hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Gary Helmling" <ghelml...@gmail.com>
Subject Re: Review Request: [HBASE-2321] [HBASE-2002] Add support for per-region dynamically registered RPC endpoints for coprocessors and allow configurable RPC client/server implementations
Date Fri, 24 Sep 2010 07:01:53 GMT

This is an automatically generated e-mail. To reply, visit:

(Updated 2010-09-24 00:01:53.015127)

Review request for hbase, Andrew Purtell and Jonathan Gray.


Changes since last:
- HTable.BatchCall and HTable.BatchCallback moved to Batch.Call and Batch.Callback
- added Batch.returning() factory method for generating common Batch.Call instances
- o.a.h.h.client.Exec now extends o.a.h.h.ipc.Invocation since there was quite a bit of overlap

To go into more detail about client usage, let me start with a slightly more applicable case
for an exported CoprocessorProtocol (taken from the HBASE-2001 patch).  Suppose a user creates
a coprocessor that returns simple stats for the rows and key values in a region.  First the
coprocessor would define an RPC interface:

  public interface CountProtocol extends CoprocessorProtocol {
    public long getRowCount();
    public long getRowCount(Filter filt);
    public long getKeyValueCount();

The coprocessor code would implement this interface and either use the region coprocessor
hooks to track state or to scan the data directly (we can happily ignore the details).

When the coprocessor is loaded with the region, the coprocessor framework identifies that
the code implements a CoprocessorProtocol and automatically registers it with the region (see

So we then need a way for clients to use a standard set of calls to invoke these registered
protocols, when we know nothing about the details of the protocols themselves.  I looked at
two options:

Option 1) Use a ioctl-type interface, passing the desired protocol class, method name (or
code), and an optional list of arguments (in some cases we might also need the list of class
types for the arguments).  Lookup the method in the protocol class based on the arguments
or argument types if available, then batch the call to the server and return the result. 
In this case, the client code might be:

  HTable table = new HTable("mytable");
  List<Row> rows = ...rows for regions to query...
  Invocation call = new Invocation(CountProtocol.class, "getRowCount");
  Map<byte[],Long> results = table.exec(rows, call);

This works alright for a basic method call, but there's no verification that the method exists
or can be called with the given arguments until runtime.  In addition, if we want to make
multiple method calls, we have to instantiate additional Invocation instances.  And if we
want to tie those results together by region, we would need to do so by cross referencing
the results in multiple maps.

Option 2) Expose a dynamic proxy of the protocol interface directly to the client code.  In
this case, the client can directly call methods on the proxy, with compile time checking and
readability as normal code, not deconstructed methods.  The Batch.Call interface allows the
user to implement the client calls against a proxy instance we hand back.  As an optimization
for the simple case from (1) we have a factory method to do the same thing with defining your
own anonymous class:

  HTable table = new HTable("mytable");
  List<Row> rows = ...rows for regions to query...
  Batch.Call<CountProtocol,Long> call = Batch.returning(CountProtocol.class, "getRowCount");
  Map<byte[],Long> results = table.exec(CountProtocol.class, rows, call);

However, if you want to combine multiple calls to the same region or do some manipulation
of the results, an anonymous class may be more useful:

  HTable table = new HTable("mytable");
  List<Row> rows = ...rows for regions to query...
  // combine row count and kv count for region
  Map<byte[],Pair<Long,Long>> results = table.exec(CountProtocol.class, rows,
      new Batch.Call<CountProtocol,Pair<Long,Long>>() {
        public Pair<Long,Long> call(CountProtocol counter) {
          return new Pair(counter.getRowCount(), counter.getKeyValueCount());

  // or return the average number of KVs per row
  Map<byte[],Double> results = table.exec(CountProtocol.class, rows,
      new Batch.Call<CountProtocol,Double>() {
        public Double call(CountProtocol counter) {
          return ((double)counter.getKeyValueCount()) / ((double)counter.getRowCount());

The anonymous class does make for more annoying boilerplate in Java, but I think the usage
of the CountProtocol interface is actually more straightforward.  And since the user code
is executed together per-region, there's no need to manually stitch together region results
from multiple method calls.  So despite the additional brace-noise, I think the proxied approach
allow for a lot more client flexibility and cleaner code in terms of the calls and manipulations
going on.

For some other examples see the org.apache.hadoop.hbase.regionserver.TestServerCustomProtocol
test case, or the TestCommandTarget test case in the HBASE-2001 patch up for review.


This is really two separate patches in one, though with some overlapping changes.  If necessary
I can split them apart for separate review.  Please let me know if that would make review

Part 1:
Port over of HADOOP-6422 to the HBase RPC code.  The goal of this change is to allow alternate
RPC client/server implementations to be enabled through a simple configuration change.  Ultimately
I would like to use this to allow secure RPC to be enabled through configuration, while not
blocking normal (current) RPC operation on non-secure Hadoop versions.

