reject write at leader if conflict

This should prevent spurious intra-cluster conflicts most of the
time. It is not true consistency, however. spurious conflicts are
still possible whenever the nodes in the cluster disagree on the
current live set of other nodes.

Author: rnewson

src/fabric/src/fabric_doc_update.erl

@@ -22,7 +22,10 @@
     doc_count,
     w,
     grouped_docs,
-    reply
+    reply,
+    update_options,
+    leaders = [],
+    started = []
 }).
 
 go(_, [], _) ->
@@ -33,25 +36,25 @@ go(DbName, AllDocs0, Opts) ->
     validate_atomic_update(DbName, AllDocs, lists:member(all_or_nothing, Opts)),
     Options = lists:delete(all_or_nothing, Opts),
     GroupedDocs = lists:map(
-        fun({#shard{name = Name, node = Node} = Shard, Docs}) ->
-            Docs1 = untag_docs(Docs),
-            Ref = rexi:cast(Node, {fabric_rpc, update_docs, [Name, Docs1, Options]}),
-            {Shard#shard{ref = Ref}, Docs}
+        fun({#shard{} = Shard, Docs}) ->
+            {Shard#shard{ref = make_ref()}, Docs}
         end,
         group_docs_by_shard(DbName, AllDocs)
     ),
     {Workers, _} = lists:unzip(GroupedDocs),
     RexiMon = fabric_util:create_monitors(Workers),
     W = couch_util:get_value(w, Options, integer_to_list(mem3:quorum(DbName))),
     Acc0 = #acc{
+        update_options = Options,
         waiting_count = length(Workers),
         doc_count = length(AllDocs),
         w = list_to_integer(W),
         grouped_docs = GroupedDocs,
         reply = dict:new()
     },
     Timeout = fabric_util:request_timeout(),
-    try rexi_utils:recv(Workers, #shard.ref, fun handle_message/3, Acc0, infinity, Timeout) of
+    Acc1 = start_leaders(Acc0),
+    try rexi_utils:recv(Workers, #shard.ref, fun handle_message/3, Acc1, infinity, Timeout) of
         {ok, {Health, Results}} when
             Health =:= ok; Health =:= accepted; Health =:= error
         ->
@@ -72,61 +75,78 @@ go(DbName, AllDocs0, Opts) ->
         rexi_monitor:stop(RexiMon)
     end.
 
