[
https://issues.apache.org/jira/browse/HBASE-12141?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14156267#comment-14156267
]
Nicolas Liochon commented on HBASE-12141:
-----------------------------------------
Yeah, the strategy was to keep the message small enough (if multiple servers
fail simultaneously, we send multiple messages instead of one). As well, we
send the message multiple times in case it got lost somewhere. I had issue with
Netty 3.x when tried to add frames. I haven't tried very hard. We could make
MAX_SERVER_PER_MESSAGE configurable for network with a very small mtu? It's
also possible to compress the message. Once again, I had issue with Netty 3.x
for this in the past.
This said, I would be interested to understand the network config.
> ClusterStatus message might exceed max datagram payload limits
> --------------------------------------------------------------
>
> Key: HBASE-12141
> URL: https://issues.apache.org/jira/browse/HBASE-12141
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.98.3
> Reporter: Andrew Purtell
>
> The multicast ClusterStatusPublisher and its companion listener are using
> datagram channels without any framing. I think this is an issue because
> Netty's ProtobufDecoder expects a complete PB message to be available in the
> ChannelBuffer yet ClusterStatus messages can be large and might exceed the
> maximum datagram payload size. As one user reported on list:
> {noformat}
> org.apache.hadoop.hbase.client.ClusterStatusListener - ERROR - Unexpected
> exception, continuing.
> com.google.protobuf.InvalidProtocolBufferException: Protocol message tag had
> invalid wire type.
> at
> com.google.protobuf.InvalidProtocolBufferException.invalidWireType(InvalidProtocolBufferException.java:99)
> at
> com.google.protobuf.UnknownFieldSet$Builder.mergeFieldFrom(UnknownFieldSet.java:498)
> at
> com.google.protobuf.GeneratedMessage.parseUnknownField(GeneratedMessage.java:193)
> at
> org.apache.hadoop.hbase.protobuf.generated.ClusterStatusProtos$ClusterStatus.<init>(ClusterStatusProtos.java:7554)
> at
> org.apache.hadoop.hbase.protobuf.generated.ClusterStatusProtos$ClusterStatus.<init>(ClusterStatusProtos.java:7512)
> at
> org.apache.hadoop.hbase.protobuf.generated.ClusterStatusProtos$ClusterStatus$1.parsePartialFrom(ClusterStatusProtos.java:7689)
> at
> org.apache.hadoop.hbase.protobuf.generated.ClusterStatusProtos$ClusterStatus$1.parsePartialFrom(ClusterStatusProtos.java:7684)
> at
> com.google.protobuf.AbstractParser.parsePartialFrom(AbstractParser.java:141)
> at
> com.google.protobuf.AbstractParser.parseFrom(AbstractParser.java:176)
> at
> com.google.protobuf.AbstractParser.parseFrom(AbstractParser.java:182)
> at
> com.google.protobuf.AbstractParser.parseFrom(AbstractParser.java:49)
> at
> org.jboss.netty.handler.codec.protobuf.ProtobufDecoder.decode(ProtobufDecoder.java:122)
> at
> org.jboss.netty.handler.codec.oneone.OneToOneDecoder.handleUpstream(OneToOneDecoder.java:66)
> at
> org.jboss.netty.channel.Channels.fireMessageReceived(Channels.java:268)
> at
> org.jboss.netty.channel.socket.oio.OioDatagramWorker.process(OioDatagramWorker.java:52)
> at
> org.jboss.netty.channel.socket.oio.AbstractOioWorker.run(AbstractOioWorker.java:73)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
> at java.lang.Thread.run(Thread.java:745)
> {noformat}
> The javadoc for ProtobufDecoder says:
> {quote}
> Decodes a received ChannelBuffer into a Google Protocol Buffers Message and
> MessageLite. Please note that this decoder must be used with a proper
> FrameDecoder such as ProtobufVarint32FrameDecoder or
> LengthFieldBasedFrameDecoder if you are using a stream-based transport such
> as TCP/IP.
> {quote}
> and even though we are using a datagram transport we have related issues,
> depending on what the sending and receiving OS does with overly large
> datagrams:
> - We may receive a datagram with a truncated message
> - We may get an upcall when processing one fragment of a fragmented datagram,
> where the complete message is not available yet
> - We may not be able to send the overly large ClusterStatus in the first
> place. Linux claims to do PMTU and return EMSGSIZE if a datagram packet
> payload exceeds the MTU, but will send a fragmented datagram if PMTU is
> disabled. I'm surprised we have the above report given the default is to
> reject overly large datagram payloads, so perhaps the user is using a
> different server OS or Netty datagram channels do their own fragmentation (I
> haven't checked).
> In any case, the server and client pipelines are definitely not doing any
> kind of framing. This is the multicast status listener from 0.98 for example:
> {code}
> b.setPipeline(Channels.pipeline(
> new
> ProtobufDecoder(ClusterStatusProtos.ClusterStatus.getDefaultInstance()),
> new ClusterStatusHandler()));
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
