KAFKA-9629 Use generated protocol for Fetch API #9008
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,93 @@ | ||
| /* | ||
| * Licensed to the Apache Software Foundation (ASF) under one or more | ||
| * contributor license agreements. See the NOTICE file distributed with | ||
| * this work for additional information regarding copyright ownership. | ||
| * The ASF licenses this file to You under the Apache License, Version 2.0 | ||
| * (the "License"); you may not use this file except in compliance with | ||
| * the License. You may obtain a copy of the License at | ||
| * | ||
| * http://www.apache.org/licenses/LICENSE-2.0 | ||
| * | ||
| * Unless required by applicable law or agreed to in writing, software | ||
| * distributed under the License is distributed on an "AS IS" BASIS, | ||
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| * See the License for the specific language governing permissions and | ||
| * limitations under the License. | ||
| */ | ||
|
|
||
| package org.apache.kafka.common.protocol; | ||
|
|
||
| import org.apache.kafka.common.record.BaseRecords; | ||
| import org.apache.kafka.common.record.MemoryRecords; | ||
| import org.apache.kafka.common.utils.ByteUtils; | ||
|
|
||
| import java.nio.ByteBuffer; | ||
|
|
||
| /** | ||
| * Implementation of Readable which reads from a byte buffer and can read records as {@link MemoryRecords} | ||
| * | ||
| * @see org.apache.kafka.common.requests.FetchResponse | ||
| */ | ||
| public class RecordsReader implements Readable { | ||
|
hachikuji marked this conversation as resolved.
Outdated
|
||
| private final ByteBuffer buf; | ||
|
|
||
| public RecordsReader(ByteBuffer buf) { | ||
| this.buf = buf; | ||
| } | ||
|
|
||
| @Override | ||
| public byte readByte() { | ||
| return buf.get(); | ||
| } | ||
|
|
||
| @Override | ||
| public short readShort() { | ||
| return buf.getShort(); | ||
| } | ||
|
|
||
| @Override | ||
| public int readInt() { | ||
| return buf.getInt(); | ||
| } | ||
|
|
||
| @Override | ||
| public long readLong() { | ||
| return buf.getLong(); | ||
| } | ||
|
|
||
| @Override | ||
| public double readDouble() { | ||
| return ByteUtils.readDouble(buf); | ||
| } | ||
|
|
||
| @Override | ||
| public void readArray(byte[] arr) { | ||
| buf.get(arr); | ||
| } | ||
|
|
||
| @Override | ||
| public int readUnsignedVarint() { | ||
| return ByteUtils.readUnsignedVarint(buf); | ||
| } | ||
|
|
||
| @Override | ||
| public ByteBuffer readByteBuffer(int length) { | ||
|
Contributor
There was a problem hiding this comment. More of a side question, but is this length guaranteed to be less than the buffer size? Wondering if it is worth adding range checking.
Member
Author
There was a problem hiding this comment. This is copied straight from ByteBufferAccessor and will probably go away in a follow-on PR. But either way, looking at it it seems it should always be in range since this is used by zero-copy byte fields in the message classes, e.g. int len = _reader.readInt();
if (len > 0) {
this.someZeroCopyField = _reader.readByteBuffer(len);
}So generally it's probably safe. In the case of a corrupt message where the length is wrong, ByteBuffer#limit will throw an error and parsing will fail. It probably would be nice to put a range check in ByteBufferAccessor so we can throw a more useful error. |
||
| ByteBuffer res = buf.slice(); | ||
| res.limit(length); | ||
|
|
||
| buf.position(buf.position() + length); | ||
|
|
||
| return res; | ||
| } | ||
|
|
||
| @Override | ||
| public BaseRecords readRecords(int length) { | ||
| if (length < 0) { | ||
| // no records | ||
| return null; | ||
| } else { | ||
| ByteBuffer recordsBuffer = readByteBuffer(length); | ||
