[HBASE-17009] Revisiting the removement of managed connection and connection caching

• Type: Task
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Component/s: Operability
• Assignee: Yu Li
• Reporter: Yu Li
• Created: 03/Nov/16 07:29
• Updated: 02/Mar/18 04:25

Description

In HBASE-13197 we have done lots of good cleanups for Connection API, but among which HBASE-13252 dropped the feature of managed connection and connection caching, and this JIRA propose to have a revisit on this decision for below reasons.

Assume we have a long running process with multiple threads accessing HBase (a common case for streaming application), let's see what happens previously and now.

Previously:
User could create an HTable instance whenever they want w/o worrying about the underlying connections because HBase client will mange it automatically, say no matter how many threads there will be only one Connection instance

  @Deprecated
  public HTable(Configuration conf, final TableName tableName)
  throws IOException {
    ...
    this.connection = ConnectionManager.getConnectionInternal(conf);
    ...
  }

  static ClusterConnection getConnectionInternal(final Configuration conf)
    throws IOException {
    HConnectionKey connectionKey = new HConnectionKey(conf);
    synchronized (CONNECTION_INSTANCES) {
      HConnectionImplementation connection = CONNECTION_INSTANCES.get(connectionKey);
      if (connection == null) {
        connection = (HConnectionImplementation)createConnection(conf, true);
        CONNECTION_INSTANCES.put(connectionKey, connection);
      } else if (connection.isClosed()) {
        ConnectionManager.deleteConnection(connectionKey, true);
        connection = (HConnectionImplementation)createConnection(conf, true);
        CONNECTION_INSTANCES.put(connectionKey, connection);
      }
      connection.incCount();
      return connection;
    }
  }

Now:
User has to create the connection by themselves, using below codes like indicated in our recommendations

    Connection connection = ConnectionFactory.createConnection(conf);
    Table table = connection.getTable(tableName);

And they must make sure only one single connection created in one process instead of creating HTable instance freely, or else there might be many connections setup to zookeeper/RS with multiple threads. Also user might ask "when I should close the connection I close?" and the answer is "make sure don't close it until the process shutdown"

So now we have much more things for user to "Make sure", but custom is something hard to change. User used to create table instance in each thread (according to which table to access per requested) so probably they will still create connections everywhere, and then operators will have to crazily resolve all kinds of problems...

So I'm proposing to add back the managed connection and connection caching support. IMHO it's something good and ever existed in our implementation, so let's bring it back and save the workload for operators when they decided to upgrade from 1.x to 2.x

Thoughts?

Issue Links

• relates to: HBASE-16713 Bring back connection caching as a client API