-handle_message({rexi_DOWN, _, {_, NodeRef}, _}, _Worker, #acc{} = Acc0) ->
+handle_message({rexi_DOWN, _, {_, NodeRef}, _}, Worker, #acc{} = Acc0) ->
     #acc{grouped_docs = GroupedDocs} = Acc0,
     NewGrpDocs = [X || {#shard{node = N}, _} = X <- GroupedDocs, N =/= NodeRef],
-    skip_message(Acc0#acc{waiting_count = length(NewGrpDocs), grouped_docs = NewGrpDocs});
+    Acc1 = Acc0#acc{waiting_count = length(NewGrpDocs), grouped_docs = NewGrpDocs},
+    Acc2 = start_followers(Worker, Acc1),
+    skip_message(Acc2);
 handle_message({rexi_EXIT, _}, Worker, #acc{} = Acc0) ->
     #acc{waiting_count = WC, grouped_docs = GrpDocs} = Acc0,
     NewGrpDocs = lists:keydelete(Worker, 1, GrpDocs),
-    skip_message(Acc0#acc{waiting_count = WC - 1, grouped_docs = NewGrpDocs});
+    Acc1 = Acc0#acc{waiting_count = WC - 1, grouped_docs = NewGrpDocs},
+    Acc2 = start_followers(Worker, Acc1),
+    skip_message(Acc2);
 handle_message({error, all_dbs_active}, Worker, #acc{} = Acc0) ->
     % treat it like rexi_EXIT, the hope at least one copy will return successfully
     #acc{waiting_count = WC, grouped_docs = GrpDocs} = Acc0,
     NewGrpDocs = lists:keydelete(Worker, 1, GrpDocs),
-    skip_message(Acc0#acc{waiting_count = WC - 1, grouped_docs = NewGrpDocs});
+    Acc1 = Acc0#acc{waiting_count = WC - 1, grouped_docs = NewGrpDocs},
+    Acc2 = start_followers(Worker, Acc1),
+    skip_message(Acc2);
 handle_message(internal_server_error, Worker, #acc{} = Acc0) ->
     % happens when we fail to load validation functions in an RPC worker
     #acc{waiting_count = WC, grouped_docs = GrpDocs} = Acc0,
     NewGrpDocs = lists:keydelete(Worker, 1, GrpDocs),
-    skip_message(Acc0#acc{waiting_count = WC - 1, grouped_docs = NewGrpDocs});
+    Acc1 = Acc0#acc{waiting_count = WC - 1, grouped_docs = NewGrpDocs},
+    Acc2 = start_followers(Worker, Acc1),
+    skip_message(Acc2);
 handle_message(attachment_chunk_received, _Worker, #acc{} = Acc0) ->
     {ok, Acc0};
 handle_message({ok, Replies}, Worker, #acc{} = Acc0) ->
     #acc{
-        waiting_count = WaitingCount,
         doc_count = DocCount,
         w = W,
         grouped_docs = GroupedDocs,
         reply = DocReplyDict0
     } = Acc0,
-    {value, {_, Docs}, NewGrpDocs} = lists:keytake(Worker, 1, GroupedDocs),
-    DocReplyDict = append_update_replies(Docs, Replies, DocReplyDict0),
-    case {WaitingCount, dict:size(DocReplyDict)} of
-        {1, _} ->
-            % last message has arrived, we need to conclude things
-            {Health, W, Reply} = dict:fold(
-                fun force_reply/3,
-                {ok, W, []},
-                DocReplyDict
-            ),
-            {stop, {Health, Reply}};
-        {_, DocCount} ->
-            % we've got at least one reply for each document, let's take a look
-            case dict:fold(fun maybe_reply/3, {stop, W, []}, DocReplyDict) of
-                continue ->
-                    {ok, Acc0#acc{
-                        waiting_count = WaitingCount - 1,
-                        grouped_docs = NewGrpDocs,
-                        reply = DocReplyDict
-                    }};
-                {stop, W, FinalReplies} ->
-                    {stop, {ok, FinalReplies}}
-            end;
-        _ ->
-            {ok, Acc0#acc{
-                waiting_count = WaitingCount - 1, grouped_docs = NewGrpDocs, reply = DocReplyDict
-            }}
+    {value, {_, Docs}, NewGrpDocs0} = lists:keytake(Worker, 1, GroupedDocs),
+    IsLeader = lists:member(Worker#shard.ref, Acc0#acc.leaders),
+    DocReplyDict = append_update_replies(Docs, Replies, W, IsLeader, DocReplyDict0),
+    Acc1 = Acc0#acc{grouped_docs = NewGrpDocs0, reply = DocReplyDict},
+    Acc2 = remove_conflicts(Docs, Replies, Acc1),
+    NewGrpDocs = Acc2#acc.grouped_docs,
+    case skip_message(Acc2) of
+        {stop, Msg} ->
+            {stop, Msg};
+        {ok, Acc3} ->
+            Acc4 = start_followers(Worker, Acc3),
+            case {Acc4#acc.waiting_count, dict:size(DocReplyDict)} of
+                {1, _} ->
+                    % last message has arrived, we need to conclude things
+                    {Health, W, Reply} = dict:fold(
+                        fun force_reply/3,
+                        {ok, W, []},
+                        DocReplyDict
+                    ),
+                    {stop, {Health, Reply}};
+                {_, DocCount} ->
+                    % we've got at least one reply for each document, let's take a look
+                    case dict:fold(fun maybe_reply/3, {stop, W, []}, DocReplyDict) of
+                        continue ->
+                            {ok, Acc4#acc{
+                                waiting_count = Acc4#acc.waiting_count - 1,
+                                grouped_docs = NewGrpDocs
+                            }};
+                        {stop, W, FinalReplies} ->
+                            {stop, {ok, FinalReplies}}
+                    end;
+                _ ->
+                    {ok, Acc4#acc{
+                        waiting_count = Acc4#acc.waiting_count - 1,
+                        grouped_docs = NewGrpDocs
+                    }}
+            end
     end;
 handle_message({missing_stub, Stub}, _, _) ->
     throw({missing_stub, Stub});
@@ -318,13 +338,91 @@ group_docs_by_shard(DbName, Docs) ->
         )
     ).
 
