Read traversal and COW about btree HOT 1 OPEN

More info about the proposal.

I'm actually using https://github.com/tidwall/btree fork and implemented the entire idea on it:

1. The equivalent to what I proposed above: gallir/tidwall.btree.old@d659c20

2. Replaced mutableChildren() for insert and remove: gallir/tidwall.btree.old@1cae1dd

3. Replaced mutableChildren in growChildAndRemove(): gallir/tidwall.btree.old@ac1d4ab

4. Replaced mutableChildren in maybeSplitChild(): gallir/tidwall.btree.old@56f9d35

I'm testing it with real data and update from our production system (at dotw.com) and getting a lot less misses during concurrent traversals.

Although during the process I learnt that most of the time the nodes were and are not being reused with the exception of the merge node due to the check of the cow pointer if you do a clone after ever serie of updates. To improve it you have to avoid cloning and by doing it the current mechanism for freeing and reusing don't make too much sense for cases where updates are atomic/mutually exclusive.

from btree.