-            Box::new(GRPCDirectoryService::from_client(
-                DirectoryServiceClient::with_interceptor(
-                    crate::tonic::channel_from_url(&url).await?,
-                    tvix_tracing::propagate::tonic::send_trace,
-                ),
-            ))
-        }
-        scheme if scheme.starts_with("objectstore+") => {
-            // We need to convert the URL to string, strip the prefix there, and then
-            // parse it back as url, as Url::set_scheme() rejects some of the transitions we want to do.
-            let trimmed_url = {
-                let s = url.to_string();
-                let mut url = Url::parse(s.strip_prefix("objectstore+").unwrap()).unwrap();
-                // trim the query pairs, they might contain credentials or local settings we don't want to send as-is.
-                url.set_query(None);
-                url
-            };
-            Box::new(
-                ObjectStoreDirectoryService::parse_url_opts(&trimmed_url, url.query_pairs())
-                    .map_err(|e| Error::StorageError(e.to_string()))?,
         )
-        }
+    })?
-        #[cfg(feature = "cloud")]
+    .0;
-        "bigtable" => {
+    let directory_service = directory_service_config
-            use super::bigtable::BigtableParameters;
+        .build("anonymous", &CompositionContext::blank())
-            use super::BigtableDirectoryService;
+        .await?;
 
-            // parse the instance name from the hostname.
-            let instance_name = url
-                .host_str()
-                .ok_or_else(|| Error::StorageError("instance name missing".into()))?
-                .to_string();
-
-            // … but add it to the query string now, so we just need to parse that.
-            url.query_pairs_mut()
-                .append_pair("instance_name", &instance_name);
-
-            let params: BigtableParameters = serde_qs::from_str(url.query().unwrap_or_default())
-                .map_err(|e| Error::InvalidRequest(format!("failed to parse parameters: {}", e)))?;
-
-            Box::new(
-                BigtableDirectoryService::connect(params)
-                    .await
-                    .map_err(|e| Error::StorageError(e.to_string()))?,
-            )
-        }
-        _ => {
-            return Err(crate::Error::StorageError(format!(
-                "unknown scheme: {}",
-                url.scheme()
-            )))
-        }
-    };
     Ok(directory_service)
 }
 

tvix/castore/src/directoryservice/grpc.rs

@@ -224,6 +224,19 @@ pub struct GRPCDirectoryServiceConfig {
     url: String,
 }
 
+impl TryFrom<url::Url> for GRPCDirectoryServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        //   This is normally grpc+unix for unix sockets, and grpc+http(s) for the HTTP counterparts.
+        // - In the case of unix sockets, there must be a path, but may not be a host.
+        // - In the case of non-unix sockets, there must be a host, but no path.
+        // Constructing the channel is handled by tvix_castore::channel::from_url.
+        Ok(GRPCDirectoryServiceConfig {
+            url: url.to_string(),
+        })
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for GRPCDirectoryServiceConfig {
     type Output = dyn DirectoryService;

tvix/castore/src/directoryservice/memory.rs

@@ -91,6 +91,17 @@ impl DirectoryService for MemoryDirectoryService {
 #[serde(deny_unknown_fields)]
 pub struct MemoryDirectoryServiceConfig {}
 
+impl TryFrom<url::Url> for MemoryDirectoryServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        // memory doesn't support host or path in the URL.
+        if url.has_host() || !url.path().is_empty() {
+            return Err(Error::StorageError("invalid url".to_string()).into());
+        }
+        Ok(MemoryDirectoryServiceConfig {})
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for MemoryDirectoryServiceConfig {
     type Output = dyn DirectoryService;

tvix/castore/src/directoryservice/object_store.rs

@@ -179,6 +179,32 @@ pub struct ObjectStoreDirectoryServiceConfig {
     object_store_options: HashMap<String, String>,
 }
 
+impl TryFrom<url::Url> for ObjectStoreDirectoryServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        // We need to convert the URL to string, strip the prefix there, and then
+        // parse it back as url, as Url::set_scheme() rejects some of the transitions we want to do.
+        let trimmed_url = {
+            let s = url.to_string();
+            let mut url = Url::parse(
+                s.strip_prefix("objectstore+")
+                    .ok_or(Error::StorageError("Missing objectstore uri".into()))?,
+            )?;
+            // trim the query pairs, they might contain credentials or local settings we don't want to send as-is.
+            url.set_query(None);
+            url
+        };
+        Ok(ObjectStoreDirectoryServiceConfig {
+            object_store_url: trimmed_url.into(),
+            object_store_options: url
+                .query_pairs()
+                .into_iter()
+                .map(|(k, v)| (k.to_string(), v.to_string()))
+                .collect(),
+        })
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for ObjectStoreDirectoryServiceConfig {
     type Output = dyn DirectoryService;

tvix/castore/src/directoryservice/sled.rs

@@ -145,6 +145,31 @@ pub struct SledDirectoryServiceConfig {
     path: Option<String>,
 }
 
