apache kafka系列之在zookeeper中儲存結構

五柳-先生發表於2015-09-22
ApacheKafka

1.topic註冊資訊

/brokers/topics/[topic] :

儲存某個topic的partitions所有分配資訊

Schema:

{
    "version": "版本編號目前固定為數字1",
    "partitions": {
        "partitionId編號": [
            同步副本組brokerId列表
        ],
        "partitionId編號": [
            同步副本組brokerId列表
        ],
        .......
    }
}

Example:

{
"version": 1,
"partitions": {
"0": [1, 2],
"1": [2, 1],
"2": [1, 2],
}
}

說明：紫紅色為patitions編號，藍色為同步副本組brokerId列表

2.partition狀態資訊

/brokers/topics/[topic]/partitions/[0...N]  其中[0..N]表示partition索引號

/brokers/topics/[topic]/partitions/[partitionId]/state

Schema:

{
"controller_epoch": 表示kafka叢集中的中央控制器選舉次數,
"leader": 表示該partition選舉leader的brokerId,
"version": 版本編號預設為1,
"leader_epoch": 該partition leader選舉次數,
"isr": [同步副本組brokerId列表]
}

 
Example:

{
"controller_epoch": 1,
"leader": 2,
"version": 1,
"leader_epoch": 0,
"isr": [2, 1]
}

3. Broker註冊資訊

/brokers/ids/[0...N]                 

每個broker的配置檔案中都需要指定一個數字型別的id(全域性不可重複),此節點為臨時znode(EPHEMERAL)

Schema:

{
"jmx_port": jmx埠號,
"timestamp": kafka broker初始啟動時的時間戳,
"host": 主機名或ip地址,
"version": 版本編號預設為1,
"port": kafka broker的服務端埠號,由server.properties中引數port確定
}

 
Example:

{
"jmx_port": 6061,

"timestamp":"1403061899859"
"version": 1,
"host": "192.168.1.148",
"port": 9092
}

4. Controller epoch: 

/controller_epoch -> int (epoch)   

此值為一個數字,kafka叢集中第一個broker第一次啟動時為1，以後只要叢集中center controller中央控制器所在broker變更或掛掉，就會重新選舉新的center controller，每次center controller變更controller_epoch值就會 + 1; 

5. Controller註冊資訊:

/controller -> int (broker id of the controller)  儲存center controller中央控制器所在kafka broker的資訊

Schema:

{
"version": 版本編號預設為1,
"brokerid": kafka叢集中broker唯一編號,
"timestamp": kafka broker中央控制器變更時的時間戳
}

 
Example:

{
"version": 1,
"brokerid": 3,
"timestamp": "1403061802981"
}



Consumer and Consumer group概念: 
a.每個consumer客戶端被建立時,會向zookeeper註冊自己的資訊;
b.此作用主要是為了"負載均衡".
c.同一個Consumer Group中的Consumers，Kafka將相應Topic中的每個訊息只傳送給其中一個Consumer。
d.Consumer Group中的每個Consumer讀取Topic的一個或多個Partitions，並且是唯一的Consumer；
e.一個Consumer group的多個consumer的所有執行緒依次有序地消費一個topic的所有partitions,如果Consumer group中所有consumer匯流排程大於partitions數量，則會出現空閒情況;
舉例說明：
kafka叢集中建立一個topic為report-log   4 partitions 索引編號為0,1,2,3
假如有目前有三個消費者node：注意-->一個consumer中一個消費執行緒可以消費一個或多個partition.
如果每個consumer建立一個consumer thread執行緒,各個node消費情況如下，node1消費索引編號為0,1分割槽，node2費索引編號為2,node3費索引編號為3
如果每個consumer建立2個consumer thread執行緒，各個node消費情況如下(是從consumer node先後啟動狀態來確定的)，node1消費索引編號為0,1分割槽；node2費索引編號為2,3；node3為空閒狀態
總結 ：
從以上可知，Consumer Group中各個consumer是根據先後啟動的順序有序消費一個topic的所有partitions的。
如果Consumer Group中所有consumer的匯流排程數大於partitions數量，則可能consumer thread或consumer會出現空閒狀態。

Consumer均衡演算法 
當一個group中,有consumer加入或者離開時,會觸發partitions均衡.均衡的最終目的,是提升topic的併發消費能力. 
1) 假如topic1,具有如下partitions: P0,P1,P2,P3 
2) 加入group中,有如下consumer: C0,C1 
3) 首先根據partition索引號對partitions排序: P0,P1,P2,P3 
4) 根據(consumer.id + '-'+ thread序號)排序: C0,C1 
5) 計算倍數: M = [P0,P1,P2,P3].size / [C0,C1].size,本例值M=2(向上取整) 
6) 然後依次分配partitions: C0 = [P0,P1],C1=[P2,P3],即Ci = [P(i * M),P((i + 1) * M -1)]

6. Consumer註冊資訊:

每個consumer都有一個唯一的ID(consumerId可以通過配置檔案指定,也可以由系統生成),此id用來標記消費者資訊.

