RUST-870 Deserialize server response directly from raw BSON bytes #389

patrickfreed · 2021-07-09T17:25:41Z

This PR updates the driver to leverage the new bson::from_slice serde functionality in the BSON library to deserialize all server response directly from BSON rather than going through Document first, resulting in significant performance improvements.

patrickfreed · 2021-07-09T17:30:17Z

src/client/executor.rs

@@ -353,22 +360,72 @@ impl Client {
        let start_time = Instant::now();
        let cmd_name = cmd.name.clone();

-        let response_result = match connection.send_command(cmd, request_id).await {
+        let command_result = match connection.send_command(cmd, request_id).await {


This match constitutes the main change in the PR. Since we want to avoid multiple deserialization steps, we need to deserialize the command response and the body of the response all in one step. Previously we would deserialize a Document, do some processing on it, and then pass that into handle_response for further deserialization of the body. Since we need to do it in one step, we need to include the "metadata" (e.g. ok, $clusterTime) information along with the "body" in the same type. To achieve this, each operation needed to specify their "body" type via an additional associated type (Operation::ResponseType).

To allow for RunCommand to fit into this scheme (which wants to preserve all the "metadata" and not separate it from the "body"), we needed a trait that the ResponseTypes implemented which could yield the ok and clusterTime values. This trait can easily be extended to extract any "metadata" we need in the future too.

patrickfreed · 2021-07-09T17:33:24Z

src/client/executor.rs

+                                        }))
+                                    }
+                                    // for ok: 1 just return the original deserialization error.
+                                    _ => Err(deserialize_error),


this case is if the user specifies a T that doesn't match the data, for example.

patrickfreed · 2021-07-09T17:33:48Z

src/client/executor.rs

+                                        if !error_response.is_success() =>
+                                    {
+                                        Err(command_error_response.into())
+                                    }


this case (and its equivalent in the other branch) really shouldn't be encountered except for when caused by bugs in the server / driver.

patrickfreed · 2021-07-09T17:34:20Z

src/client/executor.rs

+                        // if we failed to deserialize the whole response, try deserializing
+                        // a generic command response without the operation's body.
+                        match response.body::<CommandResponse<Option<CommandErrorBody>>>() {
+                            Ok(error_response) => {


most command errors will go through this path

patrickfreed · 2021-07-09T17:36:02Z

src/cmap/conn/command.rs

+
+/// A command response backed by a `Document` rather than raw bytes.
+/// Use this for simple command responses where deserialization performance is not a high priority.
+pub(crate) struct DocumentCommandResponse {


this is essentially equivalent to the old CommandResponse.

patrickfreed · 2021-07-09T17:41:32Z

src/coll/mod.rs

@@ -561,7 +561,7 @@ impl<T> Collection<T> {

 impl<T> Collection<T>
 where
-    T: DeserializeOwned + Unpin,
+    T: DeserializeOwned + Unpin + Send + Sync,


This is actually a breaking change, but it's required now that the Ts are stored on cursors and need to be held across .await points. Send and Sync are implemented on basically every type, so this shouldn't be a problem. See here for more info.

patrickfreed · 2021-07-09T17:51:53Z

src/is_master.rs

@@ -111,9 +121,7 @@ impl PartialEq for IsMasterCommandResponse {

 impl IsMasterCommandResponse {
    pub(crate) fn server_type(&self) -> ServerType {
-        if self.ok != Some(1.0) {


we were always validating the response before using this method anyways, so ok was always 1.0.

patrickfreed · 2021-07-09T17:53:20Z

src/operation/abort_transaction/mod.rs

@@ -1,12 +1,12 @@
 use crate::{


most of the operation changes are updating handle_response to work on pre-deserialized response rather than a Document-based one and updating the tests to work with that. It was a lot of refactoring, so I've updated those tests to use a shared handle_response_test function to allow for easier refactorings in the future, if necessary.

patrickfreed · 2021-07-09T17:55:00Z

src/operation/mod.rs

+/// A response to a command with a body shaped deserialized to a `T`.
+#[derive(Deserialize, Debug)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct CommandResponse<T> {


This is the type that all operations besides RunCommand use for their response type (parameterized by the generic body T).

patrickfreed · 2021-07-09T17:58:04Z

src/operation/mod.rs

@@ -229,28 +323,3 @@ pub(crate) enum Retryability {
    Read,
    None,
 }
-
-#[cfg(test)]
-mod test {


moved to operation/test.rs

abr-egn · 2021-07-12T14:10:56Z

Cargo.toml

@@ -38,7 +38,7 @@ bson-uuid-0_8 = ["bson/uuid-0_8"]
 async-trait = "0.1.42"
 base64 = "0.13.0"
 bitflags = "1.1.0"
-bson = "2.0.0-beta.2"
+bson = { git = "https://github.com/mongodb/bson-rust" }


We'll need to be sure to point this back at a published release of bson before the final 2.0.0 release.

