Prevent concurrent access to local breaker in rerank (elastic#128162) (elastic#128945)

When an async operator receives a response, we can't create new blocks
on the responding thread because multiple threads may adjust the local
breaker simultaneously, leading to a data race. To address this, we can
either use the global breaker or delay block creation in getOutput.
While using the global block factory is simpler, I prefer the second
option to use the local breaker when possible. Therefore, I opted to
keep the results in the queue and create new blocks in getOutput. Our
tests didn't catch this issue because: (1) only one block is created in
the test, and (2) there is no delay when simulating the inference
service.

Closes elastic#127638
Closes elastic#127051

Author: dnhatn

test/framework/src/main/java/org/elasticsearch/indices/CrankyCircuitBreakerService.java

@@ -29,7 +29,7 @@ public class CrankyCircuitBreakerService extends CircuitBreakerService {
      */
     public static final String ERROR_MESSAGE = "cranky breaker";
 
-    private final CircuitBreaker breaker = new CircuitBreaker() {
+    public static final class CrankyCircuitBreaker implements CircuitBreaker {
         private final AtomicLong used = new AtomicLong();
 
         @Override
@@ -82,7 +82,9 @@ public Durability getDurability() {
         public void setLimitAndOverhead(long limit, double overhead) {
 
         }
-    };
+    }
+
+    private final CrankyCircuitBreaker breaker = new CrankyCircuitBreaker();
 
     @Override
     public CircuitBreaker getBreaker(String name) {

x-pack/plugin/esql/compute/src/main/java/org/elasticsearch/compute/data/LocalCircuitBreaker.java

@@ -28,6 +28,7 @@ public final class LocalCircuitBreaker implements CircuitBreaker, Releasable {
     private final long maxOverReservedBytes;
     private long reservedBytes;
     private final AtomicBoolean closed = new AtomicBoolean(false);
+    private volatile Thread activeThread;
 
     public record SizeSettings(long overReservedBytes, long maxOverReservedBytes) {
         public SizeSettings(Settings settings) {
@@ -57,6 +58,7 @@ public void circuitBreak(String fieldName, long bytesNeeded) {
 
     @Override
     public void addEstimateBytesAndMaybeBreak(long bytes, String label) throws CircuitBreakingException {
+        assert assertSingleThread();
         if (bytes <= reservedBytes) {
             reservedBytes -= bytes;
             maybeReduceReservedBytes();
@@ -68,6 +70,7 @@ public void addEstimateBytesAndMaybeBreak(long bytes, String label) throws Circu
 
     @Override
     public void addWithoutBreaking(long bytes) {
+        assert assertSingleThread();
         if (bytes <= reservedBytes) {
             reservedBytes -= bytes;
             maybeReduceReservedBytes();
@@ -130,6 +133,7 @@ public void setLimitAndOverhead(long limit, double overhead) {
 
     @Override
     public void close() {
+        assert assertSingleThread();
         if (closed.compareAndSet(false, true)) {
             breaker.addWithoutBreaking(-reservedBytes);
         }
@@ -139,4 +143,34 @@ public void close() {
     public String toString() {
         return "LocalCircuitBreaker[" + reservedBytes + "/" + overReservedBytes + ":" + maxOverReservedBytes + "]";
     }
+
+    private boolean assertSingleThread() {
+        Thread activeThread = this.activeThread;
+        Thread currentThread = Thread.currentThread();
+        assert activeThread == null || activeThread == currentThread
+            : "Local breaker must be accessed by a single thread at a time: expected ["
+                + activeThread
+                + "] != actual ["
+                + currentThread
+                + "]";
+        return true;
+    }
+
+    /**
+     * Marks the beginning of a run loop for assertion purposes.
+     * Sets the current thread as the only thread allowed to access this breaker.
+     */
+    public boolean assertBeginRunLoop() {
+        activeThread = Thread.currentThread();
+        return true;
+    }
+
+    /**
+     * Marks the end of a run loop for assertion purposes.
+     * Clears the active thread to allow other threads to access this breaker.
+     */
+    public boolean assertEndRunLoop() {
+        activeThread = null;
+        return true;
+    }
 }

x-pack/plugin/esql/compute/src/main/java/org/elasticsearch/compute/operator/Driver.java