| return MemoryRecords.readableRecords(recordsBuffer); | ||
| } | ||
| } | ||
| } | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,141 @@ | ||
| /* | ||
| * Licensed to the Apache Software Foundation (ASF) under one or more | ||
| * contributor license agreements. See the NOTICE file distributed with | ||
| * this work for additional information regarding copyright ownership. | ||
| * The ASF licenses this file to You under the Apache License, Version 2.0 | ||
| * (the "License"); you may not use this file except in compliance with | ||
| * the License. You may obtain a copy of the License at | ||
| * | ||
| * http://www.apache.org/licenses/LICENSE-2.0 | ||
| * | ||
| * Unless required by applicable law or agreed to in writing, software | ||
| * distributed under the License is distributed on an "AS IS" BASIS, | ||
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| * See the License for the specific language governing permissions and | ||
| * limitations under the License. | ||
| */ | ||
|
|
||
| package org.apache.kafka.common.protocol; | ||
|
|
||
| import org.apache.kafka.common.network.ByteBufferSend; | ||
| import org.apache.kafka.common.network.Send; | ||
| import org.apache.kafka.common.record.BaseRecords; | ||
| import org.apache.kafka.common.utils.ByteUtils; | ||
|
|
||
| import java.io.ByteArrayOutputStream; | ||
| import java.io.DataOutput; | ||
| import java.io.DataOutputStream; | ||
| import java.io.IOException; | ||
| import java.nio.ByteBuffer; | ||
| import java.util.function.Consumer; | ||
|
|
||
| /** | ||
| * Implementation of Writable which produces a sequence of {@link Send} objects. This allows for deferring the transfer | ||
| * of data from a record-set's file channel to the eventual socket channel. | ||
| * | ||
| * Excepting {@link #writeRecords(BaseRecords)}, calls to the write methods on this class will append to a byte array | ||
| * according to the format specified in {@link DataOutput}. When a call is made to writeRecords, any previously written | ||
| * bytes will be flushed as a new {@link ByteBufferSend} to the given Send consumer. After flushing the pending bytes, | ||
| * another Send is passed to the consumer which wraps the underlying record-set's transfer logic. | ||
| * | ||
| * For example, | ||
| * | ||
| * <pre> | ||
| * recordsWritable.writeInt(10); | ||
| * recordsWritable.writeRecords(records1); | ||
| * recordsWritable.writeInt(20); | ||
| * recordsWritable.writeRecords(records2); | ||
| * recordsWritable.writeInt(30); | ||
| * recordsWritable.flush(); | ||
| * </pre> | ||
| * | ||
| * Will pass 5 Send objects to the consumer given in the constructor. Care must be taken by callers to flush any | ||
| * pending bytes at the end of the writing sequence to ensure everything is flushed to the consumer. This class is | ||
| * intended to be used with {@link org.apache.kafka.common.record.MultiRecordsSend}. | ||
| * | ||
| * @see org.apache.kafka.common.requests.FetchResponse | ||
| */ | ||
| public class RecordsWriter implements Writable { | ||
|
mumrah marked this conversation as resolved.
Outdated
|
||
| private final String dest; | ||
| private final Consumer<Send> sendConsumer; | ||
| private final ByteArrayOutputStream byteArrayOutputStream; | ||
| private final DataOutput output; | ||
|
|
||
| public RecordsWriter(String dest, Consumer<Send> sendConsumer) { | ||
| this.dest = dest; | ||
| this.sendConsumer = sendConsumer; | ||
| this.byteArrayOutputStream = new ByteArrayOutputStream(); | ||
| this.output = new DataOutputStream(this.byteArrayOutputStream); | ||
| } | ||
|
|
||
| @Override | ||
| public void writeByte(byte val) { | ||
| writeQuietly(() -> output.writeByte(val)); | ||
|
mumrah marked this conversation as resolved.
