What is the durablility guarantee?

[Jon Derrick]

What is the durability guarantee in DistributedLog? Does every write have to flush to disk? When does it flush? when calling write is completed, or in the callback of the future?

[Jia Zhai]

Here is my understanding based on Bookkeeper. :)
When you get the callback, the write is persist.  Every write will flush to disk. But the callback could be got earlier than that,
In Bookkeeper, The data is first write into a journal file, then it could do the callback. The real flush happens later. If system paniced before the real flush happened (the data is in Bookies system cache), Bookkeeper could do log-replay from journal file, which kept the write content, to make the write persist on disk. 

[Jon Derrick]

Thanks Jia.

jderrick