Details
Description
I think I have encountered one of the memory leaks described by AVRO-1283 in the way Java Avro implements field accessor caching in ReflectData. When a reflected object is serialized, the key of ClassAccessorData.bySchema (as retained by ReflectData.ACCESSOR_CACHE) retains a strong reference to the schema that was used to serialize the object, but there exists no code path for clearing these references after a schema will no longer be used.
While in most cases, a class will probably only have one schema associated with it (created and cached by ReflectData.getSchema(Type)), I experienced OutOfMemoryError when serializing generic classes with dynamically-generated schemas. The following is a minimal example which will exhaust a 50MiB heap (-Xmx50m) after about 190K iterations:
import java.io.ByteArrayOutputStream; import java.io.IOException; import java.util.Collections; import org.apache.avro.Schema; import org.apache.avro.io.BinaryEncoder; import org.apache.avro.io.EncoderFactory; import org.apache.avro.reflect.ReflectDatumWriter; public class AvroMemoryLeakMinimal { public static void main(String[] args) throws IOException { long count = 0; EncoderFactory encFactory = EncoderFactory.get(); try { while (true) { // Create schema Schema schema = Schema.createRecord("schema", null, null, false); schema.setFields(Collections.<Schema.Field>emptyList()); // serialize ByteArrayOutputStream baos = new ByteArrayOutputStream(1024); BinaryEncoder encoder = encFactory.binaryEncoder(baos, null); (new ReflectDatumWriter<Object>(schema)).write(new Object(), encoder); byte[] result = baos.toByteArray(); count++; } } catch (OutOfMemoryError e) { System.out.print("Memory exhausted after "); System.out.print(count); System.out.println(" schemas"); throw e; } } }
I was able to fix the bug in the latest 1.9.0-SNAPSHOT from git with the following patch to ClassAccessorData.bySchema to use weak keys so that it properly released the Schema objects if no other threads are still referencing them:
--- a/lang/java/avro/src/main/java/org/apache/avro/reflect/ReflectData.java +++ b/lang/java/avro/src/main/java/org/apache/avro/reflect/ReflectData.java @@ -57,6 +57,7 @@ import org.apache.avro.io.DatumWriter; import org.apache.avro.specific.FixedSize; import org.apache.avro.specific.SpecificData; import org.apache.avro.SchemaNormalization; +import org.apache.avro.util.WeakIdentityHashMap; import org.codehaus.jackson.JsonNode; import org.codehaus.jackson.node.NullNode; @@ -234,8 +235,8 @@ public class ReflectData extends SpecificData { private final Class<?> clazz; private final Map<String, FieldAccessor> byName = new HashMap<String, FieldAccessor>(); - private final IdentityHashMap<Schema, FieldAccessor[]> bySchema = - new IdentityHashMap<Schema, FieldAccessor[]>(); + private final WeakIdentityHashMap<Schema, FieldAccessor[]> bySchema = + new WeakIdentityHashMap<Schema, FieldAccessor[]>(); private ClassAccessorData(Class<?> c) { clazz = c;
Additionally, I'm not sure why an IdentityHashMap was used instead of a standard HashMap, since two equivalent schemas have the same set of FieldAccessor. Everything appears to work and all tests pass if I use a WeakHashMap instead of an WeakIdentityHashMap, but I don't know if there was some other reason object identity was important for this map. If a non-identity map can be used, this will help reduce memory/CPU usage further by not regenerating all the field accessors for equivalent schemas.
The following unit test appears to reliably catch this bug, but is non-deterministic due to the nature of garbage collection (and I'm not sure there's a way around that):
package org.apache.avro.reflect; import org.apache.avro.Schema; import org.junit.Test; import java.io.IOException; import java.lang.reflect.Field; import java.util.Collections; import java.util.Map; import static org.hamcrest.Matchers.lessThan; import static org.junit.Assert.assertThat; public class TestReflectData { /** * Test if ReflectData is leaking {@link Schema} references */ @SuppressWarnings("unchecked") @Test public void testWeakSchemaCaching() throws IOException, NoSuchFieldException, IllegalAccessException { for (int i = 0; i < 1000; i++) { // Create schema Schema schema = Schema.createRecord("schema", null, null, false); schema.setFields(Collections.<Schema.Field>emptyList()); ReflectData.get().getRecordState(new Object(), schema); } // Reflect the number of schemas currently in the cache Field cacheField = ReflectData.class.getDeclaredField("ACCESSOR_CACHE"); cacheField.setAccessible(true); Map<Class<?>, ?> ACCESSOR_CACHE = (Map) cacheField.get(null); Object classData = ACCESSOR_CACHE.get(Object.class); Field bySchemaField = classData.getClass().getDeclaredField("bySchema"); bySchemaField.setAccessible(true); Map<Schema, FieldAccessor[]> accessors = (Map) bySchemaField.get(classData); System.gc(); // Not guaranteed reliable, but seems to be reliable enough for our purposes // See if the number of schemas in the cache is less than the number we generated - if so, then they are being released. assertThat("ReflectData cache should release references", accessors.size(), lessThan(1000)); } }
(Added org.hamcrest:hamcrest-all dependency to test scope for the built-in lessThan() matcher)
The current workaround that I'm using to mitigate the leak is to cache schemas and re-use older instances when I'm about to serialize an equivalent schema. Since most of the generated schemas are equivalent, this limits the number of leaked schemas to a handful. A more permanent workaround would be to switch to using a GenericRecord instead of a generic java class for the object that is being serialized, since this cuts out the use of ReflectData entirely.