1. 13 Apr, 2017 3 commits
  2. 12 Apr, 2017 4 commits
    • Kirill Smelkov's avatar
      Simplify "splitX bug" fix · 15a07c50
      Kirill Smelkov authored
      We are simplyfing commit 5c732b36 (Fix splitX bug.) here:
      
      By always passing to splitX i of index entry that is actually going to
      be used for traversal(*) we don't need to consider edge cases, such as
      when to move to next slot, inside splitX itself and this way there is no
      need to keep relookup logic there.
      
      So inside splitX it becomes very simple:
      
      	- i <= kx  -> go left
      	- i > kx   -> go right
      
      The logic to advance i by 1 on exact key hits was already there in Set
      and Put so it is natural for them to pass correct slot index to splitX.
      
      (*) i.e. index of next slot on exact index entry key hit - see previous
          patch for explanation
      15a07c50
    • Kirill Smelkov's avatar
      Add test for splitX bug fix covering logic handling splits on edges · 2a93ae07
      Kirill Smelkov authored
      Commit 5c732b36 (Fix splitX bug.) in particular modified splitX to redo
      the search on current level when key hits exactly kx slot in splitted X.
      
      However this is not tested at all as with current tests nothing breaks with the
      following patch:
      
      	@@ -546,10 +546,11 @@ func (t *Tree) Set(k interface{} /*K*/, v interface{} /*V*/) {
      	                i, ok := t.find(q, k)
      	                if ok {
      	                        switch x := q.(type) {
      	                        case *x:
      	                                if x.c > 2*kx {
      	+                                       panic("splitX ; ok=T")
      	                                        x, i = t.splitX(p, x, pi, i)
      	                                }
      	                                pi = i + 1
      	                                p = x
      	                                q = x.x[i+1].ch
      
      The relookup handling is needed exactly for ok=true case bacause if key is in
      xk slot there are 2 cases:
      
      1) key < x.x[xk].k  : here splitX is ok to let search follow left split child
      2) key = x.x[xk].k  : here splitX needs to let search follow right split child (*)
      
      and splitX, when knowing key comes from xk entry, does not bother itself with
      deciding which way to go (1 or 2) and for this case just follows up to make the
      search relookup current level after split.
      
      We can make existing tests cover this logic by raising N in TestSetGet2.
      However for kx=32, kd=32 N has to be increased more than 100x which, on my
      computer, increases the time to run TestSetGet2 from 0.5s to more than 200s,
      and still with it it only covers case when X does not have a parent.
      
      So instead of increasing N to be very large and slow down testing, let's add
      explicit test for this particular case.
      
      (*) reminder:
      
      `i, ok := find(q, k)` finds i corresponding to entry in q with min(k' : k <= k')
      (entry at i=q.c is always handled as with k'=∞)
      
      NOTE the "or-equal" in comparision. However keys in data (d) and index (x) page
      entries are treated differently. When find finds an entry with exact match (ok=true):
      
      - for d: d[i] is used
      
      	https://github.com/cznic/b/blob/aaaa43c9/btree.go#L409
      	https://github.com/cznic/b/blob/aaaa43c9/btree.go#L558
      
      - for x: x[i+1] is used:
      
      	https://github.com/cznic/b/blob/aaaa43c9/btree.go#L406
      	https://github.com/cznic/b/blob/aaaa43c9/btree.go#L555
      
      in other words keys located at index page entries says: all keys in child are
      strictly less than this key.
      
      This is probably so organized to help splitting - so split can pick lowest data
      entry from right data page and use its key as key in parent index entry for
      left data sibling directly.
      2a93ae07
    • Kirill Smelkov's avatar
      example: Regenerate int.go + cherry-pick corresponding all_test.go updates · 0f0644ab
      Kirill Smelkov authored
      example/int.go receives updates from aaaa43c9 (Fix Prev() to behave like
      Next() in reverse).
      
      We have to also update example/all_test.go because otherwise tests stop
      passing in example/ as the above commit changed both Prev and test
      logic.
      
      There are more missed updates to example/all_test.go - here I pick only
      minimum to keep tests passing in example/ .
      0f0644ab
    • Kirill Smelkov's avatar
      . · 3bddfd4e
      Kirill Smelkov authored
      3bddfd4e
  3. 11 Apr, 2017 21 commits
  4. 10 Apr, 2017 12 commits