RUST-1842 Update prose tests for mongos deprioritization during retryable ops

src/test/spec/retryable_reads.rs

@@ -1,4 +1,4 @@
-use std::{future::IntoFuture, time::Duration};
+use std::{future::IntoFuture, sync::Arc, time::Duration};
 
 use crate::bson::doc;
 
@@ -8,6 +8,7 @@ use crate::{
         cmap::{CmapEvent, ConnectionCheckoutFailedReason},
         command::CommandEvent,
     },
+    options::SelectionCriteria,
     runtime::{self, AsyncJoinHandle},
     test::{
         block_connection_supported,
@@ -174,23 +175,40 @@ async fn retry_read_different_mongos() {
     client_options.hosts.drain(2..);
     client_options.retry_reads = Some(true);
 
-    let mut guards = vec![];
-    for ix in [0, 1] {
-        let mut opts = client_options.clone();
-        opts.hosts.remove(ix);
-        opts.direct_connection = Some(true);
-        let client = Client::for_test().options(opts).await;
+    let hosts = client_options.hosts.clone();
+    let client = Client::for_test()
+        .options(client_options)
+        .monitor_events()
+        .await;
 
+    // NOTE: This test uses a single client to set failpoints on each mongos and run the find
+    // operation. This avoids flakiness caused by a race between server discovery and server
+    // selection.
+
+    // When a client is first created, it initializes its view of the topology with all configured
+    // mongos addresses, but marks each as Unknown until it completes the server discovery process
+    // by sending and receiving "hello" messages Unknown servers are not eligible for server
+    // selection.
+
+    // Previously, we created a new client for each call to `enable_fail_point` and for the find
+    // operation. Each new client restarted the discovery process, and sometimes had not yet marked
+    // both mongos servers as usable, leading to test failures when the retry logic couldn't find a
+    // second eligible server.
+
+    // By reusing a single client, each `enable_fail_point` call forces discovery to complete for
+    // the corresponding mongos. As a result, when the find operation runs, the client has a
+    // fully discovered topology and can reliably select between both servers.
+    let mut guards = Vec::new();
+    for address in hosts {
+        let address = address.clone();
         let fail_point = FailPoint::fail_command(&["find"], FailPointMode::Times(1))
             .error_code(6)
-            .close_connection(true);
+            .selection_criteria(SelectionCriteria::Predicate(Arc::new(move |info| {
+                info.description.address == address
+            })));
         guards.push(client.enable_fail_point(fail_point).await.unwrap());
     }
 
-    let client = Client::for_test()
-        .options(client_options)
-        .monitor_events()
-        .await;
     let result = client
         .database("test")
         .collection::<crate::bson::Document>("retry_read_different_mongos")
@@ -211,6 +229,14 @@ async fn retry_read_different_mongos() {
         "unexpected events: {:#?}",
         events,
     );
+    let first_failed = events[1].as_command_failed().unwrap();
+    let first_address = &first_failed.connection.address;
+    let second_failed = events[3].as_command_failed().unwrap();
+    let second_address = &second_failed.connection.address;
+    assert_ne!(
+        first_address, second_address,
+        "Failed commands did not occur on two different mongos instances"
+    );
 
     drop(guards); // enforce lifetime
 }
@@ -235,12 +261,11 @@ async fn retry_read_same_mongos() {
         client_options.direct_connection = Some(true);
         let client = Client::for_test().options(client_options).await;
 
-        let fail_point = FailPoint::fail_command(&["find"], FailPointMode::Times(1))
-            .error_code(6)
-            .close_connection(true);
+        let fail_point = FailPoint::fail_command(&["find"], FailPointMode::Times(1)).error_code(6);
         client.enable_fail_point(fail_point).await.unwrap()
     };
 
+    client_options.direct_connection = Some(false);
     let client = Client::for_test()
         .options(client_options)
         .monitor_events()
@@ -265,6 +290,14 @@ async fn retry_read_same_mongos() {
         "unexpected events: {:#?}",
         events,
     );
+    let first_failed = events[1].as_command_failed().unwrap();
+    let first_address = &first_failed.connection.address;
+    let second_failed = events[3].as_command_succeeded().unwrap();
+    let second_address = &second_failed.connection.address;
+    assert_eq!(
+        first_address, second_address,
+        "Failed command and retry did not occur on the same mongos instance",
+    );
 
     drop(fp_guard); // enforce lifetime
 }

src/test/spec/retryable_writes.rs

@@ -1,6 +1,6 @@
 use std::{sync::Arc, time::Duration};
 
