Plans for 1.8+ (2.0?)

Matthew Dillon dillon at apollo.backplane.com
Thu Feb 1 12:53:39 PST 2007


:I can't uderstand whether snapshots are filesystems or files ?...or
:just both possible ?

    A snapshot is just a view into a historical data set.

    Lets say you have a piece of data that you are modifying every once in
    a while.  In a normal filesystem that data is overwritten on the physical
    media each time it is modified (or at least, each time the OS decides to
    flush its caches to disk).

    A filesystem which supports a historical data store works differently.
    Each flush to disk does NOT overwrite the previous data, it simply tags
    it as being replaced with a monotonically increasing transaction-id
    (which can double as a time stamp in many cases).

    So for example lets say you have the file "fubar" and you write
    "hello world" to it.  The historical data store would have something
    like this for the file 'fubar':

	created replaced	data
	------- -------- 	--------------------------
	10:00   -		hello world

    Then you modify the file to say 'goodbye world': 

	created replaced	data
	------- -------- 	--------------------------
	10:00   10:20		hello world
	10:20	-		goodbye world

    Then you modify the file to say 'hello world AGAIN!':

	created replaced	data
	------- -------- 	--------------------------
	10:00   10:20		hello world
	10:20	10:30		goodbye world
	10:30	-		hello world AGAIN!

    And so on and so forth.  Mounting a snapshot of the file is as simple
    as telling the filesystem the 'as of' date that you want your view to
    be.  The filesystem then simply ignores any records created after the
    as-of date and ignores any records replaced before the as-of date.

    And, poof, you see the filesystem as it was as-of the specified date.

    As an added bonus, the fact that the data content for any given physical
    block is not updated once written means that you can trivially
    replicate the data store.

    Deletion of old data is just as easy.  Lets say you decide you do not
    need any data you deleted before 10:31 ?  Scanning the transaction id's
    (timestamps in my example) for any record marked 'deleted' prior to
    10:31 yields a list of records which can be physically destroyed and
    their space recovered for future use.

	created replaced	data
	------- -------- 	--------------------------
	10:30	-		hello world AGAIN!

    The actual mechanism is somewhat more complex, because you need to index
    the data elements to reduce the overhead of scanning potentially thousands
    of dead records, but that is the jist of how a historical data store
    works.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>






More information about the Kernel mailing list