Uploaded image for project: 'HBase'
  1. HBase
  2. HBASE-6490

'dfs.client.block.write.retries' value could be increased in HBase

    Details

    • Type: Improvement
    • Status: Open
    • Priority: Minor
    • Resolution: Unresolved
    • Affects Version/s: 2.0.0
    • Fix Version/s: None
    • Component/s: master, regionserver
    • Labels:
      None
    • Environment:

      all

      Description

      When allocating a new node during writing, hdfs tries 'dfs.client.block.write.retries' times (default 3) to write the block. When it fails, it goes back to the nanenode for a new list, and raises an error if the number of retries is reached. In HBase, if the error is while we're writing a hlog file, it will trigger a region server abort (as hbase does not trust the log anymore). For simple case (new, and as such empty log file), this seems to be ok, and we don't lose data. There could be some complex cases if the error occurs on a hlog file with already multiple blocks written.

      Logs lines are:
      "Exception in createBlockOutputStream", then "Abandoning block " followed by "Excluding datanode " for a retry.
      IOException: "Unable to create new block.", when the number of retries is reached.

      Probability of occurence seems quite low, (number of bad nodes / number of nodes)^(number of retries), and it implies that you have a region server without its datanode. But it's per new block.

      Increasing the default value of 'dfs.client.block.write.retries' could make sense to be better covered in chaotic conditions.

        Issue Links

          Activity

          Hide
          qwertymaniac Harsh J added a comment -

          Would HDFS-4246 help here?

          Show
          qwertymaniac Harsh J added a comment - Would HDFS-4246 help here?
          Hide
          nkeywal Nicolas Liochon added a comment -

          I don't think it's an issue to increase globally. I haven't yet looked at the memstore flush, but I think it's gonna be the same or worse: we don't really expecting a write to fail.
          I need to check if it can be a fixed value or if we need to take into account the replication factor or the number of machine...

          Show
          nkeywal Nicolas Liochon added a comment - I don't think it's an issue to increase globally. I haven't yet looked at the memstore flush, but I think it's gonna be the same or worse: we don't really expecting a write to fail. I need to check if it can be a fixed value or if we need to take into account the replication factor or the number of machine...
          Hide
          stack stack added a comment -

          What should we increase it to N? We can't increase it just for WAL... it'd be globally?

          Show
          stack stack added a comment - What should we increase it to N? We can't increase it just for WAL... it'd be globally?

            People

            • Assignee:
              Unassigned
              Reporter:
              nkeywal Nicolas Liochon
            • Votes:
              0 Vote for this issue
              Watchers:
              6 Start watching this issue

              Dates

              • Created:
                Updated:

                Development