-use crate::bson::Bson;
+use crate::{bson::Bson, options::SelectionCriteria};
 use tokio::sync::Mutex;
 
 use crate::{
@@ -317,27 +317,44 @@ async fn retry_write_different_mongos() {
         );
         return;
     }
+
+    // NOTE: This test uses a single client to set failpoints on each mongos and run the insert
+    // operation. This avoids flakiness caused by a race between server discovery and server
+    // selection.
+
+    // When a client is first created, it initializes its view of the topology with all configured
+    // mongos addresses, but marks each as Unknown until it completes the server discovery process
+    // by sending and receiving "hello" messages Unknown servers are not eligible for server
+    // selection.
+
+    // Previously, we created a new client for each call to `enable_fail_point` and for the insert
+    // operation. Each new client restarted the discovery process, and sometimes had not yet marked
+    // both mongos servers as usable, leading to test failures when the retry logic couldn't insert
+    // a second eligible server.
+
+    // By reusing a single client, each `enable_fail_point` call forces discovery to complete for
+    // the corresponding mongos. As a result, when the insert operation runs, the client has a
+    // fully discovered topology and can reliably select between both servers.
     client_options.hosts.drain(2..);
     client_options.retry_writes = Some(true);
+    let hosts = client_options.hosts.clone();
+    let client = Client::for_test()
+        .options(client_options)
+        .monitor_events()
+        .await;
 
-    let mut guards = vec![];
-    for ix in [0, 1] {
-        let mut opts = client_options.clone();
-        opts.hosts.remove(ix);
-        opts.direct_connection = Some(true);
-        let client = Client::for_test().options(opts).await;
-
+    let mut guards = Vec::new();
+    for address in hosts {
+        let address = address.clone();
         let fail_point = FailPoint::fail_command(&["insert"], FailPointMode::Times(1))
             .error_code(6)
-            .error_labels(vec![RETRYABLE_WRITE_ERROR])
-            .close_connection(true);
+            .error_labels([RETRYABLE_WRITE_ERROR])
+            .selection_criteria(SelectionCriteria::Predicate(Arc::new(move |info| {
+                info.description.address == address
+            })));
         guards.push(client.enable_fail_point(fail_point).await.unwrap());
     }
 
-    let client = Client::for_test()
-        .options(client_options)
-        .monitor_events()
-        .await;
     let result = client
         .database("test")
         .collection::<crate::bson::Document>("retry_write_different_mongos")
@@ -358,6 +375,14 @@ async fn retry_write_different_mongos() {
         "unexpected events: {:#?}",
         events,
     );
+    let first_failed = events[1].as_command_failed().unwrap();
+    let first_address = &first_failed.connection.address;
+    let second_failed = events[3].as_command_failed().unwrap();
+    let second_address = &second_failed.connection.address;
+    assert_ne!(
+        first_address, second_address,
+        "Failed commands did not occur on two different mongos instances"
+    );
 
     drop(guards); // enforce lifetime
 }
@@ -384,11 +409,11 @@ async fn retry_write_same_mongos() {
 
         let fail_point = FailPoint::fail_command(&["insert"], FailPointMode::Times(1))
             .error_code(6)
-            .error_labels(vec![RETRYABLE_WRITE_ERROR])
-            .close_connection(true);
+            .error_labels(vec![RETRYABLE_WRITE_ERROR]);
         client.enable_fail_point(fail_point).await.unwrap()
     };
 
+    client_options.direct_connection = Some(false);
     let client = Client::for_test()
         .options(client_options)
         .monitor_events()
@@ -413,6 +438,14 @@ async fn retry_write_same_mongos() {
         "unexpected events: {:#?}",
         events,
     );
+    let first_failed = events[1].as_command_failed().unwrap();
+    let first_address = &first_failed.connection.address;
+    let second_failed = events[3].as_command_succeeded().unwrap();
+    let second_address = &second_failed.connection.address;
+    assert_eq!(
+        first_address, second_address,
+        "Failed commands did not occur on the same mongos instance",
+    );
 
     drop(fp_guard); // enforce lifetime
 }

src/test/util/event.rs

@@ -9,7 +9,7 @@ use crate::{
     bson::doc,
     event::{
         cmap::CmapEvent,
-        command::{CommandEvent, CommandSucceededEvent},
+        command::{CommandEvent, CommandFailedEvent, CommandSucceededEvent},
         sdam::SdamEvent,
     },
     test::get_client_options,
@@ -101,6 +101,13 @@ impl CommandEvent {
             _ => None,
         }
     }
+
+    pub(crate) fn as_command_failed(&self) -> Option<&CommandFailedEvent> {
+        match self {
+            CommandEvent::Failed(e) => Some(e),
+            _ => None,
+        }
+    }
 }
 
 #[derive(Clone, Debug)]