Improve DC and InterDC Functionality (#408)

* Test case for issue #400 * Removed stable riak metadata and improved create dc. * Changed level of pending changes log messages * Added a better named function to add nodes to a current dc * Changed description and test case for dc descriptors. Fixed #400 * Reverted handle call for periodic updates, not needed * Bumped to version 0.2.2 * Reduced severity of log message for missing external_descriptors * Update include/antidote.hrl * Added leave_dc functionality We maintain our own dc id based on the riak_core_ring:cluster_name/1 command. Riak maintains the cluster name itself, and contrary to the comment in dc_utilities, it is restored on fail and restart correctly, even if all nodes fail at the same time. This caused errors in maintaining the ring members after simplifying the create_dc function and errors in shard count meta data. Switching back to the riak internal meta data (which is managed and updated by riak_core anyway) fixes these issues. The create_dc simplification also improved joining an empty cluster significantly (10 seconds for 8 shards). This is also work towards being able to add/remove nodes dynamically to scale after some time elapsed. Co-Authored-By: Peter Zeller <[email protected]>
AntidoteDB · Nov 28, 2019 · f6bf21b · f6bf21b
1 parent a597f73
commit f6bf21b
Show file tree

Hide file tree

Showing 21 changed files with 218 additions and 353 deletions.
diff --git a/config/sys.config.src b/config/sys.config.src
@@ -28,7 +28,14 @@
   {riak_core, [
     %% riak directories
     {ring_state_dir, "${ROOT_DIR_PREFIX}${DATA_DIR_PREFIX}data_riak_core"},
-    {platform_data_dir, "${ROOT_DIR_PREFIX}${DATA_DIR_PREFIX}data_riak_core"}
+    {platform_data_dir, "${ROOT_DIR_PREFIX}${DATA_DIR_PREFIX}data_riak_core"},
+
+    %% determines how many vnodes will be used
+    %% also determines the number of files the log is sliced into
+    %% has to be an exponent of 2
+    %% low number will decrease file accesses (good for testing) and boot time
+    %% high number enables scaling and generates smaller log files
+    {ring_creation_size, 64}
   ]},
 
 

diff --git a/include/antidote.hrl b/include/antidote.hrl
@@ -185,7 +185,9 @@
 -type op() :: {op_name(), op_param()}.
 -type effect() :: term().
 
--type dcid() :: 'undefined' | {atom(),tuple()}. %% TODO, is this the only structure that is returned by riak_core_ring:cluster_name(Ring)?
+
+%% DC Id is the riak_core ring cluster name
+-type dcid() :: undefined | riak_core_ring:riak_core_ring().
 -type snapshot_time() :: 'undefined' | vectorclock:vectorclock().
 -type clock_time() :: non_neg_integer().
 -type dc_and_commit_time() :: {dcid(), clock_time()}.

diff --git a/rebar.config b/rebar.config
@@ -100,7 +100,7 @@
         ]}
 ]}.
 
-{relx, [{release, {antidote, "0.2.1"}, [antidote]},
+{relx, [{release, {antidote, "0.2.2"}, [antidote]},
     {dev_mode, false},
     % do not expect Erlang runtime at deployment site
     {include_erts, true},

diff --git a/src/antidote.app.src b/src/antidote.app.src
@@ -1,7 +1,7 @@
 %% -*- erlang -*-
 {application, antidote, [
     {description, "A transactional CRDT database"},
-    {vsn, "0.2.1"},
+    {vsn, "0.2.2"},
     {applications, [
         kernel,
         stdlib,

diff --git a/src/antidote_dc_manager.erl b/src/antidote_dc_manager.erl
@@ -28,11 +28,22 @@
 
 
 %% This module exports methods to build a data center of multiple nodes and
-%% connect data centers to start replcation among them.
-%%  Usage Example: To create 3 DCs of 2 nodes each:
-%%    create_dc(['antidote@node1', 'antidote@node2']),
-%%    create_dc(['antidote@node3', 'antidote@node4']),
-%%    create_dc(['antidote@node5', 'antidote@node6']),
+%% connect data centers to start replication among them.
+%%
+%%  Usage Example:
+%%
+%%    To create 3 DCs of 2 nodes each execute
+%%
+%%    add_nodes_to_dc(['antidote@node1', 'antidote@node2'])
+%%    add_nodes_to_dc(['antidote@node3', 'antidote@node4'])
+%%    add_nodes_to_dc(['antidote@node5', 'antidote@node6'])
+%%
+%%    on one node of the pair of nodes.
+%%    (Single) Nodes will join the data center of the node the 'add_nodes_to_dc' function is executed on.
+%%    The `add_nodes_to_dc` function is idempotent.
+%%
+%%    To connect these data centers together execute
+%%
 %%    {ok, Descriptor1} = get_connection_descriptor() % on antidote@node1
 %%    {ok, Descriptor2} = get_connection_descriptor() %% on antidote@node3
 %%    {ok, Descriptor3} = get_connection_descriptor() %% on antidote@node5
@@ -45,41 +56,28 @@
 -include("inter_dc_repl.hrl").
 -include_lib("kernel/include/logger.hrl").
 
