Merge pull request #972 from basho/mas-i970-simpleleave

Refactor leave - balance and tail violations

priv/riak_core.schema

@@ -202,6 +202,22 @@
     hidden
 ]}.
 
+%% @doc On cluster leave - force full rebalance partitions
+%% By default on a cluster leave there will first be an attempt to handoff
+%% vnodes to safe (in terms of target_n_val) locations.  In small clusters,
+%% there may be insufficient safe locations, and a temporary state can be
+%% created where a single node has a large number of vnodes.
+%% To mitigate this, a full rebalance (a re-assignment that does not optimise
+%% based on the starting position), can be forced by setting this option on
+%% all nodes.
+%% Please carefully consider any cluster plan created with this option before
+%% committing
+{mapping, "full_rebalance_onleave", "riak_core.full_rebalance_onleave", [
+    {datatype, flag},
+    {default, off}
+]}.
+
+
 %% Async Job Management
 %%
 %% This is a translation for mappings that appear in other schema files.

src/riak_core_claim.erl

@@ -64,7 +64,8 @@
          choose_claim_v2/1, choose_claim_v2/2, choose_claim_v2/3,
          choose_claim_v3/1, choose_claim_v3/2, choose_claim_v3/3,
          claim_rebalance_n/2, claim_diversify/3, claim_diagonal/3,
-         wants/1, wants_owns_diff/2, meets_target_n/2, diagonal_stripe/2]).
+         wants/1, wants_owns_diff/2, meets_target_n/2, diagonal_stripe/2,
+         sequential_claim/2, get_counts/2]).
 
 -ifdef(TEST).
 -compile(export_all).
@@ -620,6 +621,10 @@ claim_diagonal(Wants, Owners, Params) ->
            end,
     {lists:flatten([lists:duplicate(Reps, Claiming), Last]), [diagonalized]}.
 
+sequential_claim(Ring, Node) ->
+    TN = app_helper:get_env(riak_core, target_n_val, ?DEF_TARGET_N),
+    sequential_claim(Ring, Node, TN).
+
 %% @private fall back to diagonal striping vnodes across nodes in a
 %% sequential round robin (eg n1 | n2 | n3 | n4 | n5 | n1 | n2 | n3
 %% etc) However, different to `claim_rebalance_n', this function

src/riak_core_console.erl

@@ -685,6 +685,16 @@ print_plan(Changes, Ring, NextRings) ->
                       "cluster transitions~n~n", [Transitions])
     end,
 
+    Leaves = length(lists:filter(fun({_N, A}) -> A == leave end, Changes)) > 0,
+
+    case Leaves and app_helper:get_env(riak_core, full_rebalance_onleave) of
+        true ->
+            io:format("WARNING: Full rebalance forced by non-default "
+                        "option riak_core.full_rebalance_onleave = true~n~n");
+        _ ->
+            ok
+    end,
+
     _ = lists:foldl(fun({Ring1, Ring2}, I) ->
                            io:format("~79..#s~n", [""]),
                            io:format("~24.. s After cluster transition ~b/~b~n",