Yep, filed RUST-901 to remind us to do that.

abr-egn · 2021-07-12T14:40:37Z

src/client/executor.rs

@@ -393,12 +450,16 @@ impl Client {
            }
            Ok(response) => {
                self.emit_command_event(|handler| {
+                    let response_doc: Document = response


Minor: this can just be the body of the else, it doesn't need its own binding.

abr-egn · 2021-07-12T14:45:39Z

src/client/executor.rs

@@ -393,12 +450,16 @@ impl Client {
            }
            Ok(response) => {
                self.emit_command_event(|handler| {
+                    let response_doc: Document = response
+                        .raw
+                        .body()


If I understand correctly, command events need the body as a Document, so if they're enabled that deserialization is needed despite being able to go directly from the response bytes to the response struct. If that's true, it might be worth calling out in documentation that command events come at a performance penalty.

Yep that's right, and that's a good idea, added a line about this to the command_event_handler option in ClientOptions.

abr-egn · 2021-07-12T15:03:22Z

src/cursor/session.rs

@@ -55,15 +55,15 @@ where
    exhausted: bool,
    client: Client,
    info: CursorInformation,
-    buffer: VecDeque<Document>,
+    buffer: VecDeque<T>,
    _phantom: std::marker::PhantomData<T>,


Is this still needed now that buffer uses T?

nope, good catch, removed.

patrickfreed · 2021-07-14T19:23:39Z

Re-requesting review (sorry!) for the snapshot time / atClusterTime changes needed after rebasing off of master, should only need to take a look at the last two commits.

abr-egn

LGTM - thanks for handling the merge!

patrickfreed commented Jul 9, 2021

View reviewed changes

patrickfreed marked this pull request as ready for review July 9, 2021 19:18

patrickfreed requested review from isabelatkinson and abr-egn July 9, 2021 19:18

patrickfreed mentioned this pull request Jul 12, 2021

RUST-885 Support snapshot sessions #390

Merged

abr-egn reviewed Jul 12, 2021

View reviewed changes

abr-egn approved these changes Jul 12, 2021

View reviewed changes

isabelatkinson approved these changes Jul 13, 2021

View reviewed changes

patrickfreed added 2 commits July 14, 2021 13:18

deserialize responses from raw bson

c52f63d

handle atClusterTime

050a55a

patrickfreed force-pushed the RUST-870/raw-deserialize branch from a753821 to 050a55a Compare July 14, 2021 18:42

fix broken agg test

bf7a507

patrickfreed requested review from isabelatkinson and abr-egn July 14, 2021 19:22

abr-egn approved these changes Jul 14, 2021

View reviewed changes

isabelatkinson approved these changes Jul 15, 2021

View reviewed changes

patrickfreed merged commit 0fa2a90 into mongodb:master Jul 15, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

RUST-870 Deserialize server response directly from raw BSON bytes #389

RUST-870 Deserialize server response directly from raw BSON bytes #389

patrickfreed commented Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

patrickfreed Jul 9, 2021

abr-egn Jul 12, 2021

patrickfreed Jul 12, 2021

abr-egn Jul 12, 2021

patrickfreed Jul 12, 2021

abr-egn Jul 12, 2021

patrickfreed Jul 12, 2021

abr-egn Jul 12, 2021

patrickfreed Jul 12, 2021

patrickfreed commented Jul 14, 2021

abr-egn left a comment

RUST-870 Deserialize server response directly from raw BSON bytes #389

RUST-870 Deserialize server response directly from raw BSON bytes #389

Conversation

patrickfreed commented Jul 9, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

patrickfreed commented Jul 14, 2021

abr-egn left a comment

Choose a reason for hiding this comment