--export([create_dc/1,
-         get_connection_descriptor/0,
-         subscribe_updates_from/1]
-       ).
+-export([
+    leave_dc/0,
+    create_dc/1,
+    add_nodes_to_dc/1,
+    get_connection_descriptor/0,
+    subscribe_updates_from/1
+]).
+
+
+%% Command this node to leave the current data center
+-spec leave_dc() -> ok | {error, term()}.
+leave_dc() -> riak_core:leave().
 
 %% Build a ring of Nodes forming a data center
--spec create_dc([node()]) -> ok.
-create_dc(Nodes) ->
-    ?LOG_INFO("Creating DC ring with nodes ~p", [Nodes]),
-    %% Ensure each node owns 100% of it's own ring
-    _ = [[Node] = owners_according_to(Node) || Node <- Nodes],
-    %% Join nodes
-    [Node1 | OtherNodes] = Nodes,
-    case OtherNodes of
-        [] ->
-            %% no other nodes, nothing to join/plan/commit
-            ok;
-        _ ->
-            %% ok do a staged join and then commit it, this eliminates the
-            %% large amount of redundant handoff done in a sequential join
-            [staged_join(Node, Node1) || Node <- OtherNodes],
-            plan_and_commit(Node1),
-            try_nodes_ready(Nodes, 3, 500)
-    end,
-
-    ok = wait_until_nodes_ready(Nodes),
-
-    %% Ensure each node owns a portion of the ring
-    wait_until_nodes_agree_about_ownership(Nodes),
-    ok = wait_until_no_pending_changes(Nodes),
-    wait_until_ring_converged(Nodes),
-    ok = wait_until(hd(Nodes), fun wait_init:check_ready/1),
-    %% starts metadata services needed for intra-dc communication
-    ok = inter_dc_manager:start_bg_processes(stable_time_functions),
-    ok.
+-spec add_nodes_to_dc([node()]) -> ok | {error, ring_not_ready}.
+add_nodes_to_dc(Nodes) ->
+    %% check if ring is ready first
+    case riak_core_ring:ring_ready() of
+        true -> join_new_nodes(Nodes);
+        _ -> {error, ring_not_ready}
+    end.
+
 
 %% Start receiving updates from other DCs
 -spec subscribe_updates_from([descriptor()]) -> ok.
@@ -89,163 +87,96 @@ subscribe_updates_from(DCDescriptors) ->
     ok = inter_dc_manager:dc_successfully_started(),
     ok.
 
+
 %% Get the DC connection descriptor to be given to other DCs
 -spec get_connection_descriptor() -> {ok, descriptor()}.
 get_connection_descriptor() ->
     inter_dc_manager:get_descriptor().
 
+
 %% ---------- Internal Functions --------------
 
-%% @doc Return a list of nodes that own partitions according to the ring
-%%      retrieved from the specified node.
-owners_according_to(Node) ->
-    case rpc:call(Node, riak_core_ring_manager, get_raw_ring, []) of
-        {ok, Ring} ->
-            ?LOG_INFO("Ring ~p", [Ring]),
-            Owners = [Owner || {_Idx, Owner} <- riak_core_ring:all_owners(Ring)],
-            ?LOG_INFO("Owners ~p", [lists:usort(Owners)]),
-            lists:usort(Owners);
-        {badrpc, Reason} ->
-            ?LOG_INFO("Could not connect to Node ~p", [Node]),
-            {badrpc, Reason}
-    end.
+-spec join_new_nodes([node()]) -> ok.
+join_new_nodes(Nodes) ->
+    %% get the current ring
+    {ok, CurrentRing} = riak_core_ring_manager:get_my_ring(),
+
+    %% filter nodes that are not already in this nodes ring
+    CurrentNodeMembers = riak_core_ring:all_members(CurrentRing),
+    NewNodeMembers = [NewNode || NewNode <- Nodes, not lists:member(NewNode, CurrentNodeMembers)],
+    plan_and_commit(NewNodeMembers).
+
 