-append_update_replies([], [], DocReplyDict) ->
+%% use 'lowest' node that hosts this shard range as leader
+is_leader(Worker, Workers) ->
+    Worker#shard.node ==
+        lists:min([W#shard.node || W <- Workers, W#shard.range == Worker#shard.range]).
+
+start_leaders(#acc{} = Acc0) ->
+    #acc{grouped_docs = GroupedDocs} = Acc0,
+    {Workers, _} = lists:unzip(GroupedDocs),
+    LeaderRefs = lists:foldl(
+        fun({Worker, Docs}, RefAcc) ->
+            case is_leader(Worker, Workers) of
+                true ->
+                    start_worker(Worker, Docs, Acc0),
+                    [Worker#shard.ref | RefAcc];
+                false ->
+                    RefAcc
+            end
+        end,
+        [],
+        GroupedDocs
+    ),
+    Acc0#acc{leaders = LeaderRefs, started = LeaderRefs}.
+
+start_followers(#shard{} = Leader, #acc{} = Acc0) ->
+    Followers = [
+        {Worker, Docs}
+     || {Worker, Docs} <- Acc0#acc.grouped_docs,
+        Worker#shard.range == Leader#shard.range,
+        not lists:member(Worker#shard.ref, Acc0#acc.started)
+    ],
+    lists:foreach(
+        fun({Worker, Docs}) ->
+            start_worker(Worker, Docs, Acc0)
+        end,
+        Followers
+    ),
+    Started = [Ref || {#shard{ref = Ref}, _Docs} <- Followers],
+    Acc0#acc{started = lists:append([Started, Acc0#acc.started])}.
+
+start_worker(#shard{ref = Ref} = Worker, Docs, #acc{} = Acc0) when is_reference(Ref) ->
+    #shard{name = Name, node = Node} = Worker,
+    #acc{update_options = UpdateOptions} = Acc0,
+    rexi:cast_ref(Ref, Node, {fabric_rpc, update_docs, [Name, untag_docs(Docs), UpdateOptions]}),
+    ok;
+start_worker(#shard{ref = undefined}, _Docs, #acc{}) ->
+    % for unit tests below.
+    ok.
+
+append_update_replies([], [], _W, _IsLeader, DocReplyDict) ->
     DocReplyDict;
-append_update_replies([Doc | Rest], [], Dict0) ->
+append_update_replies([Doc | Rest], [], W, IsLeader, Dict0) ->
     % icky, if replicated_changes only errors show up in result
-    append_update_replies(Rest, [], dict:append(Doc, noreply, Dict0));
-append_update_replies([Doc | Rest1], [Reply | Rest2], Dict0) ->
-    append_update_replies(Rest1, Rest2, dict:append(Doc, Reply, Dict0)).
+    append_update_replies(Rest, [], W, IsLeader, dict:append(Doc, noreply, Dict0));
+append_update_replies([Doc | Rest1], [conflict | Rest2], W, true, Dict0) ->
+    %% fake conflict replies from followers as we won't ask them
+    append_update_replies(
+        Rest1, Rest2, W, true, dict:append_list(Doc, lists:duplicate(W, conflict), Dict0)
+    );
+append_update_replies([Doc | Rest1], [Reply | Rest2], W, IsLeader, Dict0) ->
+    append_update_replies(Rest1, Rest2, W, IsLeader, dict:append(Doc, Reply, Dict0)).
+
+%% leader found a conflict, remove that doc from the other (follower) workers,
+%% removing the worker entirely if no docs remain.
+remove_conflicts([], [], #acc{} = Acc0) ->
+    Acc0;
+remove_conflicts([Doc | DocRest], [conflict | ReplyRest], #acc{} = Acc0) ->
+    #acc{grouped_docs = GroupedDocs0} = Acc0,
+    GroupedDocs1 = lists:foldl(
+        fun({Worker, Docs}, FoldAcc) ->
+            case lists:delete(Doc, Docs) of
+                [] ->
+                    FoldAcc#acc{waiting_count = FoldAcc#acc.waiting_count - 1};
+                Rest ->
+                    [{Worker, Rest} | FoldAcc]
+            end
+        end,
+        [],
+        GroupedDocs0
+    ),
+    Acc1 = Acc0#acc{grouped_docs = GroupedDocs1},
+    remove_conflicts(DocRest, ReplyRest, Acc1);
+remove_conflicts([_Doc | DocRest], [_Reply | ReplyRest], #acc{} = Acc0) ->
+    remove_conflicts(DocRest, ReplyRest, Acc0);
+remove_conflicts([_Doc | DocRest], [], #acc{} = Acc0) ->
+    remove_conflicts(DocRest, [], Acc0).
 
 skip_message(#acc{waiting_count = 0, w = W, reply = DocReplyDict}) ->
     {Health, W, Reply} = dict:fold(fun force_reply/3, {ok, W, []}, DocReplyDict),