The easiest way to achieve simultaneous writing to X nodes is to copy the input stream as it comes, not waiting for it to complete. So a writer connects to one of the write-enabled nodes for a given key, starts writing and he node copies the stream to other write-enabled nodes. Follows that each write-enabled node should have a way to pick up X-1 other nodes for writing. The component picking up nodes must have at least minimal intelligence not to choose a node in the same rack or, better yet, to follow user-defined availability zone rules.
A storage node also needs to have persistent connections already established to the rest of the nodes used for writing of the key space on this node. That is, if write latency is of any importance. And of course the protocol must provide for a way to establish consistency in case of any errors. For instance, what happens if the original write node dies after sending the data object to some but not all of other data nodes?
There are also questions about the structure of data nodes key space and key space distribution: do nodes come in sets of X, where each node from the set is identical (key-space-wise) to the rest of the sets nodes or is a key space on a node made of smaller units and a copy of that smaller unit can be distributed independently of the rest of units on the node? DHT (distributed hast table) and DHT key-ring literature has examples of both approaches. I personally prefer a full back-up (identical copies) because they are easy to create and manage.
Another question is the state of node coming back from the dead and the definition of "dead". See that in the next post.
On a side: it is difficult to find time to write while vacationing on Hawaii!
A storage node also needs to have persistent connections already established to the rest of the nodes used for writing of the key space on this node. That is, if write latency is of any importance. And of course the protocol must provide for a way to establish consistency in case of any errors. For instance, what happens if the original write node dies after sending the data object to some but not all of other data nodes?
There are also questions about the structure of data nodes key space and key space distribution: do nodes come in sets of X, where each node from the set is identical (key-space-wise) to the rest of the sets nodes or is a key space on a node made of smaller units and a copy of that smaller unit can be distributed independently of the rest of units on the node? DHT (distributed hast table) and DHT key-ring literature has examples of both approaches. I personally prefer a full back-up (identical copies) because they are easy to create and manage.
Another question is the state of node coming back from the dead and the definition of "dead". See that in the next post.
On a side: it is difficult to find time to write while vacationing on Hawaii!
No comments:
Post a Comment