PaaS/云使用者常抱怨“多个进程抢同一文件导致写冲突”。在 HDFS 中，a lease is a temporary lock granted by NameNode to a client for exclusive write access.. 客户端打开文件时获得租约。并通过定期心跳续租，如果租约过期，NameNode 收回写权限并可授予其他客户端。从而防止“死锁写入”，

. 管道复制

Datanodes form a pipeline so that client sends data only once:

• The client streams packets to first DataNode.
• The node writes locally n forwards to next node.
• Acknowledgments flow back upstream to client.
• This reduces client bandwidth usage and spreads network load across nodes.
• If any node fails,pipeline is torn down and rebuilt automatically.
• Pain point：If you monitor only client‑side metrics you may miss that a middle DataNode is saturated – pipeline design pushes visibility into cluster.

. 一致性问题：部分写入

If a DataNode crashes mid‑pipeline:

• The client receives a broken ACK chain → write failure reported.
• The partially written replicas become invalid.
• NameNode marks block as corrupt and re‑assigns fresh replicas.

A two‑phase commit would require at least two round‑trips per packet plus multiple fsyncs – prohibitive for large‑block streaming workloads. HDFS assumes write failures are rare and tolerates eventual consistency via retries;higher‑level frameworks provide idempotent output handling.

三、读文件：数据局部性与机架感知

. 读取流程

sequenceDiagram participant C as 客户端 participant NN as NameNode participant DN1 as DataNode participant DN2 as DataNode 从C->NN来看，open NN-->C的观点是。返回 Block 列表 + 各副本地址 Note over C: NameNode 按距离排序 · 同节点 · 同机架 · 跨机架 说到C->DN1，读取 Block0 DN1-->C: 返回数据 --- 从C->DN2来看，读取 Block1 DN2-->C: 返回数据

. 机架感知

& // WordCount Reduce 函数
void reduce {
int sum = 0;for { sum += v.get;}
ctx.write);// }