This portion of the patch abstracts out two interfaces from the RPC code:

RpcEngine: HBaseRPC uses this to obtain proxy instances for client calls and server instances
for HMaster and HRegionServer
RpcServer: this allows differing RPC server implementations, breaking the dependency on HBaseServer

The bulk of the current code from HBaseRPC is moved into WritableRpcEngine and is unchanged
other than the interface requirements.  So the current call path remains the same, other than
the HBaseRPC.getProtocolEngine() abstraction.

Part 2:
The remaining changes provide server-side hooks for registering new RPC protocols/handlers
(per-region to support coprocessors), and client side hooks to support dynamic execution of
the registered protocols.

The new RPC protocol actions are constrained to org.apache.hadoop.hbase.ipc.CoprocessorProtocol
implementations (which extends VersionedProtocol) to prevent arbitrary execution of methods
against HMasterInterface, HRegionInterface, etc.

For protocol handler registration, HRegionServer provides a new method:

  public <T extends CoprocessorProtocol> boolean registerProtocol(
      byte[] region, Class<T> protocol, T handler)

which builds a Map of region name to protocol instances for dispatching client calls.

Client invocations are performed through HTable, which adds the following methods:

  public <T extends CoprocessorProtocol> T proxy(Class<T> protocol, Row row)

This directly returns a proxy instance to the CoprocessorProtocol implementation registered
for the region serving row "row".  Any method calls will be proxied to the region's server
and invoked using the map of registered region name -> handler instances.

Calls directed against multiple rows are a bit more complicated.  They are supported with
the methods:

  public <T extends CoprocessorProtocol, R> void exec(
      Class<T> protocol, List<? extends Row> rows,
      BatchCall<T,R> callable, BatchCallback<R> callback)

  public <T extends CoprocessorProtocol, R> void exec(
      Class<T> protocol, RowRange range,
      BatchCall<T,R> callable, BatchCallback<R> callback)

where BatchCall and BatchCallback are simple interfaces defining the methods to be called
and a callback instance to be invoked for each result.

For the sample CoprocessorProtocol interface:

  interface PingProtocol extends CoprocessorProtocol {
    public String ping();
    public String hello(String name);

a client invocation might look like:

    final Map<byte[],R> results = new TreeMap<byte[],R>(...)
    List<Row> rows = ...
    table.exec(PingProtocol.class, rows,
        new HTable.BatchCall<PingProtocol,String>() {
          public String call(PingProtocol instance) {
            return instance.ping();
        new BatchCallback<R>(){
          public void update(byte[] region, byte[] row, R value) {
            results.put(region, value);

The BatchCall.call() method will be invoked for each row in the passed in list, and the BatchCallback.update()
method will be invoked for each return value.  However, currently the PingProtocol.ping()
invocation will result in a separate RPC call per row, which is less that ideal.

Support is in place to make use of the HRegionServer.multi() invocations for batched RPC (see
the org.apache.hadoop.hbase.client.Exec class), but this does not mesh well with the current
client-side interface.

In addition to standard code review, I'd appreciate any thoughts on the client interactions
in particular, and whether they would meet some of the anticipated uses of coprocessors.

This addresses bugs HBASE-2002 and HBASE-2321.

Diffs (updated)

  src/main/java/org/apache/hadoop/hbase/client/Action.java 556ea81 
  src/main/java/org/apache/hadoop/hbase/client/Batch.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/client/Exec.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/client/ExecResult.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/client/HConnection.java 65f7618 
  src/main/java/org/apache/hadoop/hbase/client/HConnectionManager.java fbdec0b 
  src/main/java/org/apache/hadoop/hbase/client/HTable.java 0dbf263 
  src/main/java/org/apache/hadoop/hbase/client/MultiAction.java c6ea838 
  src/main/java/org/apache/hadoop/hbase/client/MultiResponse.java 91bd04b 
  src/main/java/org/apache/hadoop/hbase/client/RowRange.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/client/Scan.java 29b3cb0 
  src/main/java/org/apache/hadoop/hbase/io/HbaseObjectWritable.java 83f623d 
  src/main/java/org/apache/hadoop/hbase/ipc/ConnectionHeader.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/ipc/CoprocessorProtocol.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/ipc/ExecRPCInvoker.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/ipc/HBaseClient.java 2b5eeb6 
  src/main/java/org/apache/hadoop/hbase/ipc/HBaseRPC.java e23a629 
  src/main/java/org/apache/hadoop/hbase/ipc/HBaseServer.java e4c356d 
  src/main/java/org/apache/hadoop/hbase/ipc/HRegionInterface.java a4810a6 
  src/main/java/org/apache/hadoop/hbase/ipc/Invocation.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/ipc/RpcEngine.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/ipc/RpcServer.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/ipc/WritableRpcEngine.java PRE-CREATION 
  src/main/java/org/apache/hadoop/hbase/master/HMaster.java 36ba5c1 
  src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java 1be9cf5 
  src/main/java/org/apache/hadoop/hbase/regionserver/HRegionServer.java e9d7751 
  src/main/resources/hbase-default.xml 5452fd1 
  src/test/java/org/apache/hadoop/hbase/regionserver/TestServerCustomProtocol.java PRE-CREATION

Diff: http://review.cloudera.org/r/816/diff




View raw message