[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: LMDB stuff

Le 2/4/14 8:13 AM, Howard Chu a Ãcrit :
> Hallvard Breien Furuseth wrote:
>> On 2014-02-03 22:14, Howard Chu wrote:
>>> Was chatting with Emmanuel Lecharny (who is currently working on
>>> Mavibot for ApacheDS, an MVCC backend similar to LMDB) and had an
>>> interesting realization: we can avoid the current issue of long-lived
>>> reader txns preventing page reclamation.
>>> (...) since we're already going to add a txnID to every page's
>>> page header, we can simply add a 2nd txnID, recording the txnID of the
>>> previous change to this page's ancestor. Then, any page where this
>>> prevTxnID is >= the outstanding reader's txnID can be reclaimed.
>> Nice.  But: Seems to me the freelist needs to know when the page
>> was written, and when it was freed.  A reader older than the txn
>> which wrote the current contents of a page, is irrelevant to
>> whether the page can be overwritten.  How do ancestors matter? (Do
>> you mean a branch page? The age of the previous page contents?)
> I meant the age of the previous page contents.
>>> Still thinking about the actual implementation of this, it may make
>>> more sense to store the prevTxnID in the freeDB than in each page
>>> header. Ideally we want to be able to grab a chunk of pageIDs
>>> unambiguously, instead of having to iterate thru each page and read
>>> its header to determine if it's safe.
>> Yes, re-reading lots of pages just to find they can't be used does
>> not sound fun.  And we can't look for page headers in a broken-up
>> overflow page.  Unless we either quit breaking up freed ovpages,
>> or write a page header to the unused chunk when breaking it up.
>> Can we grab Mavibot's freeDB structure?
> Currently we're just using an array of pageIDs, we can turn that into
> an array of <pageID,prevTxnID> pairs instead.
I don't know what's inside the Free List in LMDB, but I can only assume
it's closed to what we have in Mavibot. Let me explain what we do (or
will do !) in Mavibot, and what data structure we use for that.

We have 2 special b-trees to manage the other b-trees :
- the btreeOfBtrees (BOB) which contain information about every version
of btree still in use (ie, if a search is holding rev N, then BOB will
hold btree N).
- the copiedPagesBtree (CPB) store an nformation of every copied pages,
associated with the revision which copied the page. In fact, we store a
tuple <rev, list of copied pages>

We also maintain a list of free pages (and here, free pages mean pages
that can be immediately reclaimed when we need some new page)

My guess is that the Free List is somehow a combinaison of the free
pages list and the CPB b-tree we have in Mavibot.

At this point, the mechanism we foresee for pages release is the following :
- every copied root page can be release immediately, if no search is
using the revision (or, better said, when a search is completed, and no
other search is using the revision, we can release the root page if it
has been copied). This is very basic, as no other search will ever use
this page.
- for other copied pages, it's slightly more complex. When we release a
search, and if the associated revision has not any more search
associated, then we can do some cleanup. We will brows the CPB b-tree
backward until we find a revision which is hold by a search request.
Every elements found during the CPB browse can now be processed. If a
page in this set of pages has been copied, we can safely remove the
copied page, it's not in use anymore.

So if we have these elements in the CPB, when we release N and when the
closest oldest search is N-3 :

rev(N) -> {p100, p120, p145}
rev(N-1) -> {p99, p145}
rev(N-2) -> {p45, p120, p145}
rev(N-3) -> {p99, p100, p111, p145}

we can see that p145 has been copied twice, and p120 once. We can
release both p145(N-1,N-2) pages, and p120(N-2) safely. OTOH, p100(N) is
still in use by N-3, so we have to keep it (but it will be removed if
any N+x revision is copying it)

What is needed at this point is to have a unique id for each page (and
this id should remain when we copy a page), and a way to keep the tuple
<rev, copied pages>. The algorithm is linear (O(n log(m) ) where n is
the number of revisions between the released one and the closest oldest
revision in use, and log(m) represent the number of copied pages for
each revision : it's likely to be the height of the tree)

I don't know how it fits with LMDB implementation, I just wanted to give
you an insight on what we are working on. Hope it helps.

Emmanuel LÃcharny