Skip to main content

Module rfc_3197_config

Module rfc_3197_config 

Source
Expand description

Config

§Summary

Expose services config to the user.

§Motivation

OpenDAL provides two ways to configure services: through a builder pattern and via a map.

The Builder allows users to configure services using the builder pattern:

// Create fs backend builder.
let mut builder = Fs::default();
// Set the root for fs, all operations will happen under this root.
builder.root("/tmp");

// Build an `Operator` to start operating the storage.
let op: Operator = Operator::new(builder)?.finish();

The benefit of builder is that it is type safe and easy to use. However, it is not flexible enough to configure services. Users must create a new builder for each service they wish to configure, translating user input into the API calls for each respective builder.

Consider the following real-world example from one of our users:

let mut builder = services::S3::default();

// Credential.
builder.access_key_id(&cfg.access_key_id);
builder.secret_access_key(&cfg.secret_access_key);
builder.session_token(&cfg.session_token);
builder.role_arn(&cfg.role_arn);
builder.external_id(&cfg.external_id);

// Root.
builder.root(&cfg.root);

// Disable credential loader
if cfg.disable_credential_loader {
    builder.disable_config_load();
    builder.disable_ec2_metadata();
}

// Enable virtual host style
if cfg.enable_virtual_host_style {
    builder.enable_virtual_host_style();
}

The Map approach allows users to configure services using a string-based HashMap:

let map = HashMap::from([
  // Set the root for fs, all operations will happen under this root.
  ("root".to_string(), "/tmp".to_string()),
]);

// Build an `Operator` to start operating the storage.
let op: Operator = Operator::via_map(Scheme::Fs, map)?;

This approach is simpler since it allows users to configure all services within a single map. However, it is not type safe and not easy to use. Users will need to convert their input to string and make sure the key is correct. And breaking changes could happen silently.

This is one of our limitations: We need a way to configure services that is type safe, easy to use and flexible. The other one is that there is no way for users to fetch the config of a service after it’s built. This limitation complicates the dynamic modification of a service’s root path for the user.

Our users have to wrap all our configs into an enum and store it in their own struct:

pub enum StorageParams {
    Azblob(StorageAzblobConfig),
    Fs(StorageFsConfig),
    Ftp(StorageFtpConfig),
    Gcs(StorageGcsConfig),
    Hdfs(StorageHdfsConfig),
    Http(StorageHttpConfig),
    Ipfs(StorageIpfsConfig),
    Memory,
    Moka(StorageMokaConfig),
    Obs(StorageObsConfig),
    Oss(StorageOssConfig),
    S3(StorageS3Config),
    Redis(StorageRedisConfig),
    Webhdfs(StorageWebhdfsConfig),
    Cos(StorageCosConfig),
}

So I propose to expose services config to the users, allowing them to work on config structs directly and fetch the config at runtime.

§Guide-level explanation

First of all, we will add config struct for each service. For example, Fs will have a FsConfig struct and S3 will have a S3Config. The fields within the config struct are public and marked as non-exhaustive.

#[non_exhaustive]
pub struct S3Config {
  pub root: Option<String>,
  pub bucket: String,
  pub endpoint: Option<String>,
  pub region: Option<String>,
  ...
}

Then, we will add a Config enum that contains all the config structs. The enum is public and non-exhaustive too.

#[non_exhaustive]
pub enum Config {
  Fs(FsConfig)
  S3(S3Config),
  Custom(&'static str, HashMap<String, String>),
}

Notably, a Custom variant will be added to the enum. This variant aligns with Scheme::Custom(name) and allows users to configure custom services.

At Operator level, we will add from_config and via_config methods.

impl Operator {
  pub fn via_config(cfg: impl Into<Config>) -> Result<Operator> {}
}

Additionally, OperatorInfo will introduce a new API method, config():

impl OperatorInfo {
  pub fn config(&self) -> Config {}
}

Users can use config() to fetch the config of a service at runtime and construct a new operator based on needs.

§Reference-level explanation

Every services will have a XxxConfig struct.

XxxConfig will implement the following things:

  • Default trait: All config fields will have a default value.
  • Deserialize trait: Allow users to deserialize a config from a string.
  • Into<Config> trait: All service config can be converted to Config enum.

Internally, XxxConfig will have the following traits:

  • FromMap trait: Allow users to build a config via hashmap which will replace existing from_map API in Builder.
  • Into<XxxBuilder> trait: Config can convert into corresponding builder with zero cost.

All config fields will be public and non-exhaustive, allowing users to build config this way:

let s3 = S3Config {
  bucket: "test".to_string(),
  endpoint: Some("http://localhost:9000".to_string()),
  ..Default::default()
}

The public API of existing builders will remain unchanged, although their internal implementations will be modified to utilize XxxConfig. Type that can’t be represents as String like Box<dyn AwsCredentialLoad> and HttpClient will be kept in Builder as before.

For example:

#[non_exhaustive]
pub struct S3Config {
  pub root: Option<String>,
  pub bucket: String,
  pub endpoint: Option<String>,
  pub region: Option<String>,
  ...
}

pub struct S3Builder {
    config: S3Config,
    
    customized_credential_load: Option<Box<dyn AwsCredentialLoad>>,
    http_client: Option<HttpClient>,
}

§Drawbacks

§API Surface

This modification will significantly expand OpenDAL’s public API surface, makes it harder to maintain and increases the risk of breaking changes. Also, this change will add much more work for bindings which need to implement XxxConfig for each service.

§Secrets Leakage

After our config supports Serialize, it’s possible that users will serialize the config and log it. This will lead to secrets leakage. We should add a warning in the docs to prevent this. And we should also encourage uses of services like AWS IAM instead of static secrets.

§Rationale and alternatives

§Move root out of service config to operator level

There is another way to solve the problem: Move root out of service config to operator level.

We can move root out of the service config and put it in Operator level. This way, users can configure root for all services in one place. However, this is a large breaking changes and users will need to maintain the root logic everywhere.

§Prior art

None.

§Unresolved questions

None.

§Future possibilities

§Implement FromStr for Config

We can implement FromStr for Config so that users can parse a config from a string.

let cfg = Config::from_str("s3://bucket/path/to/file?access_key_id=xxx&secret_access_key=xxx")?;

§Implement Serialize for Config

We can implement Serialize for Config so that users can serialize a config.

// Serialize
let bs = serde_json::to_vec(&cfg)?;
// Deserialize
let cfg: Config = serde_json::from_slice(&bs)?;

§Implement check for Config

Implement check for config so that users can check if a config is valid before build.