openvstorage/alba

reduce frequency of dataloss with copysets/chainsets

Opened this issue · 2 comments

domsj commented

There's a tradeoff that can be made between frequency of dataloss and amount of data lost on each occurence, see http://hackingdistributed.com/2014/02/14/chainsets/ .
(In the long run the same amount of data will still be lost on average though.)

Note that the contents of the article are applicable not only for alba but for arakoon placement too

@Romain , please share your views on this. Should we change anything on ALBA. Isn't this a very hypothetical question as dataloss is simply not acceptable.

Copysets are not hard to implement, and it's a local change. Chainsets need more changes.
This can also considered to be an argument to have wider erasure coding policies (with a higher number of parity fragments).