[HADOOP-17901] Performance degradation in Text.append() after HADOOP-16951

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 3.4.0
• Fix Version/s: 3.4.0
• Component/s: common
• Labels: pull-request-available
• Target Version/s: 3.4.0
• Hadoop Flags: Reviewed
• Assignee: Peter Bacsko
• Reporter: Peter Bacsko
• Created: 08/Sep/21 11:49
• Updated: 27/Jan/24 05:46
• Resolved: 10/Sep/21 23:03

Description

We discovered a serious performance degradation in Text.append().

The problem is that the logic which intends to increase the size of the backing array does not work as intended.
It's very difficult to spot, so I added extra logs to see what happens.

Let's add 4096 bytes of textual data in a loop:

  public static void main(String[] args) {
    Text text = new Text();
    String toAppend = RandomStringUtils.randomAscii(4096);

    for(int i = 0; i < 100; i++) {
      text.append(toAppend.getBytes(), 0, 4096);
    }
  }

With some debug printouts, we can observe:

2021-09-08 13:35:29,528 INFO  [main] io.Text (Text.java:append(251)) - length: 24576,  len: 4096, utf8ArraySize: 4096, bytes.length: 30720
2021-09-08 13:35:29,528 INFO  [main] io.Text (Text.java:append(253)) - length + (length >> 1): 36864
2021-09-08 13:35:29,528 INFO  [main] io.Text (Text.java:append(254)) - length + len: 28672
2021-09-08 13:35:29,528 INFO  [main] io.Text (Text.java:ensureCapacity(287)) - >>> enhancing capacity from 30720 to 36864
2021-09-08 13:35:29,528 INFO  [main] io.Text (Text.java:append(251)) - length: 28672,  len: 4096, utf8ArraySize: 4096, bytes.length: 36864
2021-09-08 13:35:29,528 INFO  [main] io.Text (Text.java:append(253)) - length + (length >> 1): 43008
2021-09-08 13:35:29,529 INFO  [main] io.Text (Text.java:append(254)) - length + len: 32768
2021-09-08 13:35:29,529 INFO  [main] io.Text (Text.java:ensureCapacity(287)) - >>> enhancing capacity from 36864 to 43008
2021-09-08 13:35:29,529 INFO  [main] io.Text (Text.java:append(251)) - length: 32768,  len: 4096, utf8ArraySize: 4096, bytes.length: 43008
2021-09-08 13:35:29,529 INFO  [main] io.Text (Text.java:append(253)) - length + (length >> 1): 49152
2021-09-08 13:35:29,529 INFO  [main] io.Text (Text.java:append(254)) - length + len: 36864
2021-09-08 13:35:29,529 INFO  [main] io.Text (Text.java:ensureCapacity(287)) - >>> enhancing capacity from 43008 to 49152
...

After a certain number of append() calls, subsequent capacity increments are small.

It's because the difference between two length + (length >> 1) values is always 6144 bytes. Because the size of the backing array is trailing behind the calculated value, the increment will also be 6144 bytes. This means that new arrays are constantly created.

Suggested solution: don't calculate the capacity in advance based on length. Instead, pass the required minimum to ensureCapacity(). Then the increment should depend on the actual size of the byte array if the desired capacity is larger.

Issue Links

• causes: HIVE-25686 UDFSpace result length calculation is incorrect after HADOOP-17901 and HADOOP-17905
• Dependent: HADOOP-17905 Modify Text.ensureCapacity() to efficiently max out the backing array size
• is caused by: HADOOP-16951 Tidy Up Text and ByteWritables Classes
• links to: GitHub Pull Request #3411