@@ -211,10 +211,13 @@ SubscribableListener<Void> run(TimeValue maxTime, int maxIterations, LongSupplie
         while (true) {
             IsBlockedResult isBlocked = Operator.NOT_BLOCKED;
             try {
+                assert driverContext.assertBeginRunLoop();
                 isBlocked = runSingleLoopIteration();
             } catch (DriverEarlyTerminationException unused) {
                 closeEarlyFinishedOperators();
                 assert isFinished() : "not finished after early termination";
+            } finally {
+                assert driverContext.assertEndRunLoop();
             }
             totalIterationsThisRun++;
             iterationsSinceLastStatusUpdate++;

x-pack/plugin/esql/compute/src/main/java/org/elasticsearch/compute/operator/DriverContext.java

@@ -10,9 +10,9 @@
 import org.elasticsearch.action.ActionListener;
 import org.elasticsearch.action.support.SubscribableListener;
 import org.elasticsearch.common.breaker.CircuitBreaker;
-import org.elasticsearch.common.breaker.NoopCircuitBreaker;
 import org.elasticsearch.common.util.BigArrays;
 import org.elasticsearch.compute.data.BlockFactory;
+import org.elasticsearch.compute.data.LocalCircuitBreaker;
 import org.elasticsearch.core.Releasable;
 import org.elasticsearch.core.Releasables;
 
@@ -74,14 +74,6 @@ private DriverContext(BigArrays bigArrays, BlockFactory blockFactory, WarningsMo
         this.warningsMode = warningsMode;
     }
 
-    public static DriverContext getLocalDriver() {
-        return new DriverContext(
-            BigArrays.NON_RECYCLING_INSTANCE,
-            // TODO maybe this should have a small fixed limit?
-            new BlockFactory(new NoopCircuitBreaker(CircuitBreaker.REQUEST), BigArrays.NON_RECYCLING_INSTANCE)
-        );
-    }
-
     public BigArrays bigArrays() {
         return bigArrays;
     }
@@ -208,6 +200,26 @@ public enum WarningsMode {
         IGNORE
     }
 
+    /**
+     * Marks the beginning of a run loop for assertion purposes.
+     */
+    public boolean assertBeginRunLoop() {
+        if (blockFactory.breaker() instanceof LocalCircuitBreaker localBreaker) {
+            assert localBreaker.assertBeginRunLoop();
+        }
+        return true;
+    }
+
+    /**
+     * Marks the end of a run loop for assertion purposes.
+     */
+    public boolean assertEndRunLoop() {
+        if (blockFactory.breaker() instanceof LocalCircuitBreaker localBreaker) {
+            assert localBreaker.assertEndRunLoop();
+        }
+        return true;
+    }
+
     private static class AsyncActions {
         private final SubscribableListener<Void> completion = new SubscribableListener<>();
         private final AtomicBoolean finished = new AtomicBoolean();

x-pack/plugin/esql/compute/test/src/main/java/org/elasticsearch/compute/test/OperatorTestCase.java

@@ -23,6 +23,7 @@
 import org.elasticsearch.compute.data.Block;
 import org.elasticsearch.compute.data.BlockFactory;
 import org.elasticsearch.compute.data.Page;
+import org.elasticsearch.compute.operator.AsyncOperator;
 import org.elasticsearch.compute.operator.Driver;
 import org.elasticsearch.compute.operator.DriverContext;
 import org.elasticsearch.compute.operator.DriverRunner;
@@ -249,9 +250,20 @@ public void testSimpleFinishClose() {
         try (var operator = simple().get(driverContext)) {
             assert operator.needsInput();
             for (Page page : input) {
-                operator.addInput(page);
+                if (operator.needsInput()) {
+                    operator.addInput(page);
+                } else {
+                    page.releaseBlocks();
+                }
             }
             operator.finish();
+            // for async operator, we need to wait for async actions to finish.
+            if (operator instanceof AsyncOperator<?> || randomBoolean()) {
+                driverContext.finish();
+                PlainActionFuture<Void> waitForAsync = new PlainActionFuture<>();
+                driverContext.waitForAsyncActions(waitForAsync);
+                waitForAsync.actionGet(TimeValue.timeValueSeconds(30));
+            }
         }
     }