-%% @doc Have `Node' send a join request to `PNode'
-staged_join(Node, PNode) ->
-    ?LOG_INFO("[join] ~p to (~p)", [Node, PNode]),
-    ok = rpc:call(Node, riak_core, staged_join, [PNode]),
+-spec plan_and_commit([node()]) -> ok.
+plan_and_commit([]) -> ?LOG_WARNING("No new nodes added to the ring of ~p", [node()]);
+plan_and_commit(NewNodeMembers) ->
+    lists:foreach(fun(Node) ->
+        ?LOG_INFO("Checking if Node ~p is reachable (from ~p)", [Node, node()]),
+        pong = net_adm:ping(Node)
+                  end, NewNodeMembers),
+
+    lists:foreach(fun(Node) ->
+        ?LOG_INFO("Node ~p is joining my ring (~p)", [Node, node()]),
+        ok = rpc:call(Node, riak_core, staged_join, [node()])
+                  end, NewNodeMembers),
+
+    lists:foreach(fun(Node) ->
+        ?LOG_INFO("Checking if node ring is ready (~p)", [Node]),
+        wait_until_ring_ready(Node)
+                  end, NewNodeMembers),
+
+    {ok, Actions, Transitions} = riak_core_claimant:plan(),
+    ?LOG_DEBUG("Actions planned: ~p", [Actions]),
+    ?LOG_DEBUG("Ring transitions planned: ~p", [Transitions]),
+
+    %% only after commit returns ok the ring structure will change
+    %% even if nothing changes, it returns {error, nothing_planned} indicating some serious error
+    ok = riak_core_claimant:commit(),
+    ?LOG_NOTICE("Ring committed and ring structure is changing. New ring members: ~p", [NewNodeMembers]),
+
+    %% wait until ring is ready
+    wait_until_ring_ready(node()),
+
+    %% wait until ring has no pending changes
+    %% this prevents writing to a ring which has not finished its balancing yet and therefore causes
+    %% handoffs to be triggered
+    %% FIXME this can be removed when #401 and #203 is fixed
+    wait_until_ring_no_pending_changes(),
+
+
+    %% start periodic heart beat
+    ok = inter_dc_manager:start_bg_processes(stable_time_functions),
     ok.
 
-plan_and_commit(Node) ->
-    ?LOG_INFO("Planning and committing cluster join"),
-    case rpc:call(Node, riak_core_claimant, plan, []) of
-        {error, ring_not_ready} ->
-            ?LOG_INFO("plan: ring not ready"),
-            maybe_wait_for_changes(Node),
-            plan_and_commit(Node);
-        {ok, _, _} ->
-            do_commit(Node)
-    end.
-do_commit(Node) ->
-    ?LOG_INFO("Committing"),
-    case rpc:call(Node, riak_core_claimant, commit, []) of
-        {error, plan_changed} ->
-            ?LOG_INFO("commit: plan changed"),
-            maybe_wait_for_changes(Node),
-            plan_and_commit(Node);
-        {error, ring_not_ready} ->
-            ?LOG_INFO("commit: ring not ready"),
-            maybe_wait_for_changes(Node),
-            do_commit(Node);
-        {error, nothing_planned} ->
-            %% Assume plan actually committed somehow
-            ok;
-        ok ->
-            ok
-    end.
 
-try_nodes_ready([Node1 | _Nodes], 0, _SleepMs) ->
-      ?LOG_INFO("Nodes not ready after initial plan/commit, retrying"),
-      plan_and_commit(Node1);
-  try_nodes_ready(Nodes, N, SleepMs) ->
-      ReadyNodes = [Node || Node <- Nodes, is_ready(Node) =:= true],
-      case ReadyNodes of
-          Nodes ->
-              ok;
-          _ ->
-              timer:sleep(SleepMs),
-              try_nodes_ready(Nodes, N-1, SleepMs)
-      end.
-
-maybe_wait_for_changes(Node) ->
-    wait_until_no_pending_changes([Node]).
-
-%% @doc Given a list of nodes, wait until all nodes believe there are no
+%% @doc Wait until all nodes in this ring believe there are no
 %% on-going or pending ownership transfers.
--spec wait_until_no_pending_changes([node()]) -> ok.
-wait_until_no_pending_changes(Nodes) ->
-    ?LOG_INFO("Wait until no pending changes on ~p", [Nodes]),
+-spec wait_until_ring_no_pending_changes() -> ok.
+wait_until_ring_no_pending_changes() ->
+    {ok, CurrentRing} = riak_core_ring_manager:get_my_ring(),
+    Nodes = riak_core_ring:all_members(CurrentRing),
+
+    ?LOG_DEBUG("Wait until no pending changes on ~p", [Nodes]),
     F = fun() ->
-                _ = rpc:multicall(Nodes, riak_core_vnode_manager, force_handoffs, []),
-                {Rings, BadNodes} = rpc:multicall(Nodes, riak_core_ring_manager, get_raw_ring, []),
-                Changes = [ riak_core_ring:pending_changes(Ring) =:= [] || {ok, Ring} <- Rings ],
-                BadNodes =:= [] andalso length(Changes) =:= length(Nodes) andalso lists:all(fun(T) -> T end, Changes)
+        _ = rpc:multicall(Nodes, riak_core_vnode_manager, force_handoffs, []),
+        {Rings, BadNodes} = rpc:multicall(Nodes, riak_core_ring_manager, get_raw_ring, []),
+        Changes = [ riak_core_ring:pending_changes(Ring) =:= [] || {ok, Ring} <- Rings ],
+        BadNodes =:= [] andalso length(Changes) =:= length(Nodes) andalso lists:all(fun(T) -> T end, Changes)
         end,
-    wait_until(F).
-
-%% @doc Utility function used to construct test predicates. Retries the
-%%      function `Fun' until it returns `true', or until the maximum
-%%      number of retries is reached.
-wait_until(Fun) when is_function(Fun) ->
-    MaxTime = 600000, %% @TODO use config,
-        Delay = 1000, %% @TODO use config,
-        Retry = MaxTime div Delay,
-    wait_until(Fun, Retry, Delay).
-
-%% @doc Given a list of nodes, wait until all nodes are considered ready.
-%%      See {@link wait_until_ready/1} for definition of ready.
-wait_until_nodes_ready(Nodes) ->
-    ?LOG_INFO("Wait until nodes are ready : ~p", [Nodes]),
-    [ok = wait_until(Node, fun is_ready/1) || Node <- Nodes],
-    ok.
-
-%% @private
-is_ready(Node) ->
-    case rpc:call(Node, riak_core_ring_manager, get_raw_ring, []) of
-        {ok, Ring} ->
-            case lists:member(Node, riak_core_ring:ready_members(Ring)) of
-                true -> true;
-                false -> {not_ready, Node}
-            end;
-        Other ->
-            Other
+    case F() of
+        true -> ok;
+        _ -> timer:sleep(500), wait_until_ring_no_pending_changes()
     end.
 