/consumers/[groupId]/ids/[consumerIdString]

是一個臨時的znode,此節點的值為請看consumerIdString產生規則,即表示此consumer目前所消費的topic + partitions列表.

consumerId產生規則：

   StringconsumerUuid = null;
    if(config.consumerId!=null && config.consumerId)
      consumerUuid = consumerId;
    else {
      String uuid = UUID.randomUUID()
      consumerUuid = "%s-%d-%s".format(
        InetAddress.getLocalHost.getHostName, System.currentTimeMillis,
        uuid.getMostSignificantBits().toHexString.substring(0,8));

     }
     String consumerIdString = config.groupId + "_" + consumerUuid;

Schema:

{
"version": 版本編號預設為1,
"subscription": { //訂閱topic列表
"topic名稱": consumer中topic消費者執行緒數
},
"pattern": "static",
"timestamp": "consumer啟動時的時間戳"
}

 
Example:
{
"version": 1,
"subscription": {
"open_platform_opt_push_plus1": 5
},
"pattern": "static",
"timestamp": "1411294187842"
} 
 

7. Consumer owner:

/consumers/[groupId]/owners/[topic]/[partitionId] -> consumerIdString + threadId索引編號

當consumer啟動時,所觸發的操作:

a) 首先進行"Consumer Id註冊";

b) 然後在"Consumer id 註冊"節點下注冊一個watch用來監聽當前group中其他consumer的"退出"和"加入";只要此znode path下節點列表變更,都會觸發此group下consumer的負載均衡.(比如一個consumer失效,那麼其他consumer接管partitions).

c) 在"Broker id 註冊"節點下,註冊一個watch用來監聽broker的存活情況;如果broker列表變更,將會觸發所有的groups下的consumer重新balance.

8. Consumer offset:

/consumers/[groupId]/offsets/[topic]/[partitionId] -> long (offset)

用來跟蹤每個consumer目前所消費的partition中最大的offset

此znode為持久節點,可以看出offset跟group_id有關,以表明當消費者組(consumer group)中一個消費者失效,

重新觸發balance,其他consumer可以繼續消費.

9. Re-assign partitions

/admin/reassign_partitions

{
    "fields":[
      {
         "name":"version",
         "type":"int",
         "doc":"version id"
      },
      {
         "name":"partitions",
         "type":{
            "type":"array",
            "items":{
               "fields":[
                  {
                     "name":"topic",
                     "type":"string",
                     "doc":"topic of the partition to be reassigned"
                  },
                  {
                     "name":"partition",
                     "type":"int",
                     "doc":"the partition to be reassigned"
                  },
                  {
                     "name":"replicas",
                     "type":"array",
                     "items":"int",
                     "doc":"a list of replica ids"
                  }
               ],
            }
            "doc":"an array of partitions to be reassigned to new replicas"
         }
      }
   ]
}
 
Example:
{
  "version": 1,
  "partitions":
     [
        {
            "topic": "Foo",
            "partition": 1,
            "replicas": [0, 1, 3]
        }
     ]            
}

 

10. Preferred replication election

/admin/preferred_replica_election

 

{
   "fields":[
      {
         "name":"version",
         "type":"int",
         "doc":"version id"
      },
      {
         "name":"partitions",
         "type":{
            "type":"array",
            "items":{
               "fields":[
                  {
                     "name":"topic",
                     "type":"string",
                     "doc":"topic of the partition for which preferred replica election should be triggered"
                  },
                  {
                     "name":"partition",
                     "type":"int",
                     "doc":"the partition for which preferred replica election should be triggered"
                  }
               ],
            }
            "doc":"an array of partitions for which preferred replica election should be triggered"
         }
      }
   ]
}
 
例子:
 
{
  "version": 1,
  "partitions":
     [
        {
            "topic": "Foo",
            "partition": 1         
        },
        {
            "topic": "Bar",
            "partition": 0         
        }
     ]            
}

 

11. 刪除topics
/admin/delete_topics

Schema:
{  "fields":
    [ {"name": "version", "type": "int", "doc": "version id"},
      {"name": "topics",
       "type": { "type": "array", "items": "string", "doc": "an array of topics to be deleted"}
      } ]
}
 
例子:
{
  "version": 1,
  "topics": ["foo", "bar"]
}

Topic配置

/config/topics/[topic_name]

例子

{
  "version": 1,
  "config": {
    "config.a": "x",
    "config.b": "y",
    ...
   }
}

轉載： http://blog.csdn.net/lizhitao/article/details/23744675

相關文章