• Master nodes
• Both GFS and TFS (taobao file system) uses single master node with multiple slaves. 
• My comments: facebook does not use such approaches? I think it uses P2P approach.

Differences

• Functions
• GFS/HDFS are more popular
• On top of it, we can build the big tables such as BigTable, Hypertable, HBase.
• Blob File System
• Usually used for Photos, Albums (These are called Blob Data)
• Challenges
• Blob FS
• For each write, it will request the master node to assign a blob number and machine lists to write to. 
• Challenge
• The volume of meta-data is of huge size
• E.g., Taobao has more than 10G photos, assume each photo has meta-data of size 20Bytes, the total size will be 20*10 = 200G, much more than the memory of a single machine.
• Solution
• The meta-data is not stored in Blob FS.
• The meta-data are stored in external systems.
• e.g., Taobao TFS has an id for each photo, the id are stored in external databases, such as Oracle or Mysql sharding cluster.
• Blob FS use chunk to organize data.
• Every blob file is a logical file. 
• Every chuck in a physical file.
• Multiple logical file will share a physical file, so that it can reduces the number of physical files. 
• All meta-data for physical files can be in memory, thus every read of Blob file only needs one I/O access.
• HDFS/GFS
• GFS v2 may be able to combine GFS and Blob FS into a system. It is difficult to do so, since
• It needs to support both large and small files.
• The size of meta-data is too large and thus the master nodes also need to be distributed.

• Master and Slaves
• Both GFS and HDFS use single master + multiple slaves mode.
• The master node maintains the check-point, data migration, log
• Data blocks and replication
• It maintains multiple copies (usually 3) to support better reliability and performance
• Tree structure
• It maintains a tree-structure file system, and allows operations like those under Linux system
• copy, rename, move, copy, delete etc.

• File appends
• GFS 
• allow multiple appends and allow multiple clients to append simultaneously
• if every append will visit the master node, it will be of low efficiency. 
• GFS use “Leasing Mechanism” to deliver the write permission of Chunk to Chunk Server.
• Check server can write the chunks within the lease (e.g., 12s).
• Since multiple servers may write simultaneously, and the API is asynchronous, the records might be in different order. This makes the system design very complicated.
• HDFS
• Only allow one open and data append
• The client will first write the data in local tmp file, and when the size of tmp data reach the size of a chunk (64M), then it will ask the HDFS master to assign a machine and chucn number to write the Chuck data. 
• Advantage
• The master will not be bottleneck. Since each write only occur when the data accumulated to be up to 64M.
• Disadvantage
• If the machine down in the process, some logs are not in the HDFS, and it might lose some data.
• Master failure
• GFS
• Backup master node. When the main master node fails, a new master node will be voted from the backup nodes.
• Support snapshot by using “copy on write” approach.
• HDFS
• HDFS needs human-interations in terms of failure.
• Does not support snapshot. 
• Garbage Collection (GC)
• GFS
• Lazy GC. 
• It will marks the files to be deleted (e.g., rename the file to one contains time information), thus the files will not be able to be visited by normal users.
• The master node will periodically check the files and delete the out-date ones (usually the files with more than 3 days).
• HDFS
• HDFS use simple and directly delete mechanism.

• In terms of the problem of which query should be executed first, it should not be depended in the coding process.
• But, there should be an extra phase to determine such kind of schedule. 
• Importance of which query to be executed first
• “比如我们现在有两个查询需求。一个是查询你在淘宝上买过东西的朋友，另一个是查询你在淘宝上买过保时捷的朋友。常理来说，我们会先想到查询你在淘宝上的朋友，再进行另一个条件的查询，比如这样：”

  IdList friends = waitFor(getFriends(myId));
  yield return getTaoBaoBuyers(friends);
  

  但是对于保时捷这个查询而言，这是不对的，因为淘宝上买保时捷的人是很少的，可能就一两个，而淘宝上的好友数可能有上百。因此保时捷的查询应该是这个次序比较优化：

  IdList buyers = waitFor(getPorscheBuyer());
  yield return getFriends(buyers);

• Provide scalability and high availability without compromising performance
• Linear scalability and proven fault-tolerance on commodity hardware or cloud infrastructure make it the perfect platfrom for mission critical data.
• Support replicating across multiple datacenters and providing low latency for your users
• Offer the convenience of column indexes with the performance of log-structured updates, strong support for denormalization and materialized views, and powerful built-in caching.

Current status