Uploaded image for project: 'HBase'
  1. HBase
  2. HBASE-14436

HTableDescriptor#addCoprocessor will always make RegionCoprocessorHost create new Configuration

    XMLWordPrintableJSON

    Details

    • Hadoop Flags:
      Reviewed

      Description

      HTableDescriptor#addCoprocessor will set the coprocessor value as following format:

       public HTableDescriptor addCoprocessor(String className, Path jarFilePath,
                                   int priority, final Map<String, String> kvs)
        throws IOException {
        ...
        String value = ((jarFilePath == null)? "" : jarFilePath.toString()) +
              "|" + className + "|" + Integer.toString(priority) + "|" +
              kvString.toString();
        ...
      }
      

      If the 'jarFilePath' is null, the 'value' will always has the format '|className|priority|' even if 'kvs' is null, which means no extra arguments for the coprocessor. Then, in the server side, RegionCoprocessorHost#getTableCoprocessorAttrsFromSchema will load the table coprocessors as:

        static List<TableCoprocessorAttribute> getTableCoprocessorAttrsFromSchema(Configuration conf,
            HTableDescriptor htd) {
          ...
                  try {
                    cfgSpec = matcher.group(4); // => cfgSpec will be '|' for the format '|className|priority|'
                  } catch (IndexOutOfBoundsException ex) {
                    // ignore
                  }
                  Configuration ourConf;
                  if (cfgSpec != null) {  // => cfgSpec will be '|' for the format '|className|priority|'
                    ourConf = new Configuration(false);
                    HBaseConfiguration.merge(ourConf, conf);
                  }
          ...
      }
      

      The 'cfgSpec' will be '|' for the coprocessor formatted as '|className|priority|', so that always create a new Configuration.
      In our production, there are a lot of tables having table-level coprocessors, so that the region server will create new Configurations for each region of the table, this will consume a certain number of memory when we have many such regions.
      To fix the problem, we can make the HTableDescriptor not append the '|' if no extra arguments for the coprocessor, or check the 'cfgSpec' more strictly in server side which could avoid creating new Configurations for existed such regions after the regions reopened. Discussions and suggestions are welcomed.

        Attachments

        1. HBASE-14436-trunk-v1.patch
          0.8 kB
          Jianwei Cui
        2. HBASE-14436-trunk-v2.patch
          0.8 kB
          Jianwei Cui

          Activity

            People

            • Assignee:
              stack Michael Stack
              Reporter:
              cuijianwei Jianwei Cui
            • Votes:
              0 Vote for this issue
              Watchers:
              5 Start watching this issue

              Dates

              • Created:
                Updated:
                Resolved: