From: Thompson, B. B. <BRY...@sa...> - 2006-02-17 21:32:47
|
Alex, If a reader needs to leave a timestamp mark every time it reads a record, then that is not going to be very efficient since it turns reads into read + write operations. Some kinds of MVCC require read timestamps, but I think that we need to find a variation which does not. I personally like the idea of a log structured store very much and I am also thinking in terms of how to combine MVCC with that approach. It seems a natural fit. Log structured stores can help to get rid of some hotspots in allocation of records. However, if by "log structured" store you mean an append only store (which is how I interpret it) then I think that we also need to provide for people who want a store with destructive overwrites. I.e., where the space from "old" versions eventually gets reused rather than providing an immortal history of the consistent states of the database. I personally think that we can do both. The binary formats will differ, but the main difference is in the persistent record allocation logic if the store is already using MVCC. -bryan -----Original Message----- From: Alex Boisvert To: Thompson, Bryan B. Cc: 'jdb...@li...' Sent: 2/17/2006 4:24 PM Subject: Re: [Jdbm-developer] 2PL + MVCC design Sorry, I don't have a better answer to this at the moment. Could you clarify what you mean by "is not going to be very efficient"? I'm still toying with combining MVCC with the log-structured object store idea. Transactions being organized as a multi-ended linked-list corresponding to ordering and later transactions shadowing prior transactions. alex Thompson, Bryan B. wrote: >Alex, > >One problem with some MVCC designs is that the require writing >timestamps for readers. I expect that this is not going to be >very efficient. I have seen one pure MVCC design which avoids >this by how it generates timestamps. It also seems possible >that a 2PL + MVCC design could avoid this by using the 2PL to >induce the synchronization order and corresponding timestamps. > >Do you have any insight into this matter? > >-bryan > > |