+impl TryFrom<url::Url> for SledDirectoryServiceConfig {
+    type Error = Box<dyn std::error::Error + Send + Sync>;
+    fn try_from(url: url::Url) -> Result<Self, Self::Error> {
+        // sled doesn't support host, and a path can be provided (otherwise
+        // it'll live in memory only).
+        if url.has_host() {
+            return Err(Error::StorageError("no host allowed".to_string()).into());
+        }
+
+        // TODO: expose compression and other parameters as URL parameters?
+
+        Ok(if url.path().is_empty() {
+            SledDirectoryServiceConfig {
+                is_temporary: true,
+                path: None,
+            }
+        } else {
+            SledDirectoryServiceConfig {
+                is_temporary: false,
+                path: Some(url.path().to_string()),
+            }
+        })
+    }
+}
+
 #[async_trait]
 impl ServiceBuilder for SledDirectoryServiceConfig {
     type Output = dyn DirectoryService;

tvix/castore/src/tests/import.rs

@@ -4,7 +4,6 @@ use crate::fixtures::*;
 use crate::import::fs::ingest_path;
 use crate::proto;
 
-use std::sync::Arc;
 use tempfile::TempDir;
 
 #[cfg(target_family = "unix")]
@@ -26,7 +25,7 @@ async fn symlink() {
     .unwrap();
 
     let root_node = ingest_path(
-        Arc::from(blob_service),
+        blob_service,
         directory_service,
         tmpdir.path().join("doesntmatter"),
     )
@@ -44,8 +43,7 @@ async fn symlink() {
 
 #[tokio::test]
 async fn single_file() {
-    let blob_service =
+    let blob_service = blobservice::from_addr("memory://").await.unwrap();
-        Arc::from(blobservice::from_addr("memory://").await.unwrap()) as Arc<dyn BlobService>;
     let directory_service = directoryservice::from_addr("memory://").await.unwrap();
 
     let tmpdir = TempDir::new().unwrap();
@@ -77,8 +75,7 @@ async fn single_file() {
 #[cfg(target_family = "unix")]
 #[tokio::test]
 async fn complicated() {
-    let blob_service =
+    let blob_service = blobservice::from_addr("memory://").await.unwrap();
-        Arc::from(blobservice::from_addr("memory://").await.unwrap()) as Arc<dyn BlobService>;
     let directory_service = directoryservice::from_addr("memory://").await.unwrap();
 
     let tmpdir = TempDir::new().unwrap();

tvix/serde/src/de_tests.rs

@@ -236,9 +236,8 @@ mod test_builtins {
 fn deserialize_with_extra_builtin() {
     let code = "builtins.prependHello \"world\"";
 
-    let result: String = from_str_with_config(code, |eval| {
+    let result: String =
-        eval.add_builtins(test_builtins::builtins())
+        from_str_with_config(code, |eval| eval.add_builtins(test_builtins::builtins()))
-    })
             .expect("should deserialize");
 
     assert_eq!(result, "hello world");

tvix/store/src/bin/tvix-store.rs

@@ -200,7 +200,7 @@ fn default_threads() -> usize {
 }
 
 #[instrument(skip_all)]
-async fn run_cli(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
+async fn run_cli(cli: Cli) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     match cli.command {
         Commands::Daemon {
             listen_args,
@@ -538,7 +538,7 @@ async fn run_cli(cli: Cli) -> Result<(), Box<dyn std::error::Error>> {
 }
 
 #[tokio::main]
-async fn main() -> Result<(), Box<dyn std::error::Error>> {
+async fn main() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     let cli = Cli::parse();
 
     let tracing_handle = {

tvix/store/src/utils.rs

@@ -19,19 +19,19 @@ pub async fn construct_services(
     blob_service_addr: impl AsRef<str>,
     directory_service_addr: impl AsRef<str>,
     path_info_service_addr: impl AsRef<str>,
-) -> std::io::Result<(
+) -> Result<
+    (
         Arc<dyn BlobService>,
         Arc<dyn DirectoryService>,
         Box<dyn PathInfoService>,
         Box<dyn NarCalculationService>,
-)> {
+    ),
-    let blob_service: Arc<dyn BlobService> = blobservice::from_addr(blob_service_addr.as_ref())
+    Box<dyn std::error::Error + Send + Sync>,
-        .await?
+> {
-        .into();
+    let blob_service: Arc<dyn BlobService> =
+        blobservice::from_addr(blob_service_addr.as_ref()).await?;
     let directory_service: Arc<dyn DirectoryService> =
-        directoryservice::from_addr(directory_service_addr.as_ref())
+        directoryservice::from_addr(directory_service_addr.as_ref()).await?;
-            .await?
-            .into();
 
     let path_info_service = pathinfoservice::from_addr(
         path_info_service_addr.as_ref(),