-wait_until_nodes_agree_about_ownership(Nodes) ->
-    ?LOG_INFO("Wait until nodes agree about ownership: ~p", [Nodes]),
-    Results = [ wait_until_owners_according_to(Node, Nodes) || Node <- Nodes ],
-    lists:all(fun(X) -> ok =:= X end, Results).
-
-%% @doc Convenience wrapper for wait_until for the myriad functions that
-%% take a node as single argument.
-wait_until(Node, Fun) when is_atom(Node), is_function(Fun) ->
-    wait_until(fun() -> Fun(Node) end).
-
-wait_until(Fun, Retry, Delay) when Retry > 0 ->
-    wait_until_result(Fun, true, Retry, Delay).
-
-wait_until_result(Fun, Result, Retry, Delay) when Retry > 0 ->
-    Res = Fun(),
-    case Res of
-        Result ->
-            ok;
-        _ when Retry == 1 ->
-            {fail, Res};
-        _ ->
-            timer:sleep(Delay),
-            wait_until_result(Fun, Result, Retry-1, Delay)
-    end.
 
-wait_until_owners_according_to(Node, Nodes) ->
-  SortedNodes = lists:usort(Nodes),
-  F = fun(N) ->
-      owners_according_to(N) =:= SortedNodes
-  end,
-  ok = wait_until(Node, F),
-  ok.
-
-%% @private
-is_ring_ready(Node) ->
-    case rpc:call(Node, riak_core_ring_manager, get_raw_ring, []) of
-        {ok, Ring} ->
-            riak_core_ring:ring_ready(Ring);
-        _ ->
-            false
+-spec wait_until_ring_ready(node()) -> ok.
+wait_until_ring_ready(Node) ->
+    Status = rpc:call(Node, riak_core_ring, ring_ready, []),
+    case Status of
+        true -> ok;
+        false -> timer:sleep(100), wait_until_ring_ready(Node)
     end.
 
-%% @doc Given a list of nodes, wait until all nodes believe the ring has
-%%      converged (ie. `riak_core_ring:is_ready' returns `true').
-wait_until_ring_converged(Nodes) ->
-    ?LOG_INFO("Wait until ring converged on ~p", [Nodes]),
-    [ok = wait_until(Node, fun is_ring_ready/1)|| Node <- Nodes],
-    ok.
+%% backwards compatible function for add_nodes_to_dc
+-spec create_dc([node()]) -> ok | {error, ring_not_ready}.
+create_dc(Nodes) -> add_nodes_to_dc(Nodes).
diff --git a/src/antidote_ring_event_handler.erl b/src/antidote_ring_event_handler.erl
@@ -32,6 +32,8 @@
 -include("antidote.hrl").
 -include_lib("kernel/include/logger.hrl").
 
+-export([update_status/0]).
+
 %% gen_event callbacks
 -export([init/1, handle_event/2, handle_call/2,
          handle_info/2, terminate/2, code_change/3]).