[jira] [Commented] (FLINK-18811) if a disk is damaged, taskmanager should choose another disk for temp dir , rather than throw an IOException, which causes flink job restart over and over again

["Kai Chen (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-18811?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17171367#comment-17171367 ] 

Kai Chen commented on FLINK-18811:
----------------------------------

I made some test today, and the change below looks good to me:
{code:java}
// SpillingAdaptiveSpanningRecordDeserializer.java
private FileChannel createSpillingChannel() throws IOException {
   if (spillFile != null) {
      throw new IllegalStateException("Spilling file already exists.");
   }

   // try to find a unique file name for the spilling channel
   int maxAttempts = 10;
   String[] tempDirs = this.tempDirs;
   for (int attempt = 0; attempt < maxAttempts; attempt++) {
      int dirIndex = rnd.nextInt(tempDirs.length);
      String directory = tempDirs[dirIndex];
      spillFile = new File(directory, randomString(rnd) + ".inputchannel");
      try {
         if (spillFile.createNewFile()) {
            return new RandomAccessFile(spillFile, "rw").getChannel();
         }
      } catch (IOException e) {
         // if there is no tempDir left to try
         if(tempDirs.length <= 1) {
            throw e;
         }
         LOG.warn("Caught an IOException when creating spill file: " + directory + ". Attempt " + attempt, e);
         tempDirs = (String[])ArrayUtils.remove(tempDirs,dirIndex);
      }
   }
{code}
Does anyone have any good ideas?

> if a disk is damaged, taskmanager should choose another disk for temp dir , rather than throw an IOException, which causes flink job restart over and over again
> ----------------------------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: FLINK-18811
>                 URL: https://issues.apache.org/jira/browse/FLINK-18811
>             Project: Flink
>          Issue Type: Improvement
>          Components: Runtime / Network
>         Environment: flink-1.10
>            Reporter: Kai Chen
>            Priority: Major
>         Attachments: flink_disk_error.png
>
>
> I met this Exception when a disk was damaged:
> !flink_disk_error.png!
> I think, if a disk is damaged, taskmanager should choose another disk for temp dir , rather than throw an IOException, which causes flink job restart over and over again.
> If we could just change “SpillingAdaptiveSpanningRecordDeserializer" like this:
> {code:java}
> // SpillingAdaptiveSpanningRecordDeserializer.java
> private FileChannel createSpillingChannel() throws IOException {
>    if (spillFile != null) {
>       throw new IllegalStateException("Spilling file already exists.");
>    }
>    // try to find a unique file name for the spilling channel
>    int maxAttempts = 10;
>    String[] tempDirs = this.tempDirs;
>    for (int attempt = 0; attempt < maxAttempts; attempt++) {
>       int dirIndex = rnd.nextInt(tempDirs.length);
>       String directory = tempDirs[dirIndex];
>       spillFile = new File(directory, randomString(rnd) + ".inputchannel");
>       try {
>          if (spillFile.createNewFile()) {
>             return new RandomAccessFile(spillFile, "rw").getChannel();
>          }
>       } catch (IOException e) {
>          // if there is no tempDir left to try
>          if(tempDirs.length <= 1) {
>             throw e;
>          }
>          LOG.warn("Caught an IOException when creating spill file: " + directory + ". Attempt " + attempt, e);
>          tempDirs = (String[])ArrayUtils.remove(tempDirs,dirIndex);
>       }
>    }
> {code}



--
This message was sent by Atlassian Jira
(v8.3.4#803005)