Outdated
|
||
| } | ||
|
|
||
| @Override | ||
| public void writeShort(short val) { | ||
| writeQuietly(() -> output.writeShort(val)); | ||
| } | ||
|
|
||
| @Override | ||
| public void writeInt(int val) { | ||
| writeQuietly(() -> output.writeInt(val)); | ||
| } | ||
|
|
||
| @Override | ||
| public void writeLong(long val) { | ||
| writeQuietly(() -> output.writeLong(val)); | ||
|
|
||
| } | ||
|
|
||
| @Override | ||
| public void writeDouble(double val) { | ||
| writeQuietly(() -> ByteUtils.writeDouble(val, output)); | ||
|
|
||
| } | ||
|
|
||
| @Override | ||
| public void writeByteArray(byte[] arr) { | ||
| writeQuietly(() -> output.write(arr)); | ||
| } | ||
|
|
||
| @Override | ||
| public void writeUnsignedVarint(int i) { | ||
| writeQuietly(() -> ByteUtils.writeUnsignedVarint(i, output)); | ||
| } | ||
|
|
||
| @Override | ||
| public void writeByteBuffer(ByteBuffer src) { | ||
| writeQuietly(() -> output.write(src.array(), src.position(), src.remaining())); | ||
| } | ||
|
|
||
| @FunctionalInterface | ||
| private interface IOExceptionThrowingRunnable { | ||
| void run() throws IOException; | ||
| } | ||
|
|
||
| private void writeQuietly(IOExceptionThrowingRunnable runnable) { | ||
| try { | ||
| runnable.run(); | ||
| } catch (IOException e) { | ||
| throw new RuntimeException("Writable encountered an IO error", e); | ||
| } | ||
| } | ||
|
|
||
| @Override | ||
| public void writeRecords(BaseRecords records) { | ||
| flush(); | ||
| sendConsumer.accept(records.toSend(dest)); | ||
| } | ||
|
|
||
| /** | ||
| * Flush any pending bytes as a ByteBufferSend and reset the buffer | ||
| */ | ||
| public void flush() { | ||
| ByteBufferSend send = new ByteBufferSend(dest, | ||
| ByteBuffer.wrap(byteArrayOutputStream.toByteArray())); | ||
|
Member
There was a problem hiding this comment. This creates a copy of the underlying bytes, can we avoid it?
Member
Author
There was a problem hiding this comment. Yea, it's possible, but rather complicated I think. We would need to manage our own byte array and grow it on-demand (like what happens in ByteArrayOutputStream). Then we could use
Member
There was a problem hiding this comment. Would
Member
Author
There was a problem hiding this comment. Looks like the expansion factor for ByteArrayOutputStream varies on the JDK version. In JDK 8 and 11 it's 2x, but in JDK 14 it just grows the buffer to the minimum needed size. Our growth factor of 1.1 in ByteBufferOutputStream seems reasonable . Not to mention avoiding the final copy by using slice would be nice too.
Member
There was a problem hiding this comment. @mumrah Thanks for checking this. However, the behavior in JDK 14 has not changed in that way. Performance would be atrocious if it did: private void ensureCapacity(int minCapacity) {
// overflow-conscious code
int oldCapacity = buf.length;
int minGrowth = minCapacity - oldCapacity;
if (minGrowth > 0) {
buf = Arrays.copyOf(buf, ArraysSupport.newLength(oldCapacity,
minGrowth, oldCapacity /* preferred growth */));
}The third parameter passed to
Member
There was a problem hiding this comment. Looking a bit more, it seems like this will be mostly used by the data that precedes the actual records. Do we have a sense for what's the typical size for that? If we do, we can use that in the initial size and we can keep the
Member
Author
There was a problem hiding this comment. Thanks for the explanation, @ijuma. I missed the semantics of After the initial few top-level fields, each partition will have something like 38 bytes preceding its records (at a minimum, aborted transactions could increase that). Maybe we could increase initial capacity to 64 bytes?
Member
Author
There was a problem hiding this comment. I increased the initial buffer size to 64 and also added 2x growth factor for the buffer. It occurred to me the initial size only really helps for the first partition's header fields, but beyond that (since we are reusing/growing the same ByteBufferOutputStream) we don't know what we'll need. The JMH benchmark did confirm that 2x was more performant than 1.1x for FetchResponse. Existing usages of ByteBufferOutputStream were not modified and still use 1.1x |
||
| sendConsumer.accept(send); | ||
| byteArrayOutputStream.reset(); | ||
| } | ||
| } | ||
Uh oh!
There was an error while